Innovative Systems for Intelligent Health Informatics: Data Science, Health Informatics, Intelligent Systems, Smart Computing 3030707121, 9783030707125

Table of contents :
Preface
Organization
IRICT2020 Organizing Committee
Honorary Co-chairs
Conference General Chair
Program Committee Chair
General Secretary
Technical Committee Chairs
Publications Committee
Publicity Committee
IT and Multimedia Committee
Treasure Committee
Logistic Committee Chair
Registration
International Technical Committee
Contents
Intelligent Health Informatics
Comparative Study of SMOTE and Bootstrapping Performance Based on Predication Methods
1 Introduction
2 Related Works
3 Experiments
4 Discussion
5 Conclusion
References
UPLX: Blockchain Platform for Integrated Health Data Management
1 Introduction
2 Blockchain Technology
3 Hyperledger Fabric
4 UPLX Interoperable Architecture
5 UPLX Data Structure
6 Conclusion
References
Convolutional Neural Networks for Automatic Detection of Colon Adenocarcinoma Based on Histopathological Images
1 Introduction
2 Methods and Dataset
2.1 Dataset
2.2 Convolutional Neural Networks
2.3 Proposed Model
2.4 Transfer Learning and Fine-Tune
3 Results
4 Discussion and Conclusion
References
Intelligent Health Informatics with Personalisation in Weather-Based Healthcare Using Machine Learning
1 Introduction
2 Background Study
2.1 Influence of Weather on Asthma and Eczema
2.2 Methods for Patients to Self-monitor Asthma and Eczema
2.3 Machine Learning in Weather-Based Healthcare
3 Methodology
4 Results and Discussion
5 Conclusion
References
A CNN-Based Model for Early Melanoma Detection
1 Introduction
2 Background
2.1 Related Works
2.2 Melanoma and Nevus Lesions
3 Methodology
3.1 Experimental Dataset
3.2 AlexNet Pre-trained Model
3.3 GoogleNet Pre-trained Model
4 Results and Discussion
5 Conclusion
References
SMARTS D4D Application Module for Dietary Adherence Self-monitoring Among Hemodialysis Patients
1 Introduction
1.1 Study Background
1.2 SMARTS D4D Module
2 Methodology
2.1 SMARTS D4D Module Development
2.2 Analysis of Existing Application System
2.3 Modules of Dual-Application Systems
2.4 Content and Face Validation Testing
3 Module Design, Development and Findings
3.1 Technology Deployed and System Architecture
3.2 Dual-System Implementation
3.3 Needs, Content and Face Validity Testing
4 Conclusion
References
Improved Multi-label Medical Text Classification Using Features Cooperation
1 Introduction
2 Preliminaries
2.1 Feature Extraction
2.2 Doc2Vec
2.3 Multi-label Classification
2.4 Multi-label Learning Approaches
3 Methodology
3.1 Medical Text Preprocessing
3.2 Feature Extraction
3.3 Multi-label Classification Stage
4 Experimental Results and Discussion
4.1 Used Data
4.2 Evaluation
5 Discussion
6 Conclusion
References
Image Modeling Through Augmented Reality for Skin Allergies Recognition
1 Introduction
1.1 Related Work
1.2 Proposed System
2 System Analysis, Design, and Implementation
3 System Testing and Evaluation
4 System Interface Design
5 Conclusion
References
Hybridisation of Optimised Support Vector Machine and Artificial Neural Network for Diabetic Retinopathy Classification
1 Introduction
2 Related Works
2.1 Support Vector Machine
2.2 Artificial Neural Network
2.3 Hybrid Machine Learning Algorithm
3 Proposed Work
3.1 Data from the Electronic Health Record
3.2 Performance Evaluation
3.3 The Overall Flow Architecture of Hybrid Optimised Support Vector Machine and Artificial Neural Network
4 Results
4.1 Comparison of Results Between SVM-NN, Optimised SVM, and Non-optimised SVM algorithms
5 Discussion
6 Conclusions
References
A Habit-Change Support Web-Based System with Big Data Analytical Features for Hospitals (Doctive)
1 Introduction
2 Related Work
2.1 Existing Systems
3 System Requirements and Analysis
3.1 Proposed Solutions
3.2 Development Methodology
3.3 System Architecture
4 Test and Evaluation of the System
4.1 Result
5 Conclusion and Future Work
References
An Architecture for Intelligent Diagnosing Diabetic Types and Complications Based on Symptoms
1 Introduction
2 Review of Literature
3 Research Methods
3.1 Back-Propagation Neural Networks
3.2 Data Clustering Algorithms
3.3 Fuzzy Logic Techniques
4 Result Analysis and Discussion
5 Conclusion
References
An Advanced Encryption Cryptographically-Based Securing Applicative Protocols MQTT and CoAP to Optimize Medical-IOT Supervising Platforms
1 Introduction
2 General Architecture
3 Methodology
3.1 Protocols Application IOT
4 Related Works
4.1 Comparative Study for the Proposed Protocols IOT
5 The Proposed Approach
6 Discussion
7 Conclusion and Perspectives
References
Pulmonary Nodule Classification Based on Three Convolutional Neural Networks Models
1 Introduction
2 Proposed Method
3 Experiments
3.1 Dataset and Preprocessing
3.2 Experiment Settings
3.3 Experiment Results
4 Conclusion
References
A Comparative Study on Liver Tumor Detection Using CT Images
1 Introduction
2 Related Works
3 Methodology
3.1 Data Collection
3.2 Data Augmentation
3.3 Machine Learning Techniques
3.4 Deep Learning Techniques
3.5 Transfer Learning
3.6 Deep Learning Pre-trained Model
4 Results and Discussion
5 Conclusion and Future Work
References
Brain Tumor Diagnosis System Based on RM Images: A Comparative Study
1 Introduction
2 Related Works
3 Methods
3.1 Dataset Collection Phase
3.2 Data Selection
3.3 Covert 3D MRI to JPEG Images
3.4 Deep Learning Method
3.5 Machine Learning Approach
3.6 Deep Learning Approach
3.7 Transfer Learning
3.8 Experimental Setup
4 Results and Discussion
5 Conclusion and Recommendations
References
Diagnosis of COVID-19 Disease Using Convolutional Neural Network Models Based Transfer Learning
1 Introduction
2 Related Works
3 Materials and Methods
3.1 Dataset Assembling
3.2 Proposed CNN Based Methodology
4 Results and Discussion
4.1 Training Results
4.2 Testing Results
4.3 Comparison with Related Works
5 Conclusion and Perspectives
References
Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques
1 Introduction
2 Literature Review
3 Methodology
3.1 Data Set
3.2 Proposed Model
3.3 Dimensionality Reduction
3.4 Machine Learning Algorithms
4 Results and Discussion
5 Epilogue
References
Detection of Cardiovascular Disease Using Ensemble Machine Learning Techniques
1 Introduction
2 Related Work
3 Methodology
3.1 Dataset
3.2 Classification Algorithms
3.3 Evaluation Metrics
4 Results and Evaluation
5 Conclusion
References
Health Information Management
Hospital Information System for Motivating Patient Loyalty: A Systematic Literature Review
1 Introduction
2 Methods
2.1 Literature Search
2.2 Inclusion and Exclusion Criteria
2.3 Data Extraction
3 Results
3.1 Search Results
4 Discussion
4.1 Prevalence of HIS Tasked with Motivating Patient Loyalty
4.2 Efficacy of HIS Tasked with Motivating Patient Loyalty
4.3 Gaps in the Literature
4.4 Limitations
5 Conclusions and Directions for Future Research
References
Context Ontology for Smart Healthcare Systems
1 Introduction
2 Related Work
3 Smart Healthcare Context Properties Analysis
4 Context Ontology for Mobile Service
5 The Evaluation of the Proposed Context Ontology
6 Conclusion and Future Work
References
A Modified UTAUT Model for Hospital Information Systems Geared Towards Motivating Patient Loyalty
1 Introduction
2 Literature Review
2.1 Capability
2.2 Configurability
2.3 Ease of Use/Help Desk Availability and Competence (EU)
2.4 Accessibility/Shareability
3 Modified UTAUT Model for HISPL
4 Conclusions and Directions for Future Work
References
Teamwork Communication in Healthcare: An Instrument (Questionnaire) Validation Process
1 Introduction
2 Literature Review
3 Model
4 Discussion
5 Conclusion
6 Appendix A
References
Potential Benefits of Social Media to Healthcare: A Systematic Literature Review
1 Introduction
2 Methodology
3 Results
3.1 User Classification
4 Discussion
5 Conclusion
References
Exploring the Influence of Human-Centered Design on User Experience in Health Informatics Sector: A Systematic Review
1 Introduction
2 Previous Works
3 Method
3.1 Research Questions
3.2 Search Strategy
3.3 Inclusion and Exclusion Criteria
3.4 Quality Assessment and Data Extraction
3.5 Data Analysis
4 Result
4.1 Year Published
4.2 Methodology
4.3 Publication Sources
4.4 Context of Study
4.5 Discussion of SLR
5 Discussion
6 Conclusion
References
An Emotional-Persuasive Habit-Change Support Mobile Application for Heart Disease Patients (BeHabit)
1 Introduction
1.1 Background
2 Proposed Solution
2.1 System Design
3 System Development Methodology
3.1 System Implementation
4 Testing and Evaluation
4.1 Unit Testing
4.2 Integration Testing
4.3 System Testing and User Acceptance Evaluation
5 Conclusion and Future Work
References
A Systematic Review of the Integration of Motivational and Behavioural Theories in Game-Based Health Interventions
1 Introduction
2 Integration of Theories
3 Review Procedure
3.1 Search Strategy
3.2 Data Analysis
4 Analysis
4.1 What Theories Were Targeted?
4.2 What Behaviours Were Targeted?
4.3 What Groups Were Targeted?
5 Discussion
6 Conclusion
References
Adopting React Personal Health Record (PHR) System in Yemen HealthCare Institutions
1 Introduction
2 Related Works
3 Advantages of Using and Improving the PHRs Systems
4 Methodology and System Design
4.1 Determining Level of Interest
4.2 System Technology Selection
4.3 System Framework and Users’ Interface
5 Conclusion
References
Artificial Intelligence and Soft Computing
Application of Shuffled Frog-Leaping Algorithm for Optimal Software Project Scheduling and Staffing
1 Introduction
1.1 The SFLA Algorithm
2 Related Works
3 Methodology
3.1 Frog Representation
3.2 SFLA Design
3.3 Fitness Function
4 Results and Discussion
5 Conclusion
References
A Long Short Term Memory and a Discrete Wavelet Transform to Predict the Stock Price
1 Introduction
2 Literature Survey
3 Prediction Using DWT and LSTM
3.1 Discrete Wavelet Transform (DWT)
3.2 Long Short-Term Memory
4 Experimental Setup
4.1 Data Description
4.2 Prediction Procedure
5 Results
6 Conclusions and Future Work
References
Effective Web Service Classification Using a Hybrid of Ontology Generation and Machine Learning Algorithm
1 Introduction
2 Related Work
3 Proposed Classification Approach
3.1 Dataset Collection
3.2 Data Pre-processing
3.3 Designing Context Ontology
3.4 A Hybrid Approach of Ontology and SVMs Classifier
3.5 A Hybrid Approach of Ontology and NB Classifier
3.6 Training and Testing
3.7 Evaluation Metrics
4 Result Analysis
5 Future Work
References
Binary Cuckoo Optimisation Algorithm and Information Theory for Filter-Based Feature Selection
1 Introduction
2 Background
2.1 Binary Cuckoo Optimisation Algorithm
2.2 Information Gain Based Entropy
2.3 Mutual Information
2.4 Some Related Works
3 Proposed Approaches
3.1 BCOA Based MI for FS
3.2 BCOA Based Information Gain Entropy for FS
3.3 Relevance and Redundancy Weighted Values in BCOA-MI and βCOA-E
3.4 Experimental Design
4 Discussion of Results
4.1 Results of BCOA-MI and BCOA-E
4.2 Results of BCOA-MI and BCOA-E with ΒWeighted Values
4.3 Average Fitness of BCOA-MI and BCOA-E
4.4 Convergence Trends of BCOA-MI and BCOA-E
4.5 Comparison with Other Existing Approaches
5 Conclusions and Future Works
References
Optimized Text Classification Using Correlated Based Improved Genetic Algorithm
1 Introduction
1.1 Text Classification
1.2 Genetic Optimization
2 Related Works
3 Proposed Approach
3.1 Proposed IGA
3.2 IGA Based Text Classification
4 Experimental Environment
5 Results and Discussions
6 Conclusions
References
Multi-objective NPO Minimizing the Total Cost and CO2 Emissions for a Stand-Alone Hybrid Energy System
1 Introduction
2 The Problem of Optimal Sizing
2.1 Construction of the Proposed HES
2.2 Data Collection
2.3 Modeling of the Proposed HES
3 Nomadic People Optimizer (NPO)
3.1 The Objective of the Proposed Algorithm
4 Total System Cost CT
5 Total CO2 Emissions (ECO2T)
5.1 The Constraints
6 Results and Discussions
6.1 Optimal Configuration
7 Conclusions
References
A Real Time Flood Detection System Based on Machine Learning Algorithms
1 Introduction
2 Background and Related Work
2.1 Artificial-Neural-Networks (ANNs)
2.2 Adaptive-Neuro-Fuzzy Inference System (ANFIS)
2.3 Decision Tree (DT) and Ensemble-Prediction-Systems (EPSs)
3 Proposed Methodology
3.1 Experimental Design
4 Results and Discussion
5 Conclusions
References
Extracting Semantic Concepts and Relations from Scientific Publications by Using Deep Learning
1 Introduction
2 Ontology Construction Challenges
3 Deep Belief Network
4 Methodology
4.1 Concept Relevance Measurements
4.2 Semantic Relations
4.3 Extracting Semantic Concepts and Relations
5 Conclusion
References
Effectiveness of Convolutional Neural Network Models in Classifying Agricultural Threats
1 Introduction
2 Related Work
3 The Architecture of Convolutional Neural Network (CNN)
3.1 Convnet Layers
3.2 Activation Functions
3.3 Regularization
4 Research Methodology
4.1 Dataset Collection
4.2 Data Pre-processing
4.3 Applying the Convolutional Neural Network (CNN) Models
4.4 Training and Testing
5 Result Analysis
6 Future Work
References
A Study on Emotion Identification from Music Lyrics
1 Introduction
2 Emotion Models
3 Affective Lexicon
4 Features from Music Lyrics
5 Lyrics Datasets for Emotion Identification
6 A Review of Emotion Identification from Music Lyrics
7 Discussion
8 Conclusion
References
A Deep Neural Network Model with Multihop Self-attention Mechanism for Topic Segmentation of Texts
1 Introduction
2 Related Works
3 Model Architecture
3.1 Input Representation Layer
3.2 Bidirectional LSTM (BiLSTM) Layer
3.3 Highway Network Layer
3.4 Multihop Self-attention Layer
3.5 Output and Training
4 Experiments
4.1 Datasets
4.2 Evaluation Metrics
4.3 Baselines
4.4 Preprocessing
4.5 Training and Parameters
5 Results and Discussion
5.1 Comparison with the BayesSeg Model
5.2 Comparison with Att-CNN-BiLSTM Model
6 Conclusion
References
Data Science and Big Data Analytics
Big Data Interoperability Framework for Malaysian Public Open Data
1 Introduction
2 The Proposed Big Data Interoperability Framework
3 Prototype
4 Conclusion
References
The Digital Resources Objects Retrieval: Concepts and Figures
1 Introduction
2 Digital Cultural Heritage Collections
3 Evaluation Performance Measurement in DRO Retrievals
4 Statistical Test in DROs
5 Findings
6 Summary and Conclusion
References
A Review of Graph-Based Extractive Text Summarization Models
1 Introduction
2 Extractive Text Summarization
3 Graph-Based ATS Models
3.1 Static Graph-Based Model
3.2 Dynamic Graph-Based Model
3.3 Graph Pruning-Based Model
3.4 Hypergraph-Based Model
3.5 Affinity Graph-Based Model
3.6 Semantic Graph-Based Model
3.7 Multigraph-Based Model
4 Discussion
5 Conclusion
References
Review on Emotion Recognition Using EEG Signals Based on Brain-Computer Interface System
1 Introduction
2 Brain-Computer Interface (BCI)
2.1 Medical Field
2.2 Entertainment Field
3 EEG Signals Acquisition
4 Feature Extraction
5 Classification Methods
6 Discussion of the Results of Emotion Recognition Methods Using EEG Signals
6.1 Technology Challenges
6.2 User Challenges
6.3 Discussion
7 Conclusions
References
A New Multi-resource Deadlock Detection Algorithm Using Directed Graph Requests in Distributed Database Systems
1 Introduction
2 Related Work
3 The Proposed Algorithm: Multi Resource Deadlock Detection (MRDD)
4 Analytical Model Analysis
4.1 Multi Resource Deadlock Detection in Two Sites with Eight Transactions
5 Performance Analysis
6 Conclusion
References
Big Data Analytics Model for Preventing the Spread of COVID-19 During Hajj Using the Proposed Smart Hajj Application
1 Introduction
2 Related Works
3 Methodology
3.1 First Step (Research Question)
3.2 Second Step (Data Collection)
3.3 Third Step (Investigation)
3.4 Fourth Step (Research)
3.5 Fifth Step (Results/Knowledge)
4 System Architecture
5 Conclusion
References
Financial Time Series Forecasting Using Prophet
1 Introduction
2 Literature Review and Current Practices
3 Prophet Model and Research Methodology
4 Experimental Results
4.1 Data Set and Evaluation Measure
4.2 Results
5 Conclusion
References
Facial Recognition to Identify Emotions: An Application of Deep Learning
1 Introduction
2 Background and Related Works
3 Proposed Recognition Approach
3.1 The Convolution Step
3.2 The Pooling Step
3.3 The Flattening
3.4 Optimization for Deep Learning
4 Test and Discussion
4.1 Test Data
4.2 Results
4.3 Evaluation Measure
4.4 Discussion
5 Conclusion
References
Text-Based Analysis to Detect Figure Plagiarism
1 Introduction
2 Methods
2.1 Similarity Score Based Text Comparison
3 Experimental Design and Dataset
4 Results and Evaluation
5 Conclusions and Future Work
References
A Virtual Exploration of al-Masjid al-Nabawi Using Leap Motion Controller
1 Introduction
2 Related Works
3 Methodology
3.1 Building the 3D Space
3.2 Interaction Technique
4 Results and Discussion
5 Conclusion and Future Work
References
Comparison of Data Analytic Techniques for a Spatial Opinion Mining in Literary Works: A Review Paper
1 Introduction
2 Methods for a Spatial Opinion Mining in Literary Works
2.1 Natural Language Processing and Text Analytics
2.2 Opinion Mining
2.3 Topic Modelling
2.4 Sentiment Analysis (SA)
3 Comparison and Discussion Between Sentiment Analysis and Topic Modelling Approaches
3.1 Comparison and Discussion on Different Sentiment Analysis Approaches
3.2 Comparison of Topic Modelling: Latent Dirichlet Allocation (LDA) and Latent Semantic Analysis (LSA)
4 Conclusion
References
Open Data in Prediction Using Machine Learning: A Systematic Review
1 Introduction
1.1 Background
1.2 Problem Description
2 Methodology
2.1 Research Questions (RQs)
2.2 Data Collection
2.3 Results Included
2.4 Data Extraction
2.5 Data Extraction
3 Results
3.1 Bibliometric Research Questions (BRQs)
3.2 Content Research Questions (CRQs)
4 Discussion
5 Conclusion and Research Opportunities
References
Big Data Analytics Based Model for Red Chili Agriculture in Indonesia
1 Introduction
1.1 Background
2 Methods
2.1 Review on Big Data Analytics
2.2 Proposed BDA Model.
3 Conclusion
3.1 Research Limitations and Future Research.
References
A Fusion-Based Feature Selection Framework for Microarray Data Classification
1 Introduction
2 Related Studies
3 Methods
3.1 Fusion-Based Feature Selection
3.2 Classification Phase
4 Experimental Design
4.1 Datasets
4.2 The Implemented Scenarios of the Proposed Framework
4.3 Performance Evaluation Measure
5 Results and Discussion
6 Conclusion
References
An Approach Based Natural Language Processing for DNA Sequences Encoding Using the Global Vectors for Word Representation
1 Introduction
2 Related Work
3 Method
3.1 Build the Corpus
3.2 Train the Embedding Model
3.3 Transform DNA Fragments into Numerical Vector Using Trained Embedding Model.
4 Datasets
5 Results and Discussion
6 Conclusion
References
Short-Term CO2 Emissions Forecasting Using Multi-variable Grey Model and Artificial Bee Colony (ABC) Algorithm Approach
1 Introduction
2 Grey Relational Analysis Methods
2.1 Multi-variable Grey Model GM(1,N).
2.2 ABC Algorithm Based GM(1,N) Model.
3 Results and Discussion
4 Conclusion
References
IoT and Intelligent Communication Systems
A Reliable Single Prediction Data Reduction Approach for WSNs Based on Kalman Filter
1 Introduction
2 Proposed Approach
2.1 Data Reduction Phase (DR)
2.2 Data Prediction Phase (DR)
3 Implementation and Results
3.1 Datasets
3.2 Results and Analysis
4 Conclusion and Future Work
References
A Real-Time Groundwater Level Monitoring System Based on WSN, Taiz, Yemen
1 Introduction
2 Literature Work
3 The Proposed Groundwater Framework
3.1 Case Study
3.2 Methodology
4 Results and Discussions
5 Conclusion
References
Design and Simulation of Multiband Circular Microstrip Patch Antenna with CSRR for WLAN and WiMAX Applications
1 Introduction
2 Antenna Design Methodology
2.1 Development of the Initial Antenna
2.2 Improving the Antenna Characteristics
2.3 Parametric Analysis
3 Simulation Results and Discussion
4 Conclusion
References
Reference Architectures for the IoT: A Survey
1 Introduction
2 Background
3 Reference Architectures for IoTs
3.1 IoT Reference Architecture by Guth et al. [15]
3.2 Industrial Internet Reference Architecture (IIRA) [24]
3.3 WSO2’s Reference Architecture [1]
3.4 Internet of Things Architecture (IoT-A) [4, 8, 10, 16, 17, 24, 26]
3.5 IoT Architecture Reference Model (ARM) [3, 7]
4 Analysis of the Surveyed IoT RAs
5 Related Work
6 Conclusions
References
A Circular Multiband Microstrip Patch Antenna with DGS for WLAN/WiMAX/Bluetooth/UMTS/LTE
1 Introduction
2 The Proposed Multiband Microstrip Patch Antenna
2.1 Developing the Reference Antenna
2.2 The Optimized Antenna Design
3 Simulations and Results
3.1 Evaluating the Performance of the Optimized Antenna
3.2 Parametric Analysis
4 Conclusions
References
Anomaly Intrusion Detection Systems in IoT Using Deep Learning Techniques: A Survey
1 Introduction
2 Literature Review
3 Discussion
4 Techniques Used
5 Dataset Used
6 Limitations of the Existing Studies
7 Recommendation
8 Conclusion
References
Security and Threats in the Internet of Things Based Smart Home
1 Introduction
2 Smart Home Environment
2.1 Elements of Smart Homes
2.2 Internet of Things (IoT) Architecture
3 Security Threats in the IoT
3.1 Common Threats
4 Security Measures for Smart Home
4.1 Security Components
4.2 Security Attacks and Solutions
5 Derive Security Threats Countermeasure Frameworks
6 Conclusion
References
Simulation and Control of Industrial Composition Process Over Wired and Wireless Networks
1 Introduction
2 Methodology and Design Implementation
2.1 Controller Design
2.2 TrueTime Simulator
2.3 Topology of WNCS
2.4 Controller Design in TrueTime
3 Results and Discussion
3.1 Simulation of Wired Control System
3.2 Wired vs Wireless
4 Conclusion
References
Performance Degradation of Multi-class Classification Model Due to Continuous Evolving Data Streams
1 Introduction
2 Related Work
3 Research Methodology
3.1 Tentative Solution- Overview
4 Conclusion and Future Consideration
References
Compact Wide-Bandwidth Microstrip Antenna for Millimeter Wave Applications
1 Introduction
2 Antenna Design Procedure
2.1 Design Guidelines
3 Simulation Results and Discussions
3.1 Simulated Return Loss and Bandwidth
3.2 Simulated Voltage Standing Wave Ratio (VSWR)
3.3 Simulated 3D Gain
4 Conclusions
References
Dual-Band Rectangular Microstrip Patch Antenna with CSRR for 28/38 GHz Bands Applications
1 Introduction
2 Antenna Design Methodology
2.1 The Reference Antenna Design
3 Simulation Results and Discussion
3.1 Performance of the Optimized Antenna
3.2 Parametric Analysis
3.3 Performance Comparison with Previous Published Work
4 Conclusion
References
Dual Band Rectangular Microstrip Patch Antenna for 5G Millimeter-Wave Wireless Access and Backhaul Applications
1 Introduction
2 Antenna Design and Configuration
3 Parametric Analysis and Design Optimization
4 Results and Discussions
5 Conclusions
References
Design of Wireless Local Multimedia Communication Network (WLMmCN) Based on Android Application Without Internet Connection
1 Introduction
2 Technical Background
3 WLMmCN Architecture
4 Results
5 Conclusions
References
A Statistical Channel Propagation Analysis for 5G mmWave at 73 GHz in Urban Microcell
1 Introduction
2 Channel Propagation Models
3 MIMO-OFDM Communication Scenarios
4 The Key Parameters
5 Results, Analysis and Discussion
6 Conclusion
References
Advances in Information Security
Robot Networks and Their Impact on Cyber Security and Protection from Attacks: A Review
1 Introduction
2 Related Works
3 Research Methodology
4 Importance of Cyber Security
5 Botnets Architectures
5.1 Centralized Architecture
5.2 Peer-To-Peer Architecture
5.3 Hybrid Architecture
6 How Botnets Attacks Works
7 Threats Represented by Botnets
8 Preventing Botnets Attacks
9 Improving Organizations and Individual Cyber Security
10 The Finding
11 Conclusion
References
An Efficient Fog-Based Attack Detection Using Ensemble of MOA-WMA for Internet of Medical Things
1 Introduction
2 Machine Learning Attack Detection in Security of IoMT
3 Methodology
3.1 Dataset and Methods
3.2 Research Design
3.3 Evaluation Metrics
4 Results and Discussion
5 Conclusion
References
A New DNA Based Encryption Algorithm for Internet of Things
1 Introduction
2 Background and Related Work
2.1 IoT and Encryption
2.2 Related Work
3 Proposed DNA Encryption Algorithm
4 Experimental Result and Discussion
5 Conclusion
References
Watermarking Techniques for Mobile Application: A Review
1 Introduction
2 Related Works
3 Watermarking for Mobile Applications
3.1 Mobile Architecture
4 Methodology
5 Discussion
6 Conclusion
References
Analysis and Evaluation of Template Based Methods Against Geometric Attacks: A Survey
1 Introduction
2 Literature Review
2.1 Template Based Methods in Transform Domains
2.2 Template Based Methods in Spatial Domains
3 Discussion
4 Conclusion
References
Survey of File Carving Techniques
1 Introduction
2 Traditional/Basic Carving Algorithm
2.1 Header/Footer Carving or Signature-Based Carving
2.2 Header/Maximum Size Carving’s Algorithm
2.3 Header/Embedded Length Carving
2.4 File-Structure Based Carving
3 Advance Carving Algorithm
3.1 Bifragment Gap Carving Algorithm
3.2 Aho-Corasick Algorithm
3.3 Hash-Based Carving Algorithm
3.4 Decision-Theoretic Algorithm (DECA)
3.5 Orphaned Fragment Carving Algorithm
4 Conclusion
References
Affecting Factors in Information Security Policy Compliance: Combine Organisational Factors and User Habits
1 Introduction
2 Literature Review
2.1 Organisation Commitment
2.2 Organisation Culture
2.3 Reward
2.4 Habit
3 Research Methodology
4 Research Finding
5 Discussion
6 Conclusion
References
Mitigation of Data Security Threats in Iraqi Dam Management Systems: A Case Study of Fallujah Dam Management System
1 Introduction
2 Background
2.1 Related Works
2.2 Case Study: Fallujah Dam
2.3 Security Issues
3 Method and Materials
3.1 Design of Dam Management System
3.2 Multi-Tier Secure Model
4 Findings and Discussion
5 Conclusions and Future Work
References
Advances in Information Systems
Development and Validation of a Classified Information Assurance Scale for Institutions of Higher Learning
1 Introduction
2 Related Work
3 Methodology
3.1 Selection of Experts
3.2 Invitation of Participants and Distribution of Survey Forms
3.3 Analysis of Responses
3.4 Revision and Final Scale
4 Conclusion
5 Limitations
References
Sustainable e-Learning Framework: Expert Views
1 Introduction
2 The Successful Development of Sustainable e-learning
2.1 Sustainable e-learning Framework (SeLF)
3 Methodology
4 Results and Discussion
4.1 SeLF’s Contribution Towards Sustainability Approach
4.2 Sustainable e-learning Approach
4.3 Higher Education Practices
5 Conclusion
References
Derivation of a Customer Loyalty Factors Based on Customers’ Changing Habits in E-Commerce Platform
1 Introduction
2 Background of the Problem
3 Literature Review
3.1 E-commerce
3.2 Customer Service in E-commerce
3.3 Business-to-Consumer (B2C)
3.4 Customer Satisfaction in Customer Loyalty
3.5 Loyalty in E-commerce (E-loyalty)
3.6 Customers Changing Habits
4 Methodology
5 Model Development
5.1 Derivation of Factors
6 Conclusion
References
Analysis of Multimedia Elements Criteria Using AHP Method
1 Introduction
2 Literature Review
3 Methodology
3.1 Problem and Objective
3.2 Structure Elements into Criteria, Sub-criteria and Alternatives
3.3 The Pairwise Comparison Process
3.4 Calculate Weighting and Consistency Ratio (CR) Determination for the Interaction Element Criteria
3.5 Evaluate Alternatives According to Weighting and Get Ranking
4 Result and Finding
5 Conclusion
References
The Development of a Criteria-Based Group Formation Systems for Student Group Work
1 Introduction
2 Literature Review
3 Methodology
4 Prototype Development
5 User Acceptance
6 Conclusion
References
Trusted Factors of Social Commerce Product Review Video
1 Introduction
2 Systematic Literature Review (SLR)
2.1 Purpose of the Literature Review
2.2 Practical and Training
2.3 Searching for the Literature
2.4 Practical Screen
2.5 Quality of Appraisal
2.6 Data Extraction
3 SLR Findings
3.1 Product Review Video
3.2 Online Video Sharing Platforms
3.3 Trusted Factors of Product Review Video
4 Conclusion
References
Building Information Modelling Adoption: Systematic Literature Review
1 Introduction
2 Literature Review
3 Methodology
4 Findings and Analysis
4.1 Technology Adoption Theories in BIM Adoption Research
4.2 Independent Variables in BIM Adoption Studies
4.3 Dependent Variables in BIM Adoption Studies
5 Conclusion
References
Adoption of Smart Cities Models in Developing Countries: Focusing in Strategy and Design in Sudan
1 Introduction
2 Literature Review
3 A Guide for Applying the Smart City Model
3.1 Organizational Structure for the Smart City
3.2 Vision and Goals
3.3 Evaluation of E-readiness
3.4 Documenting Operations
3.5 Redesigned Processes
3.6 Executive Plan
3.7 Management of Change
4 Framework Analysis
5 Conclusion
References
Factors Affecting Customer Acceptance of Online Shopping Platforms in Malaysia: Conceptual Model and Preliminary Results
1 Introduction
2 Theoretical and Conceptual Background
2.1 Technology Acceptance Model
2.2 Related Factors on Customer Online Shopping Acceptance
3 Conceptual Model Constructs
3.1 Website Content
3.2 Website Design
3.3 Perceived Enjoyment
3.4 Perceived Ease of Use
3.5 Perceived Usefulness
3.6 Trust
3.7 Customer Service Quality
4 Methodology
5 Results and Data Analysis
5.1 Demographic Information
5.2 Measurement Model
6 Conclusion
References
Student Compliance Intention Model for Continued Usage of E-Learning in University
1 Introduction
2 Research Problem
3 Literature Review
4 Research Methodology
5 Proposed Model
6 Conclusion
References
Digital Information and Communication Overload Among Youths in Malaysia: A Preliminary Review
1 Introduction
2 Literature Review
2.1 Digital Information and Communication
2.2 The Impact of Information and Communication Overload
2.3 Information and Communication in Modern Day
3 Methodology
4 Discussion
5 Conclusion
References
The Effect of Using Social Networking Sites on Undergraduate Students’ Perception and Academic Performance at University of Taiz-Yemen
1 Introduction
1.1 Problem Statement
1.2 Research Objectives
1.3 Research Questions
2 Literature Review
3 Research Methodology
3.1 Measurement Variables
3.2 Data Analysis
4 Findings
4.1 Profile of Respondents
4.2 Most Popular SNSs for Academic Purposes
4.3 Student Use and Academic Performance
5 Conclusion
References
Building Information Modelling Adoption Model for Malaysian Architecture, Engineering and Construction Industry
1 Introduction
2 Literature Review
3 Methodology and Proposed Model
4 Findings and Analysis
4.1 Demographic Analysis
4.2 Measurement Model
4.3 Hypothesis Testing
5 Discussion and Conclusions
References
Digital Government Competency for Omani Public Sector Managers: A Conceptual Framework
1 Introduction
2 Related Work
2.1 Digital Government and Competencies
2.2 Related Theories on Digital Competency
2.3 The Gap in the Existing DGC Study
3 Methodology
3.1 Research Design
3.2 Data Collection
3.3 Data Analysis
4 Research Conceptual Framework
4.1 Management Competency
4.2 Soft Skills Competency
4.3 Digital Literacy
4.4 Digital Creativity and Innovation
4.5 Information Security Competency
4.6 Digital Leadership Competency
5 Conclusion and Future Work
References
Computational Vision and Robotics
Landmark Localization in Occluded Faces Using Deep Learning Approach
1 Introduction
2 Related Studies
2.1 Models for Facial Landmark Localization
2.2 Dataset for Landmark Detection
3 Methodology and Implementation
4 Experimental Results
5 Conclusion and Future Work
References
Contrast Image Quality Assessment Algorithm Based on Probability Density Functions Features
1 Introduction
2 Optimization of Contrast Features in NR-IQA-CDI
2.1 Addressing the Problems of the Existing NR-IQA-CDI
2.2 Application of Monotonic PDF of Contrast Feature in NR-IQA-CDI
3 Experimental Results and Discussions
3.1 Evaluation Methodology
3.2 NR-IQA-CDI-MPCF Evaluation
3.3 Statistical Performance Analysis
4 Conclusion
References
The Impact of Data Augmentation on Accuracy of COVID-19 Detection Based on X-ray Images
1 Introduction
2 Related Work
3 Methodology and Dataset
3.1 Dataset
3.2 Convolutional Neural Networks
3.3 Data Augmentation
3.4 Proposed Model
4 Results
5 Discussion
6 Conclusion
References
A Fusion Schema of Hand-Crafted Feature and Feature Learning for Kinship Verification
1 Introduction
2 Related Works
3 Proposed Method
3.1 Hand-Crafted: Histogram of Oriented Gradients (HOG)
3.2 Feature Learning: Convolutional Neural Network (CNN)
3.3 Fusion Hand-Crafted Feature and Feature Learning
3.4 Features Subtracting Absolute Value
4 Experiments and Evaluation
4.1 KinFaceW-I and KinFaceW-II Databases
4.2 Experimental Setting
5 Results and Discussion
5.1 Experimental Results on KinFaceW-I and KinFaceW-II Databases
6 Conclusion and Future Works
References
Lossless Audio Steganographic Method Using Companding Technique
1 Introduction
2 Materials and Methods
2.1 Integer Wavelet Transform
2.2 Companding Techniques
2.3 Fredkin Gate
2.4 The Proposed Method
2.5 Performance Evaluation Metrics
3 Results and Discussion
4 Conclusions
References
Smart Traffic Light System Design Based on Single Shot MultiBox Detector (SSD) and Anylogic Simulation
1 Introduction
2 Methods
3 Results and Discussion
3.1 Design
3.2 Trial
3.3 Data Processing
3.4 Analysis
4 Conclusion
References
Learning Scope of Python Coding Using Immersive Virtual Reality
1 Introduction
2 Materials and Methods
2.1 Architecture of VR Learning Apps
2.2 Types of Interaction Techniques Used
2.3 Implementation Details
2.4 Experiments and Evaluation
3 Results and Discussion
3.1 Experiments I
3.2 Experiments II
4 Conclusion
References
Automatic Audio Replacement of Objectionable Content for Sri Lankan Locale
1 Introduction
2 Related Work
2.1 Automatic Objectionable Content Detection and Replacement in Audio
2.2 Automatic Objectionable Content Detection in Text
2.3 Libraries and Other Implementations
3 System Overview
3.1 DSP Module
3.2 Speech Recognition Module
3.3 NLP Module
3.4 Audio Replacing Module
4 Methodology
4.1 Data Corpus Collection
4.2 Preprocessing of Data
4.3 Building and Evaluating the Model
5 Evaluation
6 Conclusion
References
A Comparison of CNN and Conventional Descriptors for Word Spotting Approach: Application to Handwritten Document Image Retrieval
1 Introduction
2 Related Work
2.1 Supervised-Learning-Free Word-Spotting Techniques
2.2 Supervised-Learning Techniques
3 Methodology
3.1 Interest Points-Based Approach
3.2 CNN Based Approach
3.3 Handwritten Dataset
4 Result and Discussions
4.1 Results
4.2 Discussion
5 Conclusion
References
Handwritten Arabic Character Recognition: Comparison of Conventional Machine Learning and Deep Learning Approaches
1 Introduction
2 Overview of the Arabic Handwriting Recognition Systems
3 Proposed Methods
3.1 Method Based on Conventional Machine Learning
3.2 Method Based on Deep Learning Using CNN Models
4 Evaluation and Results
5 Conclusion
References
Document Image Edge Detection Based on a Local Hysteresis Thresholding and Automatic Setting Using PSO
1 Introduction
1.1 Overview of Edge Detection
2 Proposed Approach
2.1 Parameter Optimization Using PSO
2.2 Our Proposed Version of Pratt’s Formula
2.3 Toward an Unsupervised Segmentation Method Using an Adaptive Thresholding
3 Experimentation and Results
3.1 Evaluation of Conventional Edge Detectors Genericity
3.2 Evaluation of Our Improved Version of Pratt’s Metric
3.3 Our Proposed Method Evaluation
4 Discussion and Further Analysis
5 Conclusion
References
Fast I2SDBSCAN Based on Integral Volume of 3D Histogram: Application to Color Layer Separation in Document Images
1 Introduction
2 Existing Segmentation Methods
2.1 Segmentation Method Based on Edge/Region Detection or Binarization
2.2 Segmentation Based Elements Classification (or Clustering)
3 Our Pixel-Classification and Colorimetric Segmentation Approach
3.1 Steps of Our Approach
3.2 Our New Method of Fast Calculation of the Color Density
4 Evaluation and Results
4.1 Discussion
5 Conclusion
References
Enhancing Daily Life Skills Learning for Children with ASD Through Augmented Reality
1 Introduction
2 Related Works
3 Augmented Reality System-Overview
3.1 Analysis
3.2 Design
3.3 Prototype Development
4 Testing
5 Discussions
6 Future Work
References
Recent Computing and Software Engineering
SpaceScience App: Development of a Mobile Application for School Children
1 Introduction
2 Literature Review
2.1 Mobile Devices
2.2 Development of Mobile Applications
2.3 Engagement in Gaming
2.4 Motivation in Gaming
3 Methodology
4 Results
4.1 Interview
4.2 Flowchart of the Project
4.3 SpaceScience Mobile Application
4.4 User Experience Testing
5 Conclusion
References
Research on Online Problem-Based Learning Among Undergraduate Students: A  Systematic Review
1 Introduction
2 Method
2.1 Search Strategy
2.2 Inclusion and Exclusion Criteria
3 Results
3.1 Problem Based Learning Principles
3.2 Effects of Problem Based Learning
4 Final Remarks
5 Conclusion and Future Research
References
Derivation of Factors in Dealing Negative E-WOM for Maintaining Online Reputation
1 Introduction
2 Background of the Problem
3 Literature Review
3.1 Online Reputation
3.2 Maintaining Online Reputation
3.3 Negative E-WOM
3.4 Factors Dealing Negative E-WOM
3.5 Dealing Negative E-WOM Through CRM and Positive E-WOM in Maintaining Online Reputation
4 Methodology
5 Model in Dealing Negative E-WOM for Maintaining Online Reputation
5.1 Constructs
5.2 Model Development
6 Conclusion
References
A Terms Interrelationship Approach to Query Expansion Based on Terms Selection
1 Introduction
2 Related Works
3 Proposed Approach
3.1 Query and Text Preprocessing
3.2 Query Expansion
3.3 Documents Text Preprocessing and Indexing
4 Experiment Results and Discussion
4.1 Experiment Results
4.2 Discussion
5 Conclusion and Future Work
References
Multi-domain Business Intelligence Model for Educational Enterprise Resource Planning Systems
1 Introduction
2 Literature Review
2.1 Business Intelligence Models
2.2 Business Intelligence Tools and Applications
2.3 Enterprise Resource Planning Business Intelligence (ERPBI)
3 The Proposed Model
3.1 The Proposed BI Domains for Educational ERP System
3.2 Business Intelligence Steps
3.3 Data Warehouse
3.4 Business Intelligence Application
4 Discussion and Conclusion
References
Measuring Risk Mitigation Techniques in Agile Global Software Development
1 Introduction
2 Related Works
2.1 Risk Categories
3 Methodology
3.1 Comparison Criteria
3.2 Validation Technique
4 Results
4.1 Qualification of Panelist
4.2 Hypothesis Testing
5 Conclusion
References
Risk Mitigation Framework for Agile Global Software Development
1 Introduction
2 Related Work
3 Proposed RMAG
4 Applying RMAG in Agile Global Software Development Project Using SCRUM
5 Discussion and Result
6 Conclusion and Future Work
Appendix A
References
Re-verification of the Improved Software Project Monitoring Task Model of the Agile Kanban (i-KAM)
1 Introduction
2 Overview of i-KAM
3 Methodology
4 Results
4.1 Understandability of the Terminologies Used in i-KAM
4.2 Relevance of the i-KAM Components
4.3 Feasibility of Criteria Used in i-KAM
4.4 Organization of the Connections and Flows in i-KAM
4.5 Comprehensiveness of i-KAM
5 Discussion
6 Conclusion
References
Author Index

Citation preview

Lecture Notes on Data Engineering and Communications Technologies 72

Faisal Saeed Fathey Mohammed Abdulaziz Al-Nahari   Editors

Innovative Systems for Intelligent Health Informatics Data Science, Health Informatics, Intelligent Systems, Smart Computing

Lecture Notes on Data Engineering and Communications Technologies Volume 72

Series Editor Fatos Xhafa, Technical University of Catalonia, Barcelona, Spain

The aim of the book series is to present cutting edge engineering approaches to data technologies and communications. It will publish latest advances on the engineering task of building and deploying distributed, scalable and reliable data infrastructures and communication systems. The series will have a prominent applied focus on data technologies and communications with aim to promote the bridging from fundamental research on data science and networking to data engineering and communications that lead to industry products, business knowledge and standardisation. Indexed by SCOPUS, INSPEC, EI Compendex. All books published in the series are submitted for consideration in Web of Science.

More information about this series at http://www.springer.com/series/15362

Faisal Saeed Fathey Mohammed Abdulaziz Al-Nahari •

•

Editors

Innovative Systems for Intelligent Health Informatics Data Science, Health Informatics, Intelligent Systems, Smart Computing

123

Editors Faisal Saeed College of Computer Science and Engineering Taibah University Medina, Saudi Arabia

Fathey Mohammed School of Computing, Information Systems Department Universiti Utara Malaysia Sintok, Malaysia

Abdulaziz Al-Nahari Sanaa’a Community College Sana’a, Yemen

ISSN 2367-4512 ISSN 2367-4520 (electronic) Lecture Notes on Data Engineering and Communications Technologies ISBN 978-3-030-70712-5 ISBN 978-3-030-70713-2 (eBook) https://doi.org/10.1007/978-3-030-70713-2 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

We are honored to welcome you to the 5th International Conference of Reliable Information and Communication Technology 2020 (IRICT2020) that was conducted online on December 21–22, 2020, and organized by the Yemeni Scientists Research Group (YSRG), Information Service Systems and Innovation Research Group (ISSIRG) in Universiti Teknologi Malaysia (Malaysia), Data Science Research Group in College of Computer Science and Engineering at Taibah University (Kingdom of Saudi Arabia), School of Science and Technology in Nottingham Trent (UK), College of Engineering, IT, and Environment at Charles Darwin University (Australia) and Association for Information Systems—Malaysia Chapter (MyAIS). IRICT2020 is a forum for the presentation of technological advances in the field of information and communication technology. The main theme of the conference is “Innovative Systems for Intelligent Health Informatics.” Many researchers have been attracted to submit 140 papers to IRICT2020 from 29 countries including Algeria, Australia, China, Egypt, Fiji, Germany, India, Indonesia, Iraq, Iran, Jordon, Malaysia, Morocco, Myanmar, Nigeria, Oman, Pakistan, Saudi Arabia, Singapore, Somalia, South Africa, Sri Lanka, Sudan, Sweden, Taiwan, Tunisia, UK, USA and Yemen. Of those 140 submissions, 111 submissions have been selected to be included in this book. The book presents several research topics which include health informatics, bioinformatics, information retrieval, artificial intelligence, machine learning, data science, big data analytics, business intelligence, Internet of things (IoT), information security, intelligent communication systems, information systems theories and applications, computational vision and robotics technology, software engineering and multimedia applications and services. We would like to express our appreciations to all authors and the keynote speakers for sharing their expertise with us. And we would like to thank the organizing committee for their great efforts in managing the conference. In addition,

v

vi

Preface

we would like to thank the technical committee for reviewing all the submitted papers; Prof. Dr. Janusz Kacprzyk, AISC series editor, Prof. Dr. Fatos Xhafa, book series editor; and Dr. Thomas Ditzinger from Springer. Finally, we thank all the participants of IRICT2020 and hope to see you all again in the next conference.

Organization

IRICT2020 Organizing Committee Honorary Co-chairs Rose Alinda Alias

Ahmad Lotfi

Abdullah Alsaeedi

Association for Information Systems—Malaysian Chapter, Head of the Information Service Systems and Innovation Research Group (ISSIRG) in Universiti Teknologi Malaysia Computing and Technology School of Science and Technology, Nottingham Trent University, UK College of Computer Science and Engineering, Taibah University, Kingdom of Saudi Arabia

Conference General Chair Faisal Saeed

Yemeni Scientists Research Group (YSRG), Head of Data Science Research Group in Taibah University, Kingdom of Saudi Arabia

Program Committee Chair Fathey Mohammed

Universiti Utara Malaysia (UUM), Malaysia

General Secretary Nadhmi Gazem

Taibah University, Kingdom of Saudi Arabia

vii

viii

Organization

Technical Committee Chairs Faisal Saeed Tawfik Al-Hadhrami Mamoun Alazab

Taibah University, Kingdom of Saudi Arabia Nottingham Trent University, UK Charles Darwin University, Australia

Publications Committee Fathey Mohammed Abdulaziz Al-Nahari

Universiti Utara Malaysia Sanaá Community College

Publicity Committee Abdullah Aysh Dahawi Maged Naeser Mohammed Omar Awadh Al-Shatari Ali Ahmed Ali Salem

Universiti Teknologi Malaysia Universiti Teknologi Malaysia Universiti Teknologi PETRONAS Universiti Tun Hussein Onn Malaysia

IT and Multimedia Committee Fuad Abdeljalil Al-shamiri Mohammed Alsarem Amer Alsaket Sulaiman Mohammed Abdulrahman

Universiti Teknologi Malaysia Taibah University, KSA Sitecore, Malaysia Taibah University, KSA

Treasure Committee Abdullah Aysh Dahawi

Universiti Teknologi Malaysia

Logistic Committee Chair Wahid Al-Twaiti

Universiti Teknologi Malaysia (UTM)

Registration Sameer Hasan Albakri

Universiti Teknologi Malaysia (UTM)

Organization

ix

International Technical Committee Abdelhadi Raihani Abdelhamid Emara Abdelkaher Ait Abdelouahad Abdualmajed Ahmed Ghaleb Abdullah Almogahed Abdullah B. Nasser Abdulrahman Alsewari Abubakar Elsafi Aby Mathews Maluvelil Ahmad Fadhil Yusof Ahmed Awad Ahmed Majid Ahmed Mutahar Ahmed Rakha Ahmed Talal Alaa Alomoush Alaa Fareed Abdulateef Ali Ahmed Ameen Ba Homaid Aminu Aminu Mu’Azu Amr Tolba Amr Yassin Anton Satria Prabuwono Arwa Aleryani Auday Hashim Saeed Al-Wattar Bakar Ba Qatyan Bander Al-Rimy Bassam Al-Hameli Bouchaib Cherradi Mohammed Gamal Alsamman Haitham Alali Ehsan Othman Eissa Alshari Fadhl Hujainah Faisal Saeed Fathey Mohammed Fatma Al-Balushi Feras Zen Alden Fuad Ghaleb

Hassan II University of Casablanca, Morocco Taibah University, KSA Chouaib Doukkali University, Morocco Al-Khulaidi Sana’a University, Yemen Universiti Utara Malaysia, Malaysia University Malaysia Pahang, Malaysia University Malaysia Pahang, Malaysia University of Jeddah, KSA Independent Researcher, Canada Universiti Teknologi Malaysia, Malaysia King Abdulaziz University, KSA University of Information Technology and Communications, Iraq Management and Science University, Malaysia Al-Azhar University, Egypt Al-Iraqia University, Iraq Universiti Malaysia Pahang, Malaysia Universiti Utara Malaysia, Malaysia King Abdulaziz University, KSA University Malaysia Pahang, Malaysia Umaru Musa Yar’adua University Katsina, Nigeria King Saud University, KSA Ibb University, Yemen King Abdulaziz University, KSA Independent Researcher, Canada University of Mosul, Iraq Universiti Utara Malaysia, Malaysia UNITAR, Malaysia University Malaysia Pahang, Malaysia Hassan II University, Morocco Universiti Utara Malaysia, Malaysia Emirates College of Technology, UAE Ovgu Magdeburg, Germany Ibb University, Yemen University Malaysia Pahang, Malaysia Taibah University, KSA Universiti Utara Malaysia, Malaysia Independent Researcher, Oman Universiti Utara Malaysia, Malaysia Universiti Teknologi Malaysia, Malaysia

x

Hamzah Alaidaros Hanan Aldowah Hapini Bin Awang Hassan Silkan Hesham Alghodhaifi Hiba Zuhair Hussein Abualrejal Insaf Bellamine Insaf Bellamine Jawad Alkhateeb Kamal Alhendawi Khairul Shafee, Kalid Khaleel Bader Bataineh Khalili Tajeddine Marwa Alhadi Masud Hasan Mohamad Ghozali Hassan Mohamed Abdel Fattah Mohamed Elhamahmy Mohammed A. Al-Sharafi Mohammed Al-Mhiqani Mohammed Alsarem Mohammed Azrag Mohammed Nahid Mostafa Al-Emran Motasem Al Smadi Mustafa Ali Abuzaraida Mustafa Noori Nabil Al-Kumaim Nadhmi Gazem Naseebah Maqtary Nejood Hashim Al-Walidi Noor Akma Omar Dakkak Omar Zahour Osama Sayaydeh Qasim Alajmi Raed Aldhubhani Raghed Esmaeel Rajesh Kaluri Salah Abdelmageid Salwa Belaqziz Samar Ghazal Samar Salem Ahmed

Organization

Al-Ahgaff University, Yemen Universiti Sains Malaysia, Malaysia Universiti Utara Malaysia, Malaysia Université Chouaib Doukkali, Morocco University of Michigan, USA Al-Nahrain University, Iraq Universiti Utara Malaysia, Malaysia Chouaib Doukkali University, Morocco FSDM Fès, Morocco Taibah University, KSA Al-Quds Open University, Palestine Universiti Teknologi PETRONAS, Malaysia Amman Arab University, Jordan Hassan II University of Casablanca, Morocco Sana’a University, Yemen Taibah University, KSA Universiti Utara Malaysia, Malaysia Taibah University, KSA Egypt University Malaysia Pahang, Malaysia Universiti Teknikal Malaysia Melaka, Malaysia Taibah University, KSA University Malaysia Pahang, Malaysia Hassan II University, Casablanca, Morocco Buraimi University College, Oman Jordan University of Science and Technology, Jordan Universiti Utara Malaysia, Malaysia Middle Technical University, Iraq Universiti Teknikal Malaysia Melaka, Malaysia Taibah University, KSA University of Science and Technology, Yemen Cairo University, Egypt Universiti Malaysia Pahang, Malaysia UNIKA, Turkey Hassan II University of Casablanca, Morocco University Malaysia Pahang, Malaysia A’ Sharqiyah University, Oman University of Hafr Al Batin, KSA University of Mosul, Iraq Vellore Institute of Technology, India Taibah University, KSA Ibn Zohr University, Morocco Universiti Sains Malaysia, Malaysia International Islamic University, Malaysia

Organization

Sharaf J. Malebary Sinan Salih Soufiane Hamida Susan Abdulameer Syifak Izhar Hisham Tawfik Al-Hadhrami Waleed A. Hammood Waleed Ali Waleed Alomoush Waseem Alromimah Wasef Mater Yaqoub Sulaiman Yousef Fazea Yousif Abdullah AlRashidi Yousif Aftan Abdullah Yousif Munadhil Ibrahim Zainab Senan

xi

King Abdulaziz University, KSA Dijlah University College, Iraq Hassan II University of Casablanca, Morocco Universiti Utara Malaysia, Malaysia Universiti Malaysia Pahang, Malaysia Nottingham Trent University, UK Universiti Malaysia Pahang, Malaysia King Abdulaziz University, KSA Imam Abdulrahman bin Faisal University, KSA Taibah University, KSA University of Petra, Jordan King Abdulaziz University, KSA Universiti Utara Malaysia, Malaysia Al Yamamah University, KSA University of Baghdad, Iraq Universiti Utara Malaysia, Malaysia Universiti Utara Malaysia, Malaysia

Contents

Intelligent Health Informatics Comparative Study of SMOTE and Bootstrapping Performance Based on Predication Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abdulaziz Aborujilah, Rasheed Mohammad Nassr, Tawfik Al-Hadhrami, Mohd Nizam Husen, Nor Azlina Ali, Abdulaleem Al- Othmani, and Mustapha Hamdi

3

UPLX: Blockchain Platform for Integrated Health Data Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Omar Musa, Lim Shu Yun, and Reza Ismail

10

Convolutional Neural Networks for Automatic Detection of Colon Adenocarcinoma Based on Histopathological Images . . . . . . . . . . . . . . . Yakoop Qasim, Habeb Al-Sameai, Osamah Ali, and Abdulelah Hassan

19

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare Using Machine Learning . . . . . . . . . . . . . Radiah Haque, Sin-Ban Ho, Ian Chai, Chin-Wei Teoh, Adina Abdullah, Chuie-Hong Tan, and Khairi Shazwan Dollmat A CNN-Based Model for Early Melanoma Detection . . . . . . . . . . . . . . . Amer Sallam, Abdulfattah E. Ba Alawi, and Ahmed Y. A. Saeed SMARTS D4D Application Module for Dietary Adherence Self-monitoring Among Hemodialysis Patients . . . . . . . . . . . . . . . . . . . . Hafzan Yusoff, Nur Intan Raihana Ruhaiyem, and Mohd Hakim Zakaria Improved Multi-label Medical Text Classification Using Features Cooperation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Rim Chaib, Nabiha Azizi, Nawel Zemmal, Didier Schwab, and Samir Brahim Belhaouari

29

41

52

61

xiii

xiv

Contents

Image Modeling Through Augmented Reality for Skin Allergies Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nur Intan Raihana Ruhaiyem and Nur Amalina Mazlan

72

Hybridisation of Optimised Support Vector Machine and Artificial Neural Network for Diabetic Retinopathy Classification . . . . . . . . . . . . Nur Izzati Ab Kader, Umi Kalsom Yusof, and Maziani Sabudin

80

A Habit-Change Support Web-Based System with Big Data Analytical Features for Hospitals (Doctive) . . . . . . . . . . . . . . . . . . . . . . Cheryll Anne Augustine and Pantea Keikhosrokiani

91

An Architecture for Intelligent Diagnosing Diabetic Types and Complications Based on Symptoms . . . . . . . . . . . . . . . . . . . . . . . . . 102 Gunasekar Thangarasu, P. D. D. Dominic, and Kayalvizhi Subramanian An Advanced Encryption Cryptographically-Based Securing Applicative Protocols MQTT and CoAP to Optimize Medical-IOT Supervising Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Sanaa El Aidi, Abderrahim Bajit, Anass Barodi, Habiba Chaoui, and Ahmed Tamtaoui Pulmonary Nodule Classification Based on Three Convolutional Neural Networks Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 Enoumayri Elhoussaine and Belaqziz Salwa A Comparative Study on Liver Tumor Detection Using CT Images . . . 129 Abdulfattah E. Ba Alawi, Ahmed Y. A. Saeed, Borhan M. N. Radman, and Burhan T. Alzekri Brain Tumor Diagnosis System Based on RM Images: A Comparative Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 Ahmed Y. A. Saeed, Abdulfattah E. Ba Alawi, and Borhan M. N. Radman Diagnosis of COVID-19 Disease Using Convolutional Neural Network Models Based Transfer Learning . . . . . . . . . . . . . . . . . . . . . . . 148 Hicham Moujahid, Bouchaib Cherradi, Mohammed Al-Sarem, and Lhoussain Bahatti Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 Tariq Saeed Mian Detection of Cardiovascular Disease Using Ensemble Machine Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Fizza Kashif and Umi Kalsom Yusof

Contents

xv

Health Information Management Hospital Information System for Motivating Patient Loyalty: A Systematic Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 Saleh Nasser Rashid Alismaili, Mohana Shanmugam, Hairol Adenan Kasim, and Pritheega Magalingam Context Ontology for Smart Healthcare Systems . . . . . . . . . . . . . . . . . . 199 Salisu Garba, Radziah Mohamad, and Nor Azizah Saadon A Modified UTAUT Model for Hospital Information Systems Geared Towards Motivating Patient Loyalty . . . . . . . . . . . . . . . . . . . . . 207 Saleh Nasser Rashid Alismaili, Mohana Shanmugam, Hairol Adenan Kasim, and Pritheega Magalingam Teamwork Communication in Healthcare: An Instrument (Questionnaire) Validation Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Wasef Matar and Monther Aldwair Potential Benefits of Social Media to Healthcare: A Systematic Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230 Ghada Ahmad Abdelguiom and Noorminshah A. Iahad Exploring the Influence of Human-Centered Design on User Experience in Health Informatics Sector: A Systematic Review . . . . . . . 242 Lina Fatini Azmi and Norasnita Ahmad An Emotional-Persuasive Habit-Change Support Mobile Application for Heart Disease Patients (BeHabit) . . . . . . . . . . . . . . . . . . 252 Bhavani Devi Ravichandran and Pantea Keikhosrokiani A Systematic Review of the Integration of Motivational and Behavioural Theories in Game-Based Health Interventions . . . . . . . 263 Abdulsalam S. Mustafa, Nor’ashikin Ali, and Jaspaljeet Singh Dhillon Adopting React Personal Health Record (PHR) System in Yemen HealthCare Institutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279 Ziad Saif Alrobieh, Dhiaa Faisal Alshamy, and Maged Nasser Artificial Intelligence and Soft Computing Application of Shuffled Frog-Leaping Algorithm for Optimal Software Project Scheduling and Staffing . . . . . . . . . . . . . . . . . . . . . . . . 293 Ahmed O. Ameen, Hammed A. Mojeed, Abdulazeez T. Bolariwa, Abdullateef O. Balogun, Modinat A. Mabayoje, Fatima E. Usman-Hamzah, and Muyideen Abdulraheem A Long Short Term Memory and a Discrete Wavelet Transform to Predict the Stock Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 Mu’tasem Jarrah and Naomie Salim

xvi

Contents

Effective Web Service Classification Using a Hybrid of Ontology Generation and Machine Learning Algorithm . . . . . . . . . . . . . . . . . . . . 314 Murtoza Monzur, Radziah Mohamad, and Nor Azizah Saadon Binary Cuckoo Optimisation Algorithm and Information Theory for Filter-Based Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324 Ali Muhammad Usman, Umi Kalsom Yusof, and Syibrah Naim Optimized Text Classification Using Correlated Based Improved Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 339 Thabit Sabbah Multi-objective NPO Minimizing the Total Cost and CO2 Emissions for a Stand-Alone Hybrid Energy System . . . . . . . . . . . . . . . . . . . . . . . 351 Abbas Q. Mohammed, Kassim A. Al-Anbarri, and Rafid M. Hannun A Real Time Flood Detection System Based on Machine Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364 Abdirahman Osman Hashi, Abdullahi Ahmed Abdirahman, Mohamed Abdirahman Elmi, and Siti Zaiton Mohd Hashim Extracting Semantic Concepts and Relations from Scientific Publications by Using Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 374 Fatima N. AL-Aswadi, Huah Yong Chan, and Keng Hoon Gan Effectiveness of Convolutional Neural Network Models in Classifying Agricultural Threats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384 Sayem Rahman, Murtoza Monzur, and Nor Bahiah Ahmad A Study on Emotion Identification from Music Lyrics . . . . . . . . . . . . . . 396 Affreen Ara and Raju Gopalakrishna A Deep Neural Network Model with Multihop Self-attention Mechanism for Topic Segmentation of Texts . . . . . . . . . . . . . . . . . . . . . 407 Fayçal Nouar and Hacene Belhadef Data Science and Big Data Analytics Big Data Interoperability Framework for Malaysian Public Open Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421 Najhan Muhamad Ibrahim, Amir Aatieff Amir Hussin, Khairul Azmi Hassan, and Ciara Breathnach The Digital Resources Objects Retrieval: Concepts and Figures . . . . . . 430 Wafa’ Za’al Alma’aitah, Abdullah Zawawi Talib, and Mohd Azam Osman A Review of Graph-Based Extractive Text Summarization Models . . . . 439 Abdulkadir Abubakar Bichi, Ruhaidah Samsudin, Rohayanti Hassan, and Khalil Almekhlafi

Contents

xvii

Review on Emotion Recognition Using EEG Signals Based on Brain-Computer Interface System . . . . . . . . . . . . . . . . . . . . . . . . . . . 449 Mona Algarni and Faisal Saeed A New Multi-resource Deadlock Detection Algorithm Using Directed Graph Requests in Distributed Database Systems . . . . . . . . . . 462 Khalid Al-Hussaini, Nabeel A. Al-Amdi, and Fuaad Hasan Abdulrazzak Big Data Analytics Model for Preventing the Spread of COVID-19 During Hajj Using the Proposed Smart Hajj Application . . . . . . . . . . . 475 Ibtehal Nafea Financial Time Series Forecasting Using Prophet . . . . . . . . . . . . . . . . . 485 Umi Kalsom Yusof, Mohd Nor Akmal Khalid, Abir Hussain, and Haziqah Shamsudin Facial Recognition to Identify Emotions: An Application of Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 496 Kenza Belhouchette Text-Based Analysis to Detect Figure Plagiarism . . . . . . . . . . . . . . . . . . 505 Taiseer Abdalla Elfadil Eisa, Naomie Salim, and Salha Alzahrani A Virtual Exploration of al-Masjid al-Nabawi Using Leap Motion Controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 514 Slim Kammoun and Hamza Ghandorh Comparison of Data Analytic Techniques for a Spatial Opinion Mining in Literary Works: A Review Paper . . . . . . . . . . . . . . . . . . . . . 523 Sea Yun Ying, Pantea Keikhosrokiani, and Moussa Pourya Asl Open Data in Prediction Using Machine Learning: A Systematic Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536 Norismiza Ismail and Umi Kalsom Yusof Big Data Analytics Based Model for Red Chili Agriculture in Indonesia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554 Junita Juwita Siregar and Arif Imam Suroso A Fusion-Based Feature Selection Framework for Microarray Data Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 565 Talal Almutiri, Faisal Saeed, Manar Alassaf, and Essa Abdullah Hezzam An Approach Based Natural Language Processing for DNA Sequences Encoding Using the Global Vectors for Word Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577 Brahim Matougui, Hacene Belhadef, and Ilham Kitouni

xviii

Contents

Short-Term CO2 Emissions Forecasting Using Multi-variable Grey Model and Artificial Bee Colony (ABC) Algorithm Approach . . . . . . . . 586 Ani Shabri, Ruhaidah Samsudin, and Essa Abdullah Hezzam IoT and Intelligent Communication Systems A Reliable Single Prediction Data Reduction Approach for WSNs Based on Kalman Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601 Zaid Yemeni, Haibin Wang, Waleed M. Ismael, Younis Ibrahim, and Peng Li A Real-Time Groundwater Level Monitoring System Based on WSN, Taiz, Yemen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612 Asma’a K. Akershi, Ziad S. Arobieh, and Reayidh A. Ahmed Design and Simulation of Multiband Circular Microstrip Patch Antenna with CSRR for WLAN and WiMAX Applications . . . . . . . . . . 623 Abdulguddoos S. A. Gaid, Amer A. Sallam, Mohamed H. M. Qasem, Maged S. G. Abbas, and Amjad M. H. Aoun Reference Architectures for the IoT: A Survey . . . . . . . . . . . . . . . . . . . 635 Raghdah Saemaldahr, Bijayita Thapa, Kristopher Maikoo, and Eduardo B. Fernandez A Circular Multiband Microstrip Patch Antenna with DGS for WLAN/WiMAX/Bluetooth/UMTS/LTE . . . . . . . . . . . . . . . . . . . . . . 647 Abdulguddoos S. A. Gaid, Amer A. Sallam, Mohamed H. M. Qasem, Maged S. G. Abbas, and Amjad M. H. Aoun Anomaly Intrusion Detection Systems in IoT Using Deep Learning Techniques: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 659 Muaadh. A. Alsoufi, Shukor Razak, Maheyzah Md Siraj, Abdulalem Ali, Maged Nasser, and Salah Abdo Security and Threats in the Internet of Things Based Smart Home . . . . 676 Nor Fatimah Awang, Ahmad Fudhail Iyad Mohd Zainudin, Syahaneim Marzuki, Syed Nasir Alsagoff, Taniza Tajuddin, and Ahmad Dahari Jarno Simulation and Control of Industrial Composition Process Over Wired and Wireless Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 685 Hakim Qaid Abdullah Abdulrab, Fawnizu Azmadi Hussin, Panneer Selvam Arun, Azlan Awang, and Idris Ismail Performance Degradation of Multi-class Classification Model Due to Continuous Evolving Data Streams . . . . . . . . . . . . . . . . . . . . . . . . . . 696 Abdul Sattar Palli, Jafreezal Jaafar, and Manzoor Ahmed Hashmani

Contents

xix

Compact Wide-Bandwidth Microstrip Antenna for Millimeter Wave Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 707 Osaid Abdulrahman Saeed, Moheeb Ali Ameer, and Mansour Noman Ghaleb Dual-Band Rectangular Microstrip Patch Antenna with CSRR for 28/38 GHz Bands Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 717 Abdulguddoos S. A. Gaid, Mohamed H. M. Qasem, Amer A. Sallam, and Ebrahim Q. M. Shayea Dual Band Rectangular Microstrip Patch Antenna for 5G Millimeter-Wave Wireless Access and Backhaul Applications . . . . . . . . 728 Abdulguddoos S. A. Gaid, Amer A. Sallam, Amjad M. H. Aoun, Ahmed A. A. Saeed, and Osama Y. A. Saeed Design of Wireless Local Multimedia Communication Network (WLMmCN) Based on Android Application Without Internet Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 739 R. Q. Shaddad, F. A. Alqasemi, S. A. Alfaqih, M. F. Alsabahi, A. T. Fara, K. M. Nejad, and E. A. Albukhaiti A Statistical Channel Propagation Analysis for 5G mmWave at 73 GHz in Urban Microcell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748 Zaid Ahmed Shamsan Advances in Information Security Robot Networks and Their Impact on Cyber Security and Protection from Attacks: A Review . . . . . . . . . . . . . . . . . . . . . . . . . 759 Daniah Anwar Hasan and Linah Faisal Tasji An Efficient Fog-Based Attack Detection Using Ensemble of MOA-WMA for Internet of Medical Things . . . . . . . . . . . . . . . . . . . 774 Shilan S. Hameed, Wan Haslina Hassan, and Liza Abdul Latiff A New DNA Based Encryption Algorithm for Internet of Things . . . . . 786 Bassam Al-Shargabi and Mohammed Abbas Fadhil Al-Husainy Watermarking Techniques for Mobile Application: A Review . . . . . . . . 796 Aqilah Abd. Ghani, Syifak Izhar Hisham, Nur Alya Afikah Usop, and Nor Bakiah Abd Warif Analysis and Evaluation of Template Based Methods Against Geometric Attacks: A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 807 Tanya Koohpayeh Araghi, Ala Abdulsalam Alarood, and Sagheb Kohpayeh Araghi Survey of File Carving Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815 Nor Ika Shahirah Ramli, Syifak Izhar Hisham, and Mohd Faizal Abd Razak

xx

Contents

Affecting Factors in Information Security Policy Compliance: Combine Organisational Factors and User Habits . . . . . . . . . . . . . . . . . 826 Angraini, Rose Alinda Alias, and Okfalisa Mitigation of Data Security Threats in Iraqi Dam Management Systems: A Case Study of Fallujah Dam Management System . . . . . . . . 837 Hussam J. Ali, Hiba Zuhair, and Talib M. Jawad Advances in Information Systems Development and Validation of a Classified Information Assurance Scale for Institutions of Higher Learning . . . . . . . . . . . . . . . . . . . . . . . . 857 Bello Ahmadu, Ab Razak Che Hussin, and Mahadi Bahari Sustainable e-Learning Framework: Expert Views . . . . . . . . . . . . . . . . 869 Aidrina Binti Mohamed Sofiadin Derivation of a Customer Loyalty Factors Based on Customers’ Changing Habits in E-Commerce Platform . . . . . . . . . . . . . . . . . . . . . . 879 Mira Afrina, Samsuryadi, Ab Razak Che Hussin, and Suraya Miskon Analysis of Multimedia Elements Criteria Using AHP Method . . . . . . . 891 Nadiah Mohamad Sofian, Ahmad Sobri Hashim, and Aliza Sarlan The Development of a Criteria-Based Group Formation Systems for Student Group Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 903 Divya Gopal Mohan and Khairul Shafee Kalid Trusted Factors of Social Commerce Product Review Video . . . . . . . . . 911 Humaira Hairudin, Halina Mohamed Dahlan, and Ahmad Fadhil Yusof Building Information Modelling Adoption: Systematic Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 920 Hafiz Muhammad Faisal Shehzad, Roliana Binti Ibrahim, Ahmad Fadhil Yusof, Khairul Anwar Mohamed Khaidzir, Omayma Husain Abbas Hassan, and Samah Abdelsalam Abdalla Adoption of Smart Cities Models in Developing Countries: Focusing in Strategy and Design in Sudan . . . . . . . . . . . . . . . . . . . . . . . 933 Mohmmed S. Adrees, Abdelrahman E. Karrar, and Waleed I. Osman Factors Affecting Customer Acceptance of Online Shopping Platforms in Malaysia: Conceptual Model and Preliminary Results . . . . . . . . . . . . 945 Nabil Hasan Al-kumaim, Gan Wong Sow, and Fathey Mohammed Student Compliance Intention Model for Continued Usage of E-Learning in University . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 960 Ken Ditha Tania, Norris Syed Abdullah, Norasnita Ahmad, and Samsuryadi Sahmin

Contents

xxi

Digital Information and Communication Overload Among Youths in Malaysia: A Preliminary Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . 975 Mohamad Ghozali Hassan, Muslim Diekola Akanmu, Hussein Mohammed Esmail Abualrejal, and Amal Abdulwahab Hasan Alamrani The Effect of Using Social Networking Sites on Undergraduate Students’ Perception and Academic Performance at University of Taiz-Yemen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 987 Maged Rfeqallah, Rozilah Kasim, Faisal A. M. Ali, and Yahya Abdul Ghaffar Building Information Modelling Adoption Model for Malaysian Architecture, Engineering and Construction Industry . . . . . . . . . . . . . . 999 Hafiz Muhammad Faisal Shehzad, Roliana Binti Ibrahim, Ahmad Fadhil Yusof, Khairul Anwar Mohamed Khaidzir, Muhammad Mahboob Khurshid, and Farah Zeehan Othman Digital Government Competency for Omani Public Sector Managers: A Conceptual Framework . . . . . . . . . . . . . . . . . . . . . . . . . . 1009 Juma Al-Mahrezi, Nur Azaliah Abu Bakar, and Nilam Nur Amir Sjarif Computational Vision and Robotics Landmark Localization in Occluded Faces Using Deep Learning Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1023 Zieb Rabie Alqahtani, Mohd Shahrizal Sunar, and Abdulaziz A. Alashbi Contrast Image Quality Assessment Algorithm Based on Probability Density Functions Features . . . . . . . . . . . . . . . . . . . . . . . 1030 Ismail Taha Ahmed, Soong Der Chen, Norziana Jamil, and Baraa Tareq Hammad The Impact of Data Augmentation on Accuracy of COVID-19 Detection Based on X-ray Images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1041 Yakoop Qasim, Basheer Ahmed, Tawfeek Alhadad, Habeb Al-Sameai, and Osamah Ali A Fusion Schema of Hand-Crafted Feature and Feature Learning for Kinship Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1050 Mohammed Ali Almuashi, Siti Zaiton Mohd Hashim, Nooraini Yusoff, and Khairul Nizar Syazwan Lossless Audio Steganographic Method Using Companding Technique . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1064 Ansam Osamah Abdulmajeed

xxii

Contents

Smart Traffic Light System Design Based on Single Shot MultiBox Detector (SSD) and Anylogic Simulation . . . . . . . . . . . . . . . . 1075 E. R. Salim, A. B. Pantjawati, D. Kuswardhana, A. Saripudin, N. D. Jayanto, Nurhidayatulloh, and L. A. Pratama Learning Scope of Python Coding Using Immersive Virtual Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1086 Abdulrazak Yahya Saleh, Goh Suk Chin, Roselind Tei, Mohd Kamal Othman, Fitri Suraya Mohamad, and Chwen Jen Chen Automatic Audio Replacement of Objectionable Content for Sri Lankan Locale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1101 Gobiga Rajalingam, Janarthan Jeyachandran, M. S. M. Siriwardane, Tharshvini Pathmaseelan, R. K. N. D. Jayawardhane, and N. S. Weerakoon A Comparison of CNN and Conventional Descriptors for Word Spotting Approach: Application to Handwritten Document Image Retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1115 Ryma Benabdelaziz, Djamel Gaceb, and Mohammed Haddad Handwritten Arabic Character Recognition: Comparison of Conventional Machine Learning and Deep Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1127 Faouci Soumia, Gaceb Djamel, and Mohammed Haddad Document Image Edge Detection Based on a Local Hysteresis Thresholding and Automatic Setting Using PSO . . . . . . . . . . . . . . . . . . 1139 Mohamed Benkhettou, Nibel Nadjeh, and Djamel Gaceb Fast I2SDBSCAN Based on Integral Volume of 3D Histogram: Application to Color Layer Separation in Document Images . . . . . . . . . 1151 Zakia Kezzoula and Djamel Gaceb Enhancing Daily Life Skills Learning for Children with ASD Through Augmented Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1164 Rahma Bouaziz, Maimounah Alhejaili, Raneem Al-Saedi, Abrar Mihdhar, and Jawaher Alsarrani Recent Computing and Software Engineering SpaceScience App: Development of a Mobile Application for School Children . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1177 Wan Fatimah Wan Ahmad and Ain Fatihah Ahmad Harnaini Research on Online Problem-Based Learning Among Undergraduate Students: A Systematic Review . . . . . . . . . . . . . . . . . . . 1187 Amira Saif and Irfan Umar

Contents

xxiii

Derivation of Factors in Dealing Negative E-WOM for Maintaining Online Reputation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1198 Rizka Dhini Kurnia, Halina Mohamed Dahlan, and Samsuryadi A Terms Interrelationship Approach to Query Expansion Based on Terms Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1209 Nuhu Yusuf, Mohd Amin Mohd Yunus, Norfaradilla Wahid, Mohd Najib Mohd Salleh, and Aida Mustapha Multi-domain Business Intelligence Model for Educational Enterprise Resource Planning Systems . . . . . . . . . . . . . . . . . . . . . . . . . . 1218 Hisham Abdullah, Azman Taa, and Fathey Mohammed Measuring Risk Mitigation Techniques in Agile Global Software Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1225 Adila Firdaus Arbain, Muhammad Akil Rafeek, Zuriyaninatasa Podari, and Cik Feresa Mohd Foozy Risk Mitigation Framework for Agile Global Software Development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1233 Zuriyaninatasa Podari, Adila Firdaus Arbain, Noraini Ibrahim, and Endah Sudarmilah Re-verification of the Improved Software Project Monitoring Task Model of the Agile Kanban (i-KAM) . . . . . . . . . . . . . . . . . . . . . . . . . . . 1247 Hamzah Alaidaros, Mazni Omar, Rohaida Romli, and Adnan Hussein Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1259

Intelligent Health Informatics

Comparative Study of SMOTE and Bootstrapping Performance Based on Predication Methods Abdulaziz Aborujilah1(B) , Rasheed Mohammad Nassr1 , Tawfik Al-Hadhrami2 , Mohd Nizam Husen1 , Nor Azlina Ali1 , Abdulaleem Al- Othmani1 , and Mustapha Hamdi3 1 University Kuala Lumpur, 50250 Kuala Lumpur, Malaysia

[email protected] 2 Nottingham Trent University, Nottingham NG1 4FQ, UK 3 Edge IA, IoT, Nottingham, UK

Abstract. Recently, there has been a renewed interest in smart health systems that aim to deliver high quality healthcare services. Prediction methods are very essential to support these systems. They mainly rely on datasets with assumptions that match the reality. However, one of the greatest challenges to prediction methods is to have datasets which are normally distributed. This paper presents an experimental work to implement SMOTE (Synthetic Minority Oversampling Technique) and bootstrapping methods to normalize datasets. It also measured the impact of both methods in the performance of different prediction methods such as Support vector machine (SVM), Naive Bayes, and neural network(NN) The results showed that bootstrapping with native bays yielded better prediction performance as compared to other prediction methods with SMOTE. Keywords: Datasets normalization · Prediction systems · Dataset redistribution methods · SMOTE-Bootstrapping

1 Introduction Healthcare systems demand for accurate data to handle all aspects of healthcare tasks from making the policies until delivering the end services [1]. Data normality is necessary for efficient healthcare systems. Data mining methods are mainly used in healthcare applications such as disease predictions. The performance of such applications is basically influenced by the issue of classes’ distributions. These methods presumably deal with balanced datasets. However, most real datasets are not balanced. This causes poor performance of prediction systems. Thus, datasets balancing issues are getting more attention from researchers [1]. The issue of imbalanced dataset highly impacts the sensitivity of prediction methods and creates a bias in prediction performance [2]. For example, misclassification of the minority class causes serious consequences in detecting fraud, intrusion, and chronic diseases [1]. Imbalanced data makes classification algorithms more sensitive towards majority classes. This is due to the performance © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 3–9, 2021. https://doi.org/10.1007/978-3-030-70713-2_1

4

A. Aborujilah et al.

of the classification algorithms which are highly biased towards the majority classes, at the same time performing poorly towards the minority classes. Reducing such biases requires techniques to balance the data such as Cost-Sensitive Approach, samplingbased techniques, algorithm modifications [3], and attributes ranking [4]. This paper demonstrated how sampling methods such as SMOTE and bootstrapping influence the production performance. The remainder of this paper is organized with the second section presenting the related studies followed by the experiment section. The next section is results’ discussion while the last section is on conclusion and future work.

2 Related Works An imbalanced dataset has a direct impact in the accuracy of prediction systems. This impact results in the form of abnormality of classes’ distribution. It is relatively difficult to have accurate prediction systems with imbalanced classes [1]. Similarly, it is not easy to have balanced datasets from real life situations. This is due to the ubiquitous nature of real-life datasets. Therefore, it is a researchable issue of improving the predictive performance of classification algorithms and reducing the deficiency of class imbalance [1]. An imbalanced dataset consists of data that belong to both majority and minority classes. From this view, two main approaches to deal with imbalanced data have been identified. First approach is under-sampling which aims to reduce the number of majority samples [6] while second is oversampling which aims to increase the amount of minority samples [7]. Another method [8] for sampling is Bagging method which uses random sampling with the replacement of the original dataset. The optimal dataset is chosen based on the average results of all the models (ensemble model). This method relies on random sampling and does not rely on training results. It also performs well with generalization and has no direct impact on imbalance unless some other factors are included. In [9] and [10], a modification of oversampling was suggested by selecting the training sample randomly. This method has a positive impact in improving accuracy of the classification model. However, it causes delay in processing large datasets. Another method of sampling was proposed in [11] and [12] where it leads to new data generation from the training dataset. It keeps the basic characteristics of the original dataset but lowers the classification model accuracy as the original dataset maintains its integrity. Another sampling method is called under-sampling which focuses on reducing the majority class and finds a balance with minority class. It has gained more attention among academics [1]. The study [13] presented two methods of sampling, under-sampling as random and informative. Informative under-sampling looks to achieve the data balance by eliminating dates from the training dataset based on predefined criteria. Deep neural network is used in different domains. It is also used widely in security applications. SMOTE was proposed by Chawla et al. [7]. Its core idea is constructing the synthetic minority samples through the interpolation between minority training data and its k-nearest neighbourhoods [14]. SMOTE is an oversampling technique that is used to increase the minority class samples by generating data artificially. It continuously increases the minority until the dataset reaches an acceptable ratio where the minority class and majority class become approximately equal [15]. SMOTE is an acceptable

Comparative Study of SMOTE and Bootstrapping Performance

5

method for oversampling. On the other hand, it suffers from high ratio of misclassification [16]. It uses minority class to generate new synthetic examples based on the randomly selected k-nearest neighbors.The number of generated samples depends on a predefined over-sampling ratio [17]. This pre-processing method is widely used to enhance the dataset balancing. More works on the extensions are on-going [18] such as generating and removing the samples simultaneously based on their influences on the model performance, such as SPIDER [19]. Bootstrapping technique resamples each training record by applying the probability with replacements based on the bagging algorithm concept [14]. The bagging algorithm creates a random forest of samples and features used in training that are selected randomly [20]. Bootstrapping is a data replacement method that replaces the actual dataset items with their statistical interests such as mean as shown in the following steps: 1. Given a data set of size n, replacement is applied on the sample from the data set n times 2. Repeat step 1 m times (e.g., m = 10,000) 3. For each vector produced by step 1, calculate the statistic of interest (e.g., the mean) 4. The result is a distribution of the statistic (e.g., if the statistic of interest is the mean, with the assumptions that you picked the right m and n is large enough, step 3 should result in a normal distribution) [22].

3 Experiments The goal of these experiments was to investigate how to handle the problem of imbalanced dataset through the refinement of training dataset. To carry out the experiment RapidMiner studio 9.3 [1] have been used. Well-known dataset which is Breast Cancer Wisconsin (Diagnostic) dataset has been selected [21]. It consists of 569 patient records of breast cancer with 12 attributes. Table 1 shows the used attributes and their definition. The steps of conducting the experiments are consists of three main phases, training, testing and evaluation. At the training phase the dataset (2/3 of whole dataset) is extracted form CSV file to rapid miner data store then diagnosis feature was selected as the class label. Then SMOTE method is used to normalize dataset, next SVM, Naïve Bayes and Neural network classification models are created. Afterword, at the testing phase (1/3 of whole dataset) is extracted and the models are applied. Then confusion matrix and measurements are calculated. The same steps are repeated and SMOTE was replaced with Bootstrapping method. Finally, the measurements values of both methods are compared. Figure 1 show the steps of conducting the experiments. A predictor attribute is a diagnosis which takes two values either M = malignant or B = benign. The number of records in M class was 212 (37%) and B class was 357 (63%). This shows that the dataset was not balanced. Two types of experiments were done. The first experiment aimed to examine how bootstrapping and SMOTE methods impact the prediction performance while the second aimed to evaluate how the prediction performance is affected by using the oversampling method. Three types of prediction methods were used in this evaluation: SVM, neural networks, and Naïve Bayes. Table 2 shows the results of the prediction performance by using bootstrapping and SMOTE with SVM, neural networks, and Naïve Bayes.

6

A. Aborujilah et al.

Fig. 1. Experiments phases flow diagram

Table 1. Breast cancer Wisconsin dataset attributes Features

Date type Meaning

1.

ID number

Real

ID

2.

Diagnosis

Real

M = malignant, B = benign

3.

Radius

Real

Mean of distances from centre to points on the perimeter

4.

Texture

Real

Standard deviation of gray-scale values

5.

Perimeter

Real

Perimeter

6.

Area

Real

Area

7.

Smoothness

Real

Local variation in radius lengths

8.

Compactness

Real

Perimeter2 / area - 1.0

9.

Concavity

Real

Severity of concave portions of the contour

10. Concave points

Real

Number of concave portions of the contour

11. Symmetry

Real

Symmetry

12. Fractal dimension Real

“Coastline approximation” -

Table 2 compares the prediction performance of three methods: SVM, Naïve Bayes, and neural networks using SMOTE and bootstrapping. The results showed that the bootstrapping generally did better than SMOTE. For example, SVM with bootstrapping achieved the best results as compared to the prediction method that used SMOTE. The results of prediction sensitivity, specificity, precision and negative predictive, and accuracy reached 100%. F1 Score, Matthews Correlation Coefficient, False Positive Rate,

Comparative Study of SMOTE and Bootstrapping Performance

7

Table 2. Performance comparison of SMOTE and Bootstrap methods Measurements

SMOTE

Bootstrapping

SVM

Neural networks

Naive bayes

SVM

Neural networks

Naive bayes

Sensitivity

0.972

0.9813

0.8879

1

0.9848

0.8939

Specificity

0.9813

0.9813

0.972

1

0.9905

0.9714

Precision

0.9811

0.9813

0.9694

1

0.9848

0.9516

Negative predictive value

0.9722

0.9813

0.8966

0

0.9905

0.9358

False positive rate

0.0187

0.0187

0.028

0

0.0095

0.0286

False discovery rate

0.0189

0.0187

0.0306

0

0.0152

0.0484

False negative rate

0.028

0.0187

0.1121

0

0.0152

0.1061

Accuracy

0.9766

0.9813

0.9299

1

0.9883

0.9415

F1 Score

0.9765

0.9813

0.9268

1

0.9848

0.9219

Matthews correlation coefficient

0.9533

0.9626

0.8629

1

0.9753

0.8763

False Discovery Rate, and False Negative Rate reached 0%. Naïve with bootstrapping did not achieve good prediction results as compared to the other prediction methods that used SMOTE. For example, the prediction sensitivity, specificity, precision, negative predictive, accuracy, F1 Score, and Matthews Correlation Coefficient values were less than their pairs with SMOTE method. Similarly, the false positive rate, false discovery rate, and false negative rate values were higher than their pairs in SMOTE method.

4 Discussion The initial objective of the project is to explore how SMOTE and bootstrapping methods repair imbalance data. It compared the performance of three prediction methods of SVM, Naïve Bayes, and neural networks with SMOTE and bootstrapping methods. Figure 2 shows that SMOTE generally performed poorer than bootstrapping with SVM and neural networks methods. This is because SMOTE does not consider the neighboring records that can belong to other classes which increase overlapped classes and add new noisy data. In contrast, bootstrapping is a straightforward way to redistribute data to become normal via calculating a statistic of interest such as mean. SVM with bootstrapping reached the best performance because of its ability to avoid the direct probability estimates and insistence that it relies on soft margin classification concept.

8

A. Aborujilah et al. 1.2 1 0.8 0.6 0.4 0.2 0 Sensivity Specificity Precision Negave Predicve Value

False Posive Rate

False False Accuracy Discovery Negave Rate Rate

SVM_with_SMOTE

Deep learning_with_SMOTE

Naive Bayes_with_SMOTE

SVM_with_bootstrapping

Deep learning _with_bootstrapping

Naivewith_bootstrapping

F1 Score Mahews Correlaon Coefficient

Fig. 2. Performance comparison of SMOTE and Bootstrapping with prediction methods

5 Conclusion The purpose of the current study is to compare the impact of samples redistribution methods such as SMOTE and bootstrapping on predication methods performance. So, SVM, NN, and Naïve Bayes prediction methods have been used. Overall, these results indicate that SMOTE generally performs poorer than bootstrapping with SVM and neural networks methods. This is because SMOTE does not take into account the neighboring records that can belong to other classes which increase overlapped classes and add new noisy data. This research has strengthened our understanding of how to optimise prediction processes with imbalanced datasets through employing the sample normalization methods. The major limitation of this study is that the evaluation was done by using a health dataset. Training other datasets may improve the understanding of this problem. More works need to be done to determine how to handle imbalanced datasets using other optimization methods such as ad-hoc heuristics-based methods. Further research should be carried out to establish the comparative approach to handle the imbalanced dataset problem including selecting the most efficient features and prediction methods.

References 1. Ebenuwa, S.H., Sharif, M.S., Alazab, M., Al-Nemrat, A.: Variance ranking attributes selection techniques for binary classification problem in imbalance data. IEEE Access 7, 24649–24666 (2019) 2. Longadge, R., Dongre, S.: Class imbalance problem in data mining review. arXiv Prepr. arXiv1305.1707 (2013)

Comparative Study of SMOTE and Bootstrapping Performance

9

3. López, V., Fernández, A., García, S., Palade, V., Herrera, F.: An insight into classification with imbalanced data: Empirical results and current trends on using data intrinsic characteristics. Inf. Sci. (Ny) 250, 113–141 (2013) 4. Meidan, Y., et al.: N-BaIoT—network-based detection of IoT botnet attacks using deep autoencoders. IEEE Pervasive Comput. 17(3), 12–22 (2018) 5. Nguyen, G.H., Bouzerdoum, A., Phung, S.L.: Learning pattern classification tasks with imbalanced data sets. In: Pattern recognition, IntechOpen (2009) 6. Luo, M., Wang, K., Cai, Z., Liu, A., Li, Y., Cheang, C.F.: Using imbalanced triangle synthetic data for machine learning anomaly detection. Comput. Mater. Contin. 58(1), 15–26 (2019) 7. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P.: SMOTE: synthetic minority over-sampling technique. J. Artif. Intell. Res. 16, 321–357 (2002) 8. Caminero, G., Lopez-Martin, M., Carro, B.: Adversarial environment reinforcement learning algorithm for intrusion detection. Comput. Netw. 159, 96–109 (2019) 9. Seiffert, C., Khoshgoftaar, T.M., Van Hulse, J.: Improving software-quality predictions with data sampling and boosting. IEEE Trans. Syst. Man, Cybern. Syst. Humans 39(6), 1283–1294 (2009) 10. Drummond, C., Holte, R.C.: “C4. 5, class imbalance, and cost sensitivity: why under-sampling beats over-sampling. In: Workshop on learning from imbalanced datasets II, vol. 11, pp. 1–8 (2003) 11. Liu, A., Ghosh, J., Martin, C.E.: Generative Oversampling for Mining Imbalanced Datasets. In: DMIN, pp. 66–72 (2007) 12. Huda, S., Yearwood, J., Jelinek, H.F., Hassan, M.M., Fortino, G., Buckland, M.: A hybrid feature selection with ensemble classification for imbalanced healthcare data: a case study for brain tumor diagnosis. IEEE Access 4, 9145–9154 (2016) 13. Team, A.V.C.: Practical guide to deal with imbalanced classification problems in R. Analytics Vidhya (2016) 14. Wang, Q., Luo, Z., Huang, J., Feng, Y., Liu, Z.: A novel ensemble method for imbalanced data learning: bagging of extrapolation-SMOTE SVM. Comput. Intell. Neurosci. 2017, (2017) 15. Liu, R., Hall, L.O., Bowyer, K.W., Goldgof, D.B., Gatenby, R., Ben Ahmed, K.: Synthetic minority image over-sampling technique: How to improve AUC for glioblastoma patient survival prediction. In: 2017 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 1357–1362 (2017) 16. Wijermans, N., Conrado, C., van Steen, M., Martella, C., Li, J.: A landscape of crowdmanagement support: an integrative approach. Saf. Sci. 86, 142-164 (2016) 17. Batista, G.E., Prati, R.C., Monard, M.C.: A study of the behavior of several methods for balancing machine learning training data. ACM SIGKDD Explor. Newsl. 6(1), 20–29 (2004) 18. Maciejewski, T., Stefanowski, J.: Local neighbourhood extension of SMOTE for mining imbalanced data. In: 2011 IEEE Symposium on Computational Intelligence and Data Mining (CIDM), pp. 104–111 (2011) 19. Stefanowski, J., Wilk, S.: Selective pre-processing of imbalanced data for improving classification performance. In: International Conference on Data Warehousing and Knowledge Discovery, pp. 283–292 (2008) 20. Liaw, A., Wiener, M.: Classification and regression by randomForest. R news 2(3), 18–22 (2002) 21. Lavanya, D., Rani, D.K.U.: Analysis of feature selection with classification: breast cancer datasets. Indian J. Comput. Sci. Eng. 2(5), 756–763 (2011)

UPLX: Blockchain Platform for Integrated Health Data Management Omar Musa1(B) , Lim Shu Yun1 , and Reza Ismail2 1 Faculty of Business and Technology, UNITAR International University,

47300 Petaling Jaya, Malaysia {omarm,lim_sy}@unitar.my 2 LedgerX International Sdn Bhd, T3-20-3A Icon City Trade Center, 47300 Petaling Jaya, Malaysia [email protected]

Abstract. Health data management currently needs a technology refresh in order to provide accurate, reliable and verifiable data for doctors and researchers to decide on the best medications and for the public to have their own dependable health information history as they continue with their daily lives. We propose UPLX (Unified Patient Ledger) as a blockchain-based data platform to securely record, “anonymize” and store patient health data for medical, academic and pharmaceutical research. UPLX is a blockchain powered health data platform which is designed to be interoperable and can be integrated with any hospital information systems (HIS) through API (Application Programming Interface) technology. To this end, a Hyperledger Fabric implementation is described to demonstrate the feasibility of the proposal and its use in healthcare organization. Successful implementation will accelerate the acceptance of Blockchain technology in protecting recorded health data while increasing the efficiency of healthcare delivery. Keywords: Blockchain · Interoperable · Medical ledger · Hyperledger fabric

1 Introduction The ongoing COVID-19 pandemic has exposed the need of a better data platform to manage patient health information [1]. The health care industry has transitioned from paper based to digital record keeping through the introduction of modern Electronic Medical Records (EMR) and Electronic Health Record (EHR) systems. Together, these two digital record keeping systems are able to chart a patient’s medical history and overall health. However, digital record keeping comes with its own share of advantages and disadvantages. While digital health records have proven to be cost effective, efficient and has greatly improved the accessibility of health information, recent security breaches has also introduced some ethical and security issues in managing patient confidentiality and privacy [2].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 10–18, 2021. https://doi.org/10.1007/978-3-030-70713-2_2

UPLX: Blockchain Platform for Integrated Health Data Management

11

In 2019, it was found that records for 14,000 HIV-Positive people in Singapore was leaked after a major cyber attack on Singapore’s health database [3]. In 2013, a former medical technician at Howard University Hospital in Washington U.S., pleaded guilty for violating the Health Insurance Portability and Accountability Act (HIPAA) by selling identifiable patient information to third parties [4]. A HIPAA Compliance Study found that 73% of physicians text other physicians about work [5]. Text messages, although fast and efficient can be easily accessed by third parties if the mobile device is not secured properly, misplaced, lost or stolen. According to a report from the Business Insider Intelligence, health care was exposed to more cybersecurity breaches than any other industry in 2018, accounting for 25% of 750 reported hacks. The numbers were particularly high in the U.S., where health firms suffered a record 365 data breaches in 2018 in comparison with 2017 s high of 358. During the hacks patients frequently lose their social security numbers, names and addresses. Sometimes, the information is more sensitive, such as health insurance information and medical histories [6]. UPLX or Unified Patient Ledger is a blockchain powered health data network designed to securely record and store patient consultation, prescription and treatment data. Data recorded in UPLX is securely “anonymized” using blockchain encryption methodology, enabling unprecedented access to patient data without exposing patient identity and confidentiality. UPLX is designed to be interoperable and easily integrated with any hospital system via API or data entry process. This work extends and improves on related works such as [7, 8] and [9]. In [7], the authors proposed the adoption of Blockchain technology as a disruptive technology for health data management. In [8], the authors proposed a secure electronic health record management on a public blockchain and the attendant consensus algorithm. In addition, simulations were performed to evaluate the scalability of their proposal instead of actual implementation. In [9], the authors proposed a model of a private consortium Blockchain model which assumes the presence of a cloud “Blockchain-As-A-Service (BaaS)” platform that integrates the off-chain and on-chain transactions. The remainder of this paper is structured in the following manner: In Sect. 2, we introduced the salient points of Blockchain Technology. In Sect. 3 we present the Hyperledger Fabric Blockchain model which utilizes permissioned blockchain network and how this is applicable for our UPLX implementation. In Sect. 4, we outlined the UPLX Interoperable Architecture and how it can be the foundation of our use case scenario of collaborations among relevant stakeholders. In Sect. 5, we presented the Data Structure Design and how the read and write processes of patient transactions are captured and stored. Finally, Sect. 6 concludes the paper.

2 Blockchain Technology The Blockchain concept was first published by a (fictional) person named Satoshi Nakamoto in 2009 in the form of a white paper about a peer-to-peer electronic cash system called Bitcoin [10]. It is a distributed P2P (Peer to Peer) ledger technology to process transactions in immutable blocks of data using cryptography. The blockchain is deployed as a distributed and decentralized network that processes, verifies and maintains (multiple copies of) its own data; autonomously.

12

O. Musa et al.

The blockchain records transactions in the form of an immutable ledger. It is deployed via a distributed network of untrusting peers, each maintaining a copy of the ledger [11]. Data is created in the blockchain in the form of a time stamped ledger which cannot be changed, updated or deleted in anyway because the provisions to do so simply does not exist. This data is therefore termed immutable. Its time stamped ledger also enables provenance, enabling us to trace the origin of the data and its evolution over time. Computable blockchains are able to execute pre-programmable code called smart contracts, which are preprogrammed code that can be written and deployed into the blockchain with rules to self-execute and self-enforce itself [9]. Smart contracts are transparent, run autonomously and once deployed cannot be changed or manipulated. The blockchain architecture then allows untrusting parties with common interests to co-create a permanent, unchangeable and transparent record of exchange and processing without relying on a central authority [12]. UPLX or Unified Patient Ledger is built on Hyperledger Fabric and designed as a blockchain platform to securely record and store patient data.

3 Hyperledger Fabric Hyperledger Fabric is a modular and extensible open source system for deploying permissioned blockchain networks [13]. Permissioned blockchain networks operate with a set of known, identified and verified participants. Public or permissionless blockchain networks such as Bitcoin or Ethereum allows anyone to participate in its network without verifying their identity. Public blockchains are usually deployed for cryptocurrency and usually require a consensus algorithm such as Proof of Work (POW) in Bitcoin coupled with fee-based incentives to ensure its transactions are made without the need for any centralized authority for verification. In such distributed, consensus environments, transactions are executed through an Order-Execute architecture (Fig. 1) in which it is broadcasted to all peers and it is executed sequentially [14].

Fig. 1. Public/Permissionless order-execute architecture

This results in some limitations: • Consensus has to be hard-coded within the platform • Transactions need to be broadcasted to all peers and executed sequentially • Smart contracts are programmed via a fixed, non-standard domain specific programming language and need to be run at all peers

UPLX: Blockchain Platform for Integrated Health Data Management

13

In a permissioned blockchain network such as Hyperledger Fabric, participants are groups of organizations that even though do not fully trust each other are able to exchange information and validate transactions because they share a common goal. Hyperledger Fabric executes transactions in a different Execute-Order-Validate (Fig. 2) architecture [14].

Fig. 2. Permissioned Execute-Order-Validate Architecture

Transactions are executed and endorsed first before they are ordered and validated in the chain. This resolves some of the limitations of a Order-Execute based public blockchain network: • All peers validate transactions, but not all peers need to execute it • Endorsement policies can be created and customized to determine which peers execute smart contracts • Executing transactions before ordering them allows peers to execute transactions in parallel • Smart contracts can be written using non-deterministic code such as Go, JAVA and Node.js

4 UPLX Interoperable Architecture UPLX is a blockchain powered interoperable health data platform and can be integrated with any hospital information systems (HIS) through API (Application Programming Interface) technology. To provide real-time data, UPLX can also be integrated with health tracking apps, wireless enabled wearables or IoT devices. Blockchain based data architecture is the leading candidate to enhance interoperability whether for existing application systems, IoT platforms or other smart devices because it is able to ensure security, privacy and performance [14]. UPLX is divided into two phases: A Write Phase and a Read Phase. In the Write Phase where each medical or health institution participating in UPLX network is represented as an “Organization” object, with rights to create and endorse transactions. The Write Phase (Fig. 3) provides tools for organizations to record their patient data via integrating UPLX APIs into their information systems. Their data is encrypted and patient information is anonymized before being recorded into the blockchain. UPLX anonymizes patient identity by applying an SHA-256 cryptographic hash function; utilizing information such as patient name and identity number combined with the organization’s private keys to create a unique representation of that data (Fig. 4). Since the private keys are unique to each organization, their patient information cannot be read by other organizations within UPLX network.

14

O. Musa et al.

Fig. 3. UPLX write phase

Fig. 4. Organization based “Anonymized” patient asset

The Read Phase (Fig. 5) enables recorded data to be accessed by third parties to perform various actions such as big data analytics, machine learning and Artificial Intelligence (AI). Read access to the UPLX network allows access to readable data structures which store health records within UPLX.

Fig. 5. UPLX read phase

UPLX: Blockchain Platform for Integrated Health Data Management

15

These APIs can be integrated into various analytical systems which enables access to data being recorded in real time from the health organization’s internal systems as the patient goes through their consultation processes. The access to the data may be permissioned with time restrictions as well as a subset of the data depending on the use cases scenarios. The validity and success of research studies, big data analytics and artificial intelligence activities depend a whole lot of the source data that it uses. The data should ideally come of its original source, unedited and untouched, for a valid and meaningful research and analysis to be performed.

5 UPLX Data Structure UPLX utilizes the Hyperledger Fabric model to map its health-based data structure design. UPLX data structure can be summarized into two main components, Assets and Transactions. In Hyperledger Fabric, an asset is defined as a collection of key-value pairs while a transaction are chaincode executables written to change or modify the state of an asset [14]. An asset in UPLX can be generalized as a binary representation of a patient’s identity. While transactions are a set of activities or actions that can be performed upon a patient. UPLX focuses on recording patient data and is designed specifically to integrate patient records and medical workflow into the blockchain (Fig. 6).

Fig. 6. UPLX data structure summary

UPLX transactions that directly affect the patient’s status are categorized into four main types: • • • •

Constant Physical Data (e.g. ethnicity, blood type) Variable Physical Data (e.g. height, weight) Location Data (e.g. city, state) Social Data (e.g. marriage status, number of children)

These categories are then paired with 8 medical workflows defined as transactions; or medical activities that affects or cause changes to the patient asset: • Consultations • Triage • Issues

16

• • • • •

O. Musa et al.

Tests Outpatient/Inpatient Prescriptions Procedures History

Each transaction has its own set of data structures which describes how it relates to and affects an asset (Fig. 7).

Fig. 7. Example - create patient data structure JSON API

A combination of APIs allow for any health and medical record system to integrate their workflows and participate in UPLX blockchain platform. Similar to any other blockchain platforms, each medical organization participating in UPLX generates their own private key which is used to encrypt their patient data. UPLX’s comprehensive data structure allows for a robust implementation of blockchain based health data records. Data structures are designed to enable tracking the health records of a person’s whole life, from simple medical consultations, inpatient treatments, health checkups to disbursement and consumption of medication and specific drugs.

UPLX: Blockchain Platform for Integrated Health Data Management

17

There is great potential of UPLX use cases in areas of drug research and traceability, clinical trials, disease tracking and even indirect processes such as health insurance claims.

6 Conclusion As blockchain technology continues to gain more exposure, the adoption of blockchain technology into industry applications is something that should be looked into to fully reap its benefits. Many blockchain projects are kicking into high gear as researchers and practitioners continue to experiment with its capabilities and limits. The blockchain at its core, generates a set of secure, immutable and trusted data. Data is the foundation of research and analysis in many areas of industry and academia. The validity, accuracy and reliability of any research stems from the trustworthiness of its data source. We believe our blockchain powered UPLX platform as described in this paper is able to function as a unified platform to securely consolidate, record and store patient medical data. It is able to operate as a trusted source of health data and anonymously distribute health records for the purpose of improving the reliability, accuracy and validity of research and development of medications and vaccines, while protecting the patient’s confidentiality and privacy. In addition, by having a permissioned blockchain architecture, a consensus algorithm is not needed thus increasing the scalability and efficiency of UPLX. The UPLX is currently undergoing prototyping and deploying to pilot institutions that have indicated willingness to participate. Other than finetuning the prototype based on feedbacks from the pilot, future work will focus on developing an improved dynamic and distributed trust model for blockchain based self -sovereign identity management. This will enhance the security and privacy requirements of autonomous users of UPLX. Acknowledgment. The authors acknowledge the support by Malaysia’s Fundamental Research Grant Scheme under FRGS/1/2018/ICT04/UNITAR/03/1.

References 1. Hamzah, F.A., et al.: CoronaTracker: Worldwide COVID-19 Outbreak Data Analysis and Prediction. World Health Organisation, Bull cE-publication (2020) 2. Ozair, F., Jamshed, N., Sharma, A., Aggarwal, P.: Ethical issues in electronic health records: a general overview. Perspect. Clin. Res. 6(2), 73 (2015) 3. Reuters. https://www.reuters.com/article/us-singapore-health/us-finds-american-guilty-insingapore-hiv-data-leak-case-idUSKCN1T709J. Accessed 20 June 2020 4. FBI. https://archives.fbi.gov/archives/washingtondc/press-releases/2012/former-howarduniversity-hospital-employee-pleads-guilty-to-selling-personal-information-about-patients. Accessed 20 June 2020 5. Greene, A.H.: HIPAA compliance for clinician texting. J. AHIMA 83(4) 34–36 (2012) 6. Healthcare Compliance Analytics-Protenus. https://www.protenus.com/

18

O. Musa et al.

7. Magyar, G.: Blockchain: solving the privacy and research availability tradeoff for EHR data: a new disruptive technology in health data management. In: IEEE 30th Neumann Colloquium (NC), Budapest, pp. 000135–0001402017 (2017) 8. de Oliveira, M.T. et al.: Towards a Blockchain-based secure electronic medical record for healthcare applications. In: ICC 2019-2019 IEEE International Conference on Communications (ICC), Shanghai, China, pp. 1–6 (2019) 9. Theodouli, A., et.al.: On the design of a blockchain-based system to facilitate healthcare data sharing. In: 2018 17th IEEE International Conference On Trust, Security And Privacy In Computing And Communications/12th IEEE International Conference On Big Data Science And Engineering (TrustCom/BigDataSE), New York, NY, pp. 1374–1379 (2018) 10. S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash system”, 28 (2008). 11. Androulaki, E., Barger, A., Bortnikov, V., et al.: Hyperledger Fabric: A Distributed Operating System for Permissioned Blockchains. [4] (2018) 12. Levi, S.D., et.al.: An Introduction to Smart Contracts and Their Potential and Inherent Limitations, [5] (2019) 13. IBM Research Group. https://www.ibm.com/blogs/research/2018/02/architecture-hyperl edger-fabric “Behind The Architecture of Hyperledger Fabric”. Accessed 20 June 2020 14. Dorri, A., Kanhere, S.S., Jurdak, R., Gauravaram, P.: Blockchain for IoT security and privacy: the case study of a smart home. In: IEEE International Conference on Pervasive Computing and Communications Workshops, pp. 618–623 (2017)

Convolutional Neural Networks for Automatic Detection of Colon Adenocarcinoma Based on Histopathological Images Yakoop Qasim(B) , Habeb Al-Sameai, Osamah Ali, and Abdulelah Hassan Department of Mechatronics and Robotics Engineering, Taiz University, Taiz, Yemen

Abstract. Colorectal cancer is the second type of cancer that causes death and the third in terms of prevalence and number of cases. Due to the absence of symptoms in the early stages of the injury, several types of tests must be performed to discover the cancer, but these methods take a lot of time, cost and require a specialized expert. So in this paper, we proposed a Convolutional Neural Network (CNN) model that characterized by speed of diagnosis and high accuracy with few number of parameters for diagnosing colon adenocarcinoma since it is the most common of colorectal cancer, where it represents 95% of the total cases of colorectal cancer, depending on dataset of 10000 histopathological images divided into 5000 images for colon adenocarcinoma and 5000 images for benign colon. Our model consists of two paths each path is responsible for creating 256 feature maps to increase the number of features at different level in order to improve the accuracy and sensitivity of the classification. To compare the performance of the proposed model, Visual geometry Group (VGG16) model was prepared and trained on the same dataset. After training the two models we obtained an accuracy of 99.6%, 96.2% for the proposed model and VGG16 respectively, we also obtained from the proposed model a sensitivity of 99.6% and Area Under Curve (AUC) of 99.6% which indicates the effectiveness of this model in diagnosing colon adenocarcinoma. Keywords: Deep learning · Colorectal cancer · Convolutional neural networks

1 Introduction Colorectal cancer is a cancer that arises in the colon and rectum, the colon is also known as the large intestine, while the rectum is the last part of the colon. According to the World Health Organization (WHO), colorectal cancer is the second most common type of cancer leading to death and the third at the most common cancer cases list, in 2018 [1]. There are many types of colorectal cancer such as adenocarcinoma, carcinoid tumors, gastrointestinal stromal tumors and colorectal lymphoma, whereas adenocarcinoma is the most common type of colorectal cancer and it represents about 95−98% of the total cases of colorectal cancer [2]. There are various symptoms of colorectal cancer such as constipation, diarrhea, changes in stool color, blood in the stool and bleeding from the rectum, often these symptoms do not appear on the patient in the early stages, and here lies the danger of colorectal cancer [2]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 19–28, 2021. https://doi.org/10.1007/978-3-030-70713-2_3

20

Y. Qasim et al.

To diagnose colorectal cancer, doctors use a Colonoscopy which is a long, thin and flexible tube attached to a camera and screen to view inside the colon and rectum, if doctors found any suspected areas with the disease, surgical instruments can be inserted through the Colonoscopy to take tissue samples (biopsies) [3], but the histopathological diagnosis requires an expert and a pathologist who is able to distinguish the size of cell nuclei as well as the shape of cells and where they are in the tissue, which may lead to an increasing in the diagnostic time and cost. With artificial intelligence algorithms, diagnosis process becomes easier as it possible to use deep learning algorithms to diagnose colorectal cancer based on the histopathological images at a faster speed and lower cost, there are many studies in this field. In [4] the authors used two approaches for transfer learning which are (i) CNN as a fixed feature generator, in this approach the CNN model which is VGG16 [5] is used to extract features and then fed them into a separate machine learning algorithm to complete the classification process, (ii) Fine-tuning the CNN in this approach the low layers of VGG16 are fixed, and the top layers of VGG16 are changed to adapt the new classification task. The authors training and testing their models on a dataset of 13500 whole-slide images of colorectal tissues distributed over three classes adenocarcinoma, tubuvillous adenoma and healthy tissue, they obtained an accuracy of 96%. In [6] the authors used different strategies of transfer learning by changing the number of layers which are frozen, they used four popular models for training and testing a dataset of 1577 confocal laser microscopy images speared over four classes healthy colon, malignant colon, healthy peritoneum and malignant peritoneum, they obtained Area under curve of 97.1% for classification metastases in the peritoneum. One of the disadvantages of this paper was that the dataset used was small and belonged to rats. In [7] a CNN model consisting of 43 convolution layers was presented to classify three classes which are adenocarcinoma, adenomatous polyps and normal on images, after training the model it was tested on 410 images achieving an accuracy of 94.4%. In [8] the authors used two approaches which are traditional approaches and transfer learning approach. In the traditional approach they used 5 state-of-the-art feature extraction techniques followed by an Support Vector Machine (SVM) classifier, while in the second approach a CNN model which is InceptionV3 [9] used as feature extraction and classifier. Both approaches are used to classify four classes normal, hyperplastic polyps, tubular adenoma and carcinoma based on histology images consists of 4000 images at rate 1000 images for each class. The best accuracy and sensitivity are 94.5% and 95.21% respectively. The main contributions of this paper is a CNN model has the ability of diagnosis colon adenocarcinoma based on histopathological images with high accuracy, as this model was trained on a large number of data, which gives reliability to the performance of the model, in addition this model is characterized by having a small number of parameters which can be used in any platform or framework since it does not need a large storage space.

Convolutional Neural Networks for Automatic Detection

21

2 Methods and Dataset 2.1 Dataset The dataset used in this work was extracted from LC25000 dataset which is available online [10]. The folder of colorectal cancer consists of two subfolders, the first subfolder is colon_aca with 5000 histopathological images of colon adenocarcinoma, and the second subfolder is colon _n with 5000 histopathlogical images of benign colonic tissue. Before training we divided the dataset into 70% for training and 30% for validation. 2.2 Convolutional Neural Networks A Convolutional Neural Network (CNN) is a type of neural networks that specializes in image classification and computer vision tasks, a typical CNN architecture has a convolution layer which extracts the features from the input array by applying a different filters on the input array pixels producing what is known as a convolved feature map [11], a pooling layer which reduces the size of feature map to make the computational easier, a Rectified Linear Unit (ReLU) which is an activation function can be represented as f(x) = max(0,x), the ReLU function accelerates the training speed and solves the vanishing gradient problem [12], a fully connected layer which takes the outputs of the previous layers, turn them into a single vector and gives the predictions for each class. CNN architecture can be divided into two parts, the first part consists of convolution layer, pooling layer and ReLU, is responsible for extraction the features, the second part consists of fully connected layer, is responsible for classification tasks. Figure 1 shows the process of extracting the features and creating feature maps by the convolution layer, and reducing the size of feature map by the max pooling layer.

Fig. 1. The process of extracting and reducing the feature maps.

2.3 Proposed Model As is known that increasing the depth of the model improves the accuracy of classification [13]. However, this is may not apply to medical diagnostics, as increasing the depth of

22

Y. Qasim et al.

the model may not improves its performance in the biological images [14], instead, the number of feature maps must be increased in order to include the most important and necessary details of the diagnostic process, therefore, we present the proposed model which consists of two paths in order to obtain the most high-level and low-level features with a suitable depth to diagnose medical images, and has a low number of parameters for reducing the computational resources and the time consuming in training. Table 1. The architecture of the proposed model for each path. Layer

Input feature

Stride

Padding

Output feature

Input layer

50 × 50 × 3

−

−

50 × 50 × 3

Parameters 0

Conv1

50 × 50 × 3

1

Same

50 × 50 × 32

896

Max pooling

50 × 50 × 32

2

−

25 × 25 × 32

0

Drop out

25 × 25 × 32

−

−

25 × 25 × 32

0

Conv2

25 × 25 × 32

1

Same

25 × 25 × 64

18,496

Max pooling

25 × 25 × 64

2

−

12 × 12 × 64

0

Drop out

12 × 12 × 64

−

−

12 × 12 × 64

0

Conv3

12 × 12 × 64

1

Same

12 × 12 × 128

Max pooling

12 × 12 × 128

2

−

6 × 6 × 128

0

Drop out

6 × 6 × 128

−

−

6 × 6 × 128

0

73,856

Conv4

6 × 6 × 128

1

Same

6 × 6 × 256

295,168

Max pooling

6 × 6 × 256

2

−

3 × 3 × 256

0

Drop out

3 × 3 × 256

−

−

3 × 3 × 256

0

GAP

3 × 3 × 256

None

Valid

256

−

Drop Out

1024

Dense2

1024

Dense1

256

0

−

1024

263,168

−

−

1024

0

−

−

2

2050

As shown in Table 1 and Fig. 1 our model consists of two paths, each path consists of four blocks and each block consists of combination of convolution + ReLU, max pooling layer and dropout layer [15] which prevents over-fitting [16], the two paths of the proposed model is followed by Global Average Pooling which creates a feature map for each category in the last layers [17]. The input layer of the proposed model is fixed with the size 50 × 50 × 3 pixels, the kernel size of all convolution layers is 3 × 3 because this size is the smallest possible size that can extract the features left/right and up/down, and it is also known that in medical diagnostics based on histopathological images, the smallest details are important to the success of the classification process, the filters number that applied in the first, second, third, fourth convolution layers are 32, 64, 128, 256 respectively. The output or dense layer consists of two neurons and the activation function was applied on it is Softmax function. Also as shown in Table 1 each path is responsible for creating 256 feature maps at different levels, and the total number of the

Convolutional Neural Networks for Automatic Detection

23

parameters is 653,634 and this is a small number compared to the number of parameters of popular models, which leads to a decrease in the time required for training. To show the effective of the proposed model, the VGG16 model was prepared and trained on the same dataset, VGG16 was chosen because it is one of the most commonly used model for image recognition, and also due to its simple structure (Fig. 2).

Fig. 2. The architecture of the proposed model

2.4 Transfer Learning and Fine-Tune Transfer learning is a technique of taking weights learned one problem and applying them to a new, similar problem [18], while fine-tune is the way of applying or utilizing transfer learning and adapting the pre-trained models to our classification task. In our work, we have used the transfer learning to re-train the VGG16 model, by keeping the weights of low layers or blocks fixed and fine-tuning the weights of the top layers. Table 2 shows the hyper-parameters for the two models. Table 2. Hyper-parameters for the two models. Hyper-parameters Value Batch size

32

Epochs

30

Image sized

50 × 50

Optimizer

Adam

Loss function

Categorical cross-entropy

Learning rate

1e−3

3 Results After training the proposed model we got Training and Validation curves for an accuracy and loss as shown in the figures below (Figs. 3 and 4).

24

Y. Qasim et al.

Fig. 3. Accuracy curves for the proposed model

Fig. 4. Loss curves for the proposed model.

For evaluating the performance of our model and other model, we calculated the Confusion matrix, and plotted the Receiver Operating Characteristics (ROC) so we can calculate the Area Under Curve (AUC). Confusion Matrix By using confusion matrix, we can calculate the evaluation metrics based on four parameters which are True Positive (TP) which is the correctly predicts of colon adenocarcinoma cases, False Positive (FP) which is the cases of benign colon that were classified as colon adenocarcinoma, True Negative (TN) which is the cases of benign colon and were classified as benign colon, False Negative (FN) which is the cases of colon adenocarcinoma that were classified as benign colon, Fig. 5 shows the confusion matrix for

Convolutional Neural Networks for Automatic Detection

25

the two models, Table 3 shows the four parameters for the two models, the evaluation metrics which can found by the four parameters are accuracy, sensitivity, specificity, precision and F-Measures, As the sensitivity is the percentage of cases with colon adenocarcinoma were correctly classified, specificity is the percentage of cases with benign colon were correctly classified, precision is the percentage of cases that actually belong to the colon adenocarcinoma cases from all cases that were classified as colon adenocarcinoma, and F-measure is the harmonic mean of sensitivity and precision. Table 4 shows the evaluation metrics for the two models.

Fig. 5. Confusion Matrix for the Proposed Model (left) and for the VGG16 Model (right).

Table 3. The four parameters for the two models. Model

TP

VGG16

1494

FP TN 6 1494

FN 6

The proposed 1435 49 1451 65

Table 4. Evaluation metrics for the two models. Model

Accuracy

Sensitivity

Specificity

Precision

F-Measure

VGG16

96.2

95.67

96.73

96.7

96.18

The proposed

99.6

99.6

99.6

99.6

99.6

TP + TN TP + TN + FP + FN

(1)

Sensitivity =

TP TP + FN

(2)

Specificity =

TN TN + FP

(3)

Accuracy =

26

Y. Qasim et al.

Precision = F − Measure =

TP TP + FP

2 × Senstivity × Precision Senstivity + Precision

(4) (5)

ROC and AUC ROC curve is a 2-D graph shows the performance of the classification model at all the classification thresholds [19], it can also be defined as the trade-off between sensitivity and specificity [20]. ROC is plotted between the True Positive Rate (TPR) or sensitivity and False Positive Rate (FPR) or 1-specificity, we have plotted ROC curve for the two models so we can calculate AUC, where AUC measures the area under the ROC curve and ranges from zero to one, if the predictions of the model are completely true then AUC is one, and if the predictions of the model are completely false then AUC is zero, whereas the higher the value of AUC means the lower the values of FP and FN. In medical diagnostic, the value of AUC is closer to one is required [22, 23]. Figure 6 shows ROC and AUC from the figure we obtained AUC of 99.6% for the proposed model and 96.2% for the VGG16 model.

Fig. 6. ROC and AUC for the two models.

4 Discussion and Conclusion In this work, we proposed a deep learning model consisting of two paths to diagnose colon adenocarcinoma based on the histopathological images with low number of parameters in order to reduce the computational resources and time of training. The proposed model was tested on 3000 images and performed exceptionally well. Our model achieved an overall accuracy, sensitivity, specificity, precision and F-Measure of 99.6% for the all

Convolutional Neural Networks for Automatic Detection

27

metrics outperforming the VGG16 model, from Table 4 we note that the sensitivity is very high, which means the model is very sensitive to images of colon adenocarcinoma and suitable to be used as a diagnostic colon cancer. Our model also achieved an AUC of 99.6%, this value is considered to be perfect in the field of medical diagnosis, it should be noted that higher AUC means low FP and FN cases, and low FP and FN cases means better classification and perfect diagnostic results. From the above we conclude that the models which are built in this way are very effective in medical diagnosing based on histopatological images. This model has one drawback or limitation, which that is not trained on all types of colorectal cancer, the focus has been on colon adenocarcinoma, as it constitutes the vast majority of people with colorectal cancer. In the future we aspire to train the model on all types of colorectal cancer and create a diagnostic platform or framework. In Table 5 we compared between the result obtained from the previous studies and the results which we obtained. Table 5. Compression between the previous studies and our study. Study

Dataset used

Approaches-Models

Results

Francesco et al. [4]

Consists of 13500 Whole-slide-images(WSIs)

i-CNN as feature generatorVGG16 + SVM ii-Fine tuning -VGG16

Accuracy = 96%

Nills et al. [6]

Confocal laser microscopy (CLM) which consist of 1577 images

InceptionV3 Densenet121 SE-Resnext50 & VGG-16

AUC = 97.1%

Hyun et al. [7]

Consists of 49458 of endoscopy images

CNN consists of 43 layers

Accuracy = 94.39%

Junaid et al. [8]

QU-AHLI which Consists of i- Traditional approach4000 images rLPQ + SVM rlbp + svm Uniform Rlbp + SVM Haralick + SVM (rLPQ + rLbp) + SVM ii- Transfer learning approach-InceptionV3

Accuracy = 94.4% AUC = -

This study

LC25000

Accuracy = 99.6% AUC = 99.6%

The proposed model

References 1. WHO Cancer-World Accessed 28 Jul 2020

Health

Organization,

http://www.who.int/health-topics/cancer.

28

Y. Qasim et al.

2. Cancer.org, What is Colorectal cancer, http://www.google.com/amp/s/amp.cancer.org/cancer/ colon-rectal-cancer/about/what-is-colorectal-cancer.html Accessed 2 Aug 2020 3. Christina Chun, Colorectal cancer: Symptoms, treatment, risk factors, and causes. http://www. medicalnewstoday.com/articles/155598 Accessed 2 Aug 2020 4. Ponzio, F., Enrico, M., Elisa, F., Santa, D.: Colorectal Cancer Classification Using Deep Convolutional Networks, In: Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies, vol. 2, (2018) 5. Simonyan, K., Andrew, Z.: Very Deep Convolutional Networks for Large-Scale Image Recognition, arXiv preprint arXiv. pp. 1409–1556 (2014) 6. Streiner, D.L., John, C.: What’s under the roc? An introduction to receiver operating characteristics curves. Can. J. Psychiatry 52(2), 121–128 (2007) 7. Gessert, N., Marcel, B., Lukas, W., Daniel, D.: Deep transfer learning methods for colon cancer classification in confocal laser microscopy images. Int. J. Comput. Assis. Radiol. Surg. 14(11), 1837–1845 (2019) 8. Park, H., Yoon, K., Sang, L.: Adenocarcinoma recognition in endoscopy images using optimized convolutional neural networks. Appl. Sci. 10(5), 1650 (2020) 9. Malik, J., Serkan, K., Suchitra, K., Turker, I., Somaya, A., Ridha, H., Moncef, G.: Colorectal Cancer Diagnosis from Histology Images, A Comparative Study, arXiv preprint arXiv. pp. 1903–11210 (2019) 10. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z.: Rethinking the inception architecture for computer vision. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2818–2826 (2016) 11. Borkowski, A., Marilyn, M., Brannon, T., Catherine, P., Lauren, A., Stephen, M.: Lung and Colon Cancer Histopathological Image Dataset (Lc25000), arXiv, preprint arXiv, pp. 1912– 12142 (2019) 12. Pinaya, W., Garcia-Dias, S., Mechelli, A.: Convolutional neural networks. https://dpi.org/10. 1016/B978-0-12-815739-8.00010-9 13. Ide, H., Takio, K.: Improvement of Learning for CNN with ReLU Activation by Sparse Regularization. In: 2017 International Joint Conference on Neural Networks (IJCNN), pp. 2684–2691. IEEE (2017) 14. LeCun, Y., Yoshua, B.: Geoffrey H. Deep Learn. Nat. 521(7553), 436–444 (2015) 15. Min, S., Byunghan, L., Sungroh, Y.: Deep learning in bioinformatics. Briefings Bioinf. 18(5), 851–869 (2017) 16. Hinton, G., Srivastava, N., Krizhevsky, A.: Improving Neural Networks by Preventing CoAdaption of Feature Detectors [R/Ol], (2015) 17. Hawkins, D.M.: The problem of overfitting. J. Chem. Inf. Comput. Sci. 44(1), 1–12 (2004) 18. Chris. What are Max Pooling, Average Pooling, Global max Pooling and Global Average Pooling. https://www.machinecurve.com/index.php/2020/01/30/what-are-max-poolingaverage-pooling-global-max-pooling-and-global-average-pooling/. Accessed 27 Jun 2020 19. Cheng, B., Liu, M., Zhang, D., Munsell, B.C., Shen, D.: Domain Transfer Learning for MCI conversion Prediction. IEEE Trans. Biomed. Eng. 62(7), 1805–1817 (2015) 20. Developers.google. Classification: ROC curve and AUC. https://developers.google.com/mac hine-learning/crash-course/classification/roc-and-auc. Accessed 16 Aug 2020 21. Korsten, MA.: Application of Summary Receiver Operating Characteristics (Sroc) Analysis to Diagnostic Clinical Testing. In: 7th Reflections on the Future of Gastroenterology–unmet Needs vol. 52, p. 76, (2007) 22. Streiner, D.L., Cairney, J.: What’s under the roc? an introduction to receiver operating characteristics curves. Can. J. Psychiatry 52(2), 121–128 (2007) 23. Siddiqui, M.K„ Morales-Menendez, R., Ahmad, S.: Application of Receiver Operating Characteristics (Roc) on the Prediction of Obesity. Brazilian Arch. Biol. Technol. 63, (2020)

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare Using Machine Learning Radiah Haque1 , Sin-Ban Ho1(B) , Ian Chai1 , Chin-Wei Teoh1 , Adina Abdullah2 , Chuie-Hong Tan3 , and Khairi Shazwan Dollmat1 1 Faculty of Computing and Informatics, Multimedia University, 63100 Cyberjaya, Malaysia

{sbho,ianchai,shazwan.dollmat}@mmu.edu.my 2 Department of Primary Care Medicine, Faculty of Medicine, University of Malaya,

50603 Kuala Lumpur, Malaysia [email protected] 3 Faculty of Management, Multimedia University, 63100 Cyberjaya, Malaysia [email protected]

Abstract. Enhancing personalisation is important for productive collaboration between humans and machines. This is because the integration of human intelligence with cognitive computing would provide added value to healthcare. While the well-being and human health can be profoundly affected by weather, the effect of machine learning on personalised weather-based healthcare for selfmanagement is unclear. This paper seeks to understand how machine learning use affects the personalisation of weather-based healthcare. Based on the Uses and Gratifications Theory (UGT), new constructs are incorporated (demography, weather and effectiveness) in order to propose a model for health science with machine learning use, weather-based healthcare, and personalisation. Subsequently, this paper proposes building a system that can predict the symptoms of two diseases (asthma and eczema) based on weather triggers. The outcome from this paper will provide deeper understanding of how personalisation is impacted by machine learning usage and weather-based healthcare for individual patients’ self-management and early prevention. The findings in this paper will also assist machine learning facilitators design effective use policies for weatherbased healthcare that will have new fundamental knowledge with personalisation to enhance the future of intelligent health informatics, and artificial intelligence. Keywords: Machine learning · Intelligent health informatics · Artificial intelligence · Weather-based healthcare · Mobile application

1 Introduction Weather-based healthcare, which refers to self-management of chronic diseases that are affected by the weather, helps patients avoid weather triggers that can worsen their symptoms by changing their lifestyle. However, it is difficult for individual patients © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 29–40, 2021. https://doi.org/10.1007/978-3-030-70713-2_4

30

R. Haque et al.

to change their lifestyle for self-management and early prevention of worsening their disease symptoms based on the weather, because many of them are unaware of which weather triggers they are vulnerable to, which may depend on their demographic characteristics (e.g. age and disease severity level). Many weather-based healthcare systems for self-management fail to be adopted widely because these systems lack the capability to support personalisation (i.e. providing feedback based on individual patients’ demographic characteristics and weather triggers). Fortunately, machine learning can solve this problem because it teaches computers to learn from relevant weather and demographic data of individual patients and provide personalised prediction results. Machine Learning is a scientific study that involves statistical models and algorithms. It is used to implement tasks by systems without specific instructions, but by relying on inference and patterns instead [1]. In other words, machine learning is a subset of data science and artificial intelligence, where computers learn through algorithms based on implications of predictive power from a set of data [2]. Thereby, machine learning works in systems that can discover and learn from data patterns, and use them to make independent decisions. Consequently, the application of machine learning is driven by the availability of large amounts of data and lower cost computation. By using this technology, a lot more topics can be researched and can produce results or decisions that are more accurate and useful to the community. One of the topics is healthcare where machine learning techniques have made advances in the healthcare domain. This is because healthcare data provides vast opportunities for development of learning the patterns between individual patient’s symptom triggers, medical history and demographic characteristics, which then assists in processing automation and personalised predictions. Meanwhile, well-being and human health can be profoundly affected by the weather. The weather may also be associated with allergies and respiratory diseases, which can be possibly linked to concentrations of pollution levels and pollen grains [3]. However, the machine learning impact on personalised weather-based healthcare is still unclear at the moment. As compared to the effectiveness of machine learning techniques in other commercial domains, such as face and image recognition, weather-based analytics and prediction for self-management still lags behind. Therefore, this paper is to explore further the relationship between weather conditions and chronic diseases that are affected by weather. Two diseases have been considered for this study; asthma and eczema. Furthermore, a model of intelligent health informatics with personalised healthcare will be proposed using machine learning techniques to predict asthma attack and worsening of eczema for individual patients based on weather triggers. This paper will benefit the information and communication technologies and healthcare sectors. The outcome of this paper will assist machine learning developers and researchers design effective use policies for weather-based healthcare, and provide invaluable feedback to government healthcare systems.

2 Background Study 2.1 Influence of Weather on Asthma and Eczema The influence of weather and climate changes is critical to patients who have chronic asthma. The study by Alharbi and Abdullah [4] has stated that asthma attacks are affected

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare

31

by the changes of weather temperature and humidity. Another study by Asthma and Allergy Foundation of America (AAFA) [5] has elaborated that increased humidity in the atmosphere and thunderstorms can trigger asthma attacks. This is because humidity helps dust mites thrive and the number of pollen particles increases in the air which can aggravate asthma. Heavy rain and strong wind caused by thunderstorms can break pollen grains into smaller sizes, which makes them transmit more easily through the air. It has a critical effect as asthma patients inhale the polluted pollen-laden air into their lungs. One of the significant incidents that reflect the influence of humidity and thunderstorms for asthma triggers is the 2016 Melbourne, Australia event when a thunderstorm affected thousands of asthma patients living in the city. In fact, this particular event has been labelled “thunderstorm asthma” [6]. Moreover, analysis of 15,678 asthmatic hospital admissions in Shanghai, China found that cold temperatures can trigger attacks in asthma [7]. Eczema, or skin allergies, is a skin barrier dysfunction which causes skin dryness, itchiness and irritation. Since skin is mostly exposed to the environment, the skin barrier is important to block allergens and other germs in the air from entering the skin surface. Eczema causes the skin to lose the ability to adapt to climate changes [8]. Vocks et al. [9] found that patients with eczema suffer more itchiness and irritation in winter than in summer, and during thunderstorms, due to colder temperatures and windy conditions. The increased pollen grains in the air due to winds enter the skin cells and cause dryness and irritation. Respiratory health and skin diseases affect the general population around the world, and the level of severity varies from patient to patient based on different weather conditions. Thus, weather is an important factor that must be monitored by individual patients with asthma and eczema. 2.2 Methods for Patients to Self-monitor Asthma and Eczema For personalised healthcare, patients need to perform control tests enabling them to conduct self-monitoring on the seriousness of their asthma or eczema on their own from their location. The AAFA has introduced the Asthma Control Test (ACT) as the standard test to monitor asthma. The ACT is recommended by medical experts [5]. Asthma patients can use the ACT to identify the severity of their asthma, which, in turn, is useful for doctors and nurses to determine the required treatment. The ACT has a scaling index for patients to record the severity of their asthma easily. Meanwhile, Charman et.al [10] has suggested the Patient-Oriented Eczema Measure (POEM) as the standard assessment to identify the severity of eczema. The POEM assessment has a scaling index for patients to record the severity of their eczema easily. The National Institute for Health and Care Excellence (NICE) at the University of Nottingham [11] recommends the use of POEM in clinical guidelines. In eczema trials, the HOME (Harmonizing Outcome Measures for Eczema) initiative recommends POEM to be used as the essential instrument to measure patient-reported symptoms of skin allergies. Thus, both recommended methods allow patient self-monitoring for asthma and eczema. The score obtained on both tests helps medical personnel to classify the severity and determine the effective treatment required for their patients effectively and efficiently. In addition to the wide use of ACT and POEM for self-monitoring asthma and

32

R. Haque et al.

eczema respectively, this paper aims to relate them to weather. This can allow patients to identify the triggers of their disease based on the weather conditions in their location. 2.3 Machine Learning in Weather-Based Healthcare The effect of weather conditions on asthma and eczema necessitated further research into how computers can assist personalised healthcare awareness for self-management, and provide assessment of temperature, exposure to allergens, changes in barometric pressure, humidity, and wind. Information extraction and machine learning promise lower costs, besides being capable of discovering patterns in large amounts of data, dealing with uncertainties and probabilities. The working of a machine learning application is completely different from a regular application [12]. Figure 1 illustrates how machine learning fits with weather-based healthcare under data science [13]. In the context of information extraction and machine learning, there is a need to revisit regression analytics challenges for two reasons. Firstly, machine learning relies extensively on statistical methods. The ultimate goal is to model real-world weather-based healthcare system, with mathematical relationships. Secondly, the lack of standards for measuring the relationships among the variables makes utilising the most effective method even harder. The many capabilities for regression are based on training data uniformity [14], making fitting the regressions to non-uniform data challenging.

Fig. 1. Machine learning in weather-based healthcare.

In recent years, with the expansion of computer-assisted systems in health, attention is driven on developing tools such as mobile health (mHealth) applications for selfmanagement by providing early prevention using machine learning. Table 1 summarises examples of machine learning use in weather-based healthcare for self-management of asthma and eczema, and highlights the limitations of these proposed models. From the background study, it was identified that implementing a weather-based healthcare system with personalisation support for self-management has proven to be a challenge. This is because it is difficult to identify the impact of weather on an individual patient’s symptoms, since weather attributes affect each patient differently based on demographic characteristics and severity level. Fortunately, machine learning algorithms, such as neural networks, can be developed to predict the impact of weather on an individual

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare

33

patient’s symptoms and provide personalised feedback for self-management and early prevention of asthma attacks and worsening eczema symptoms. Table 1. Machine learning use in weather-based healthcare. Ref.

Contribution

Limitation

[15]

Proposed machine learning technique for an mHealth application for predicting and providing early prevention of asthma attacks. The model can provide real-time feedback to patients based on weather conditions in the user’s location

The proposed model does not include collecting demography data and does not provide personalised feedback to users. Moreover, there is a need to classify the severity level of asthma for individual patients

[16]

Developed a prototype for an mHealth application which contains knowledge generation via machine learning to sense the surrounding environment and weather conditions in the user’s location. The model can predict the weather triggers for asthmatic patients and provide feedback and early prevention of asthma attacks

The model does not propose identifying the severity level of asthma for individual patients. Furthermore, the proposed model does not provide personalised prediction based on individual user’s demographic characters and weather triggers

[17]

Developed a machine learning model to predict eczema severity level on a daily basis. The proposed model is design through Bayesian inference to provide probabilistic prediction to patients with eczema for early prevention of worsening symptoms

The proposed model does not identify the impact of weather attributes on individual patient’s eczema severity level and does not collect demography data for personalised feedback

3 Methodology This paper aims to propose a weather-based healthcare system that can predict chances of triggering asthma attacks or worsening eczema symptoms of its users based on daily weather forecasts for their location using machine learning techniques. The conceptual representation of the proposed system in Fig. 2 presents how a machine learning model for developing algorithms can be used to provide recommendations for patients based on the current weather forecast [18, 19]. These machine learning algorithms shall predict how weather conditions affect a patient’s asthma or eczema. The algorithms and predictions [20] are based on the analysis of user-reported data and the weather forecast. Consequently, the system can continuously improve the accuracy of the prediction of these diseases and the algorithms through machine learning [21]. Typically, a machine learning process includes collecting data, preparing the data and applying algorithms to train and test the data [13]. In order to collect the data, a weatherbased mHealth application with a user personalisation feature is developed. The mHealth application provides the latest weather forecast to help users with asthma and eczema

34

R. Haque et al.

estimate the chances of asthma triggers or to measure the severity level of their eczema based on the weather conditions of the day from their location. Table 2 summarises the core functions of the weather-based mHealth application, while Fig. 3 illustrates the data flow between the user and the system. Figure 4 shows the Entity Relationship Diagram (ERD) of the weather-based mHealth application. In this case, the system comprises of four entities including UserInfo to store user profile and demography data, ReportDiseaseAsthma and ReportDiseaseEczema to store user results of the ACT and POEM respectively, plus ReportWeather to store weather forecast information when the user submits the ACT or POEM results.

Fig. 2. Conceptual representation of the proposed system. Table 2. Core functions of the weather-based healthcare mobile application. Core functions

Description

Sign in and register

Provide a function to allow the user to sign in and register with email address and password

Location retrieval

Detect the user’s current location and retrieve the weather forecast information for that location

Weather conditions

Provide weather forecast information including temperature, wind speed, pressure, humidity and rain

Main forecast activity

Provide weather forecast information based on the user location

Daily forecast

Provide daily updated weather forecast information for the current day, the next day, and the next 5 days

Hourly forecast

Provide hourly weather forecast information for the next 3 h

Report asthma

Provide the Asthma Control Test (ACT) with questions for the user to answer and submit the results (continued)

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare

35

Table 2. (continued) Core functions

Description

Report eczema

Provide the Patient-Oriented Eczema Measure (POEM) test with questions for the user to answer and submit the results

Report weather

Provide a function to collect the weather information as a timestamp when the user submits the ACT or POEM answers

Data storage

Store user answers with the timestamp in a real-time system database

Personalised settings

Provide several application settings options to allow the user to personalise preferences, such as weather forecast unit, display format and application theme

Graph activity

Provide a function to display the weather forecast information in a graphical form

Air quality index (AQI)

Provide a function to display AQI reading from the user location

Weather map activity

Provide a function to allow the user to interact and view the weather forecast information in a more creative way

System requirements

Provide a system that can work on mobile devices with Android version 6.0.1 and above. The application functional module must meet the system functional and non-functional requirements

Fig. 3. Context diagram of the system.

4 Results and Discussion The proposed weather-based healthcare model is based on the Uses and Gratification Theory (UGT) that incorporates weather attributes, demographic characters and personalisation effectiveness. UGT helps to identify the media and the elements that benefit users’ social need [22], thus ensuring sustained user engagement and wide adoption. This model suggests developing an mHealth application (media) using machine learning techniques (elements) for self-management (social need). Smartphones and tablets

36

R. Haque et al.

Fig. 4. Entity Relationship Diagram (ERD).

introduce lighter-weight operating systems and user interfaces with gesture-based interactions which helps building interactive mHealth application for weather-based healthcare. Figure 5 shows the capability of disease severity prediction based on weather conditions. The application provides a reliable and easy-to-use interface for asthma and eczema patients to stay updated with the weather forecast in their location.

Fig. 5. The forecast interface with asthma and eczema precautions.

Figure 6 provides examples of the current hourly and daily weather forecast extracted from the two weather resources, namely Wunderground [18] and DarkSky [19]. This is

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare

37

to provide a weather forecast report offline viewing feature, together with asthma and eczema precautions. Users are able to report and record the severity of diseases in the application using ACT for asthma and POEM for eczema. ACT is mainly used because it provides a numerical score to determine the severity level of asthma for individual patients. There are two tests available for ACT. One for those 12 years or older, and another, the Childhood ACT, for those under 12 years old. Meanwhile, POEM is used for monitoring atopic eczema severity. This weather-based healthcare mobile application is on the Android platform, and developed with Android Studio and the Firebase database. Data has been collected from users with asthma or eczema who reported their conditions through ACT or POEM reporting interfaces, by answering the Multiple-Choice Questions (MCQ) (see Fig. 7). Once their answers are submitted, a timestamp is created with the weather forecast information of that day and time. This timestamp, along with the pre-assigned number to each MCQ answer, is stored in the database. From the weather-based healthcare application, it was identified that both asthma and eczema cause a variety of symptoms that can worsen based on different weather conditions. The ACT and POEM results indicated that on some days, a user may not have symptoms, but on other days the user shows strong symptoms, depending on the weather conditions of those days. By analysing the results, it has been identified that cold temperature and thunderstorms were among the common causes of triggering asthma as well as worsening eczema for the majority of the users. This result agrees with the findings in the literature which show that weather has an apparent effect on asthma and eczema. This can be tracked for individual patients, who can then take necessary precautions based on the predicted weather conditions for self-management and early prevention, which can lead to personalised healthcare.

Fig. 6. Hourly and daily weather forecasts from two weather resources.

38

R. Haque et al.

Fig. 7. ACT and POEM reporting interfaces with MCQs.

This weather-based healthcare mobile application can be useful from the perspectives of both the user and machine learning developer. While the user obtains information and submits data through the application, the developer performs data cleansing to apply machine learning algorithms that can provide predictions to individual users of their asthma or eczema condition based on daily weather forecast in each user’s location. This is important because the severity level of these diseases differ among users and the extent of weather impact on each user’s condition also varies based on their location. In light of this, the on-going investigation is focused on how to benefit from machine learning use in the context of personalisation in weather-based healthcare. The machine learning techniques used for the personalised weather-based healthcare model include a regression technique to predict asthma attacks and eczema severity based on daily weather forecast. The model used in machine learning includes a recommendation technique to associate users’ activities or preferences with their situation and providing predictive feedback to them. This model recommends precautions to individual patients in certain weather conditions. A Recurrent Neural Network (RNN) [23] combines all the machine learning techniques mentioned above, so is considered for integration into the mHealth application. Specifically, RNN is suitable for modelling the personalised weather-based healthcare system because it can cluster a dataset with many variables in functional groups for individual patients. To accomplish personalisation in the proposed weather-based healthcare system, it is important to identify similar patterns and regular trends in the dataset for individual patients over a period of time. Consequently, a ‘many-to-one’ RNN is used with multiple input neurons at the input layer, including weather and demography input, and one output neuron, which is the chance rate of triggering an asthma attack/worsening eczema symptom. Once this rate is identified through an RNN as output for each patient, prediction results using the machine learning recommendation technique will be given to individual users on the mHealth application’s forecast interface. Figure 5 illustrates an example of this output,

Intelligent Health Informatics with Personalisation in Weather-Based Healthcare

39

where a list of precautions is provided for self-management and early prevention of asthma attacks or worsening eczema symptom based on weather triggers.

5 Conclusion Machine learning has become a universal and promising technology that can be researched and improved continuously and has the potential to contribute in many significant studies such as personalised weather-based healthcare for self-management and early prevention of worsening chronic disease symptoms. Personalisation is important in weather-based healthcare as it can offer prognosis information of health condition based on the weather to facilitate self-monitoring. Throughout this paper, it was identified that human health can be profoundly affected by the weather, which can trigger chronic diseases such as asthma and eczema. The effect of weather conditions on these diseases underlined the importance of personalised weather-based healthcare. It necessitated further research into how machine learning techniques can assist raising self-management and early prevention awareness in healthcare and provide predictions on how certain weather temperatures, exposure to allergens, barometric pressure changes, humidity and wind may trigger attacks in asthma, or worsen the symptoms of eczema. This paper proposes a meaningful way of collecting weather-based healthcare data through developing a mobile health (mHealth) application. The application provides daily weather forecast information retrieved from the user’s location. The main goal of the mHealth application is to allow patients to self-manage their condition and prevent from getting worse based on the provided weather forecast in their location. To do this, users can go through the Asthma Control Test (ACT) for asthma symptoms tracking or the Patient Oriented Eczema Measure (POEM) for monitoring atopic severity in skin allergy. The main limitation of the current version of the weather-based mHealth application is that it does not provide daily predictions to individual users of their asthma or eczema severity based on the weather forecast in their location, which can facilitate personalisation. Having said that, the development of this application as a data collection mechanism for machine learning process is revolutionary, due to its ability to connect and link different datasets into one self-adjusted training and testing dataset for Recurrent Neural Network (RNN) modelling. For further development, the absence of classification for prognosis through machine learning needs to be addressed. With the annotated dataset, the intended result of the proposed system using RNN can be properly measured and documented to assist machine learning researchers to develop algorithms through proper machine learning frameworks and design effective personalised weather-based healthcare systems for self-management and early prevention. Acknowledgment. The authors appreciate the financial support given by the Fundamental Research Grant Scheme, FRGS/1/2019/SS06/MMU/02/4 and Multimedia University, Cyberjaya, Malaysia (Project ID: MMUE/190031).

References 1. Akhil, J., Samreen, S., Aluvalu, R.: The future of healthcare: machine learning. Int. J. Eng. Technol. (UAE) 7, 23–25 (2018)

40

R. Haque et al.

2. Panch, T., Szolovits, P., Atun, R.: Artificial intelligence, machine learning and health systems. J. Glob. Health 8(2), 020303 (2018). https://doi.org/10.7189/jogh.08.020303 3. Lepeule, J., Litonjua, A.A., Gasparrini, A., Koutrakis, P., Sparrow, D., Vokonas, P.S., Schwartz, J.: Lung function association with outdoor temperature and relative humidity and its interaction with air pollution in the elderly. Environ. Res. 165, 110–117 (2018) 4. Alharbi, E., Abdullah, M.: Asthma attack prediction based on weather factors. Periodicals Eng. Nat. Sci. 7, 408–419 (2019). https://doi.org/10.21533/pen.v7i1.422 5. AAFA.: Weather can trigger asthma. Asthma and Allergy Foundation of America (2017). https://www.aafa.org/weather-triggers-asthma. Accessed 6 Aug 2020 6. D’Amato, G., Pawankar, R., Vitale, C., Lanza, M., Molino, A., Stanziola, A., Sanduzzi, A., Vatrella, A., D’Amato, M.: Climate change and air pollution: Effects on Respiratory Allergy. Allergy, Asthma Immunol. Res. 8(5), 391–395 (2016) https://doi.org/10.4168/aair.2016.8. 5.391 7. Zhang, Y., Peng, L., Kan, H., Xu, J., Chen, R., Liu, Y., Wang, W.: Effects of meteorological factors on daily hospital admissions for asthma in adults: A time-series analysis. PLoS One 9(7), e102475 (2014). https://doi.org/10.1371/journal.pone.0102475 8. Balato, N., Megna, M., Ayala, F., Balato, A., Napolitano, M., Patruno, C.: Effects of climate changes on skin diseases. Expert Rev. Anti. Infect. Ther. 12, 171–181 (2014) 9. Vocks, E., Busch, R., Frölich, C., Borelli, S., Mayer, H., Ring, J.: Influence of weather and climate on subjective symptom intensity in atopic dermatitis. Int. J. Biometeorol. 45, 27–33 (2001) 10. Charman, C., Venn, A., Ravenscroft, J., Williams, H.: Translating patient-oriented eczema measure (POEM) scores into clinical practice by suggesting severity strata derived using anchor-based methods. Br. J. Dermatol. 169(6), 1326–1332 (2013) 11. POEM. Patient Oriented Eczema Measure (2020). https://www.nottingham.ac.uk/research/ groups/cebd/resources/poem.aspx. Accessed 6 Aug 2020 12. Scarpino, M.: Tensor Flow for Dummies. Wiley, Hoboken, pp. 8–43, pp. 201–224, USA (2018) 13. Kurata, J.: Understanding machine learning with python (2016). Pluralsight. https://app.plu ralsight.com. Accessed 30 Apr 2020 14. Bassi, S.: Python for Bioinformatics. 2nd ed. CRC Press, Taylor & Francis Group, Boca Raton, pp. 30–37, pp. 158–208 (2018) 15. Tsang, K., Pinnock, H., Wilson, A., Shah, S.: Application of machine learning to support self-management of asthma with mHealth. In: 42nd Annual Int’l Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Montreal, QC, Canada (2020) 16. Gaynor, M., Schneider, D., Seltzer, M., Crannage, E., Barron, M.L., Waterman, J., Obersle, A.: A user-centered learning asthma smartphone application for patients and providers. Learn. Health Syst. 4(3), e10217 (2020) 17. Hurault, G., Domínguez-Hüttinger, E., Langan, S.M., Williams, H.C., Tanaka, R.J.: Personalized prediction of daily eczema severity scores using a mechanistic machine learning model. Clin. Exp. Allergy (2020). https://doi.org/10.1111/cea.13717 18. Wunderground. Weather Underground Application Programming Interface (API) (2020). https://www.wunderground.com/weather/api/d/pricing.html. Accessed 6 Aug 2020 19. DarkSky. DarkSky API (2020). https://darksky.net/dev/docs. Accessed 6 Aug 2020 20. VanderPlas, J.: Python data science handbook. O’Reilly Media Inc, Sebastopol (2017) 21. Geron, A.: Hands-on machine learning with scikit-learn and tensorflow. O’Reilly Media Inc, Sebastopol (2017) 22. Hossain, M.: Effects of uses and gratifications on social media use. PSU Res. Rev. 3(1), 16–28 (2019). https://doi.org/10.1108/prr-07-2018-0023 23. Phan, D., Yang, N., Kuo, C., Chan, C.: Deep learning approaches for sleep disorder prediction in an asthma cohort. J. Asthma (2020) https://doi.org/10.1080/02770903.2020.1742352

A CNN-Based Model for Early Melanoma Detection Amer Sallam1(B) , Abdulfattah E. Ba Alawi2 , and Ahmed Y. A. Saeed2 1 Computer Network and Distributed Systems Department, Taiz University, Taiz, Yemen

[email protected] 2 Software Engineering Department, Taiz University, Taiz, Yemen

Abstract. Melanoma is a serious form of skin cancer that develops from pigmentproducing cells known as melanocytes, which in turn produce melanin that gives your skin its color. Early detection of these symptoms will certainly help affected people to overcome their suffering and find appropriate solutions for their treatment methods. That is why researchers have tried in many studies to provide technical solutions to help early detection of skin cancer. In this paper, a smart pre-trained model based on deep learning techniques for the early detection of Melanoma and Nevus has been proposed. It is designed to track and divide the dynamic features of the dermoscopic ISIC dataset into two distinguished classes Melanoma and Nevus of epidermal pathologies. AlexNet and GoogLeNet are used to classify each cancer type according to their profile features. It was found that the average classification accuracy for the above-mentioned algorithms is 90.2% and 89% respectively, providing plausible results when comparing to other existing models. Keywords: Skin diseases · Dermoscopic · Dermatologist · CAD · Melanoma · GoogLeNet · AlexNet

1 Introduction Skin is the first defense line in the human body. It plays a critical and vital role in protecting the body from infections, injuries, UV radiation, harmful radiation, and temperature control. Such importance has been attracting many researches in the field of computer science (i.e. data mining, computer vision, and pattern recognition). So many studies are constantly investigating and developing for such diseases. Malign Malignancy [1] is one of the leading causes of skin cancer that risks people’s lives severely. It can scatter over time rapidly. The rapid growth of the melanoma cases makes it a very vigilant form of cancer that has been receiving wide attention. Increasing survival rates for patients are critical to the early diagnosis of melanoma [2]. In this study, a CNN-based model is proposed to recognize melanoma disease on dermoscopic images. GoogLeNet and AlexNet have been employed to make a decision on the recognition process. The remaining parts of this paper are arranged in the following structure: Sect. 2 presents a background of the problem domain and introduces a brief description of the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 41–51, 2021. https://doi.org/10.1007/978-3-030-70713-2_5

42

A. Sallam et al.

related works. The proposed methods are elaborated in Sect. 3. Then, the obtained results and observations are discussed and analyzed in Sect. 4. The last section (Sect. 5) is the conclusion and the directions of the future work are briefly outlined.

2 Background The nightmare of skin diseases in general and the skin cancer in particular, still causing a suffering and death of millions compelled humankind to shudder. For such reasons, a several attempts try to diminish such suffering. Dermoscopy is a trustworthy screening tool for malignant melanoma. Skin experts or dermatologists can greatly minimize the risk of malignant melanoma wrong diagnosis by integrating skin pathology with macroscopic and microscopic clinical dermatology. Expert dermatologists, however, demonstrate variance among observers and can lead, in presenting the same lesion [3], to different outcomes. 2.1 Related Works Pattern recognition is a hot topic for researchers that has inspired many scholars to find solutions for real-world problems including skin diseases. The manual inspection method of a skin lesion, which cannot be possible over time, is generally dependent on a qualified specialist; therefore, machine learning-based methods are proposed by a group of researchers working in this domain. Codella et al. [4] implemented an ensemble of deep learning CNN-based models to identify lesions through three separate architectures, which are integrated to form a standard architecture of pre-trained models. With the aid of 1279 images of the International Symposium on Biomedical Imaging (ISBI) 2016, the model is tested with a fully CNN model, achieving 76% accuracy. Li and Shen [5] proposed a reliable DL approach for the identification of lesion problems by segmentation methods, function estimation, and eventually recognition. The approach is intended to address lesion extraction issues. For lesion segmentation and classification, two fully residual network layers are used and enhanced through the measurement unit of lesion index. On the (ISIC 2017) dataset, the proposed approach achieved 91.2%. Adjed et al. [6] implemented a fusion of wavelet, curvelet, and two local binary pattern descriptors with structural and textural characteristics. The suggested treatment is carried out by using 200 dermoscopic photographs from the PH2 collection, including 160 non-melanoma and 40 melanoma photographs. The validated results were very promising a random cross-validation approach of Support Vector Machine (SVM) success at 78.93%, 93.25%, and 86.07% for the sensitivity, specificity, and accuracy respectively. Mukherjee et al. [7] designed a Deep Convolution Neural Networks DCNN method based on MEDNOE and Dermofit datasets in two phases. Another different performance is calculated at the early stage and later combined both datasets and achieves 83.07% in terms of accuracy. Mahbod et al. [8] proposed an automated model for recognizing skin lesions with deep features optimization. Three CNN models including ResNet18,

A CNN-Based Model for Early Melanoma Detection

43

AlexNet, and VGG are utilized. Then, the SVM classifier used and achieved 83.33% melanoma recognition accuracy. Abbas and Celebi [9] introduced a model that can recognize pigmented skin lesions using a new method named DermoDeep. The DermoDeep technique consists of five architectural layers and involves the fusion of Visual and DNN elements. This model trained using 2800 images. The efficiency of the model is verified and obtained 93% and 95% sensitivity and specificity respectively. In the face of the eclecticism on the basis of what has been obtained by the aforementioned methods, but they still exist some prominence and imperfections in terms of generalization as a result of a difference in dermoscopic scans and poor resolution of datasets. In addition, the collection of most discriminatory characteristics is not sufficient. Many Computer-Aided Diagnosis Systems (CADs) are also used for recognizing the skin lesions, and to effectively assist the dermatologists’ clinical diagnosis [10]. Therefore, it is very significant to develop an efficient computer-aided Diagnosis system for melanoma classification. At present, one of the most commonly used diagnostic features of melanomas in the CAD systems is the ABCD rule [11], Menzies method [12], the seven-point checklist [13], and the CASH method [14]. These approaches focus primarily on global characteristics such as color, texture, and form that distinguish melanomas, which are difficult to react and represent the same identical appearances of distinct lesions (e.g. melanoma and nevus). This study suggests a method for diagnosing melanoma and resolving the issue of melanoma skin disease. This approach can be described in four important ways. First-ly, this method proposes an effective algorithm for extracting features from dermoscopic images. Secondly, the end-to-end classification method can be applied without a need for any feature selection experience. Thirdly, this method treats all local regional features of skin image equally; it does not consider any specific portion of the image. It does not also require any complex hardware, which is undoubtedly cost-effective by proposing a robust and effective tool that may help in diagnosing skin lesions. Two types of lesions are considered in this study. They are Melanoma and Nevus. 2.2 Melanoma and Nevus Lesions Melanoma [15] is a type of cancer that starts from unchecked cells in nearly any region of the body; cells can become cancer and infects others. Thus, cancer begins and grows. Several other kinds of skin cancers are more rarely reported than melanoma. However, melanoma is serious since it is much more likely to spread to other areas of the body unless early detection takes place. Melanoma [16] is a developing cancer of the melanocytes. Many of the melanoma cells create a dark or brown tumor. Any melanomas, however, do not contain melanin and can look black, brown, or green. Melanomas are most likely to arise with the trunk (thorn as well as back) or anywhere on the skin, in both men and women. Melanoma lesion mostly appears in the face and the neck. Nevus (plural: nevi) is the scientific name for the mole of skin [17, 18]. It is very common. Simple nevi collections of colored cells are harmless. It usually tends to be tiny gray, tan, or purple. It may be born before or without moles. The lumps you are born with are referred to as congenital lumps. Throughout puberty and adolescence, however,

44

A. Sallam et al.

most moles develop; and they are considered a nevus gained. Because of sun exposure, moles will also grow later in life. There are many other nevi types. Some are innocuous and the others are more serious. As shown in Fig. 1, visual appearance between different skin lesions, especially melanoma and benign lesions can be indistinguishable. There are several visual methods used by dermatologists to diagnose melanoma without biopsy, but the accuracy of these methods is poor; it is around 60% [19]. Due to the high similarity between nevus and melanoma, the classification of these diseases is difficult to be done using visual differences. The following figure shows samples for melanoma and nevus.

Fig. 1. The pictures of A, B, C, and D are melanoma samples whereas E, F, G, and H are Nevi cases.

3 Methodology The proposed model of recognizing melanoma disease in early-stage can be depicted as shown in Fig. 2 below.

Fig. 2. The proposed model for recognizing melanoma.

A CNN-Based Model for Early Melanoma Detection

45

Figure 2 clearly depicts the proposed melanoma recognition model. The acquired images are forwarded to the deep neural network after they have been pre-processed. Then in the deep convolution layers, the features of data are extracted and obtained. Finally, the classification results of the input image are shown. 3.1 Experimental Dataset The dataset that is used is ISIC dataset is downloaded from ISIC 2019 Challenge [20– 22]. In total, it includes about 10275 images for both melanoma and nevus classes. About 4275 images represent melanoma class and 6000 images for nevus class. The following figure (Fig. 3) shows the main steps that have been followed to build the proposed model.

Fig. 3. The steps that are done for building the proposed model.

Figure 3 shows the steps that can be followed to train and evaluate the proposed melanoma recognition model. During the training phase, the two classifiers that have been involved are AlexNet and GoogleNet. In the pre-processing phase, it is essential to make the images fit the first layer of the pre-trained models used. The images are resized to 224 × 224 in the case of GoogleNet and 227 × 227 for AlexNet, and the features of each class have been extracted successfully. Later, trained classifiers are employed in the testing phase. 3.2 AlexNet Pre-trained Model AlexNet [23] is a well-known pre-trained model in the ImageNet Larger Visual Recognition Challenge (ILSVRC) in autumn September 2012. AlexNet demonstrated a superior deep learning ability of GPUs. AlexNet is built with 25 layers and it has been commonly used for image classification tasks.

46

A. Sallam et al.

3.3 GoogleNet Pre-trained Model GoogleNet [24] is also a competed and superior pre-trained model since it is implemented as a state of the art responsible for image detection and classification task (ILSVRC 2014). The key mark of this designed model is the increased use of computational power within the network. Due to its engineered architecture that enables the depth and breadth of the network to be expanded while retaining a constant computational budget [25]. The structure decisions were based on the Hebbian theory and the multiple-scale processing insight to maximize efficiency [24]. GoogleNet is constructed of 22 deep-layer networks, whose attributes are evaluated in the sense of classification.

4 Results and Discussion By using deep learning techniques, the outcome obtained from this model in terms of training accuracy is illustrated in Fig. 4.

Fig. 4. The obtained training accuracy.

In Fig. 4, the obtained training accuracy using AlexNet and GoogLeNet pre-trained models is 89.12% and 90% respectively, though many images have been used as an experimental dataset. The efficiency and the robustness of both models can be observed from the outcome of the training loss as reflected in Fig. 5. GoogLeNet pre-trained model showed very promising results; it achieved less than 0.25 loss while AlexNet achieved 0.3. The performance of pre-trained models in terms of accuracy during the validation phase is shown in the following figure (Fig. 6).

A CNN-Based Model for Early Melanoma Detection

47

Fig. 5. The obtained loss during the training phase.

Fig. 6. The obtained accuracy during the validation phase.

In the context, again, the GoogLeNet performs well and competes with the validation accuracy of AlexNet pre-trained model. The validation losses of the pre-trained models are illustrated below in (Fig. 7).

48

A. Sallam et al.

Fig. 7. The performance of the used pre-trained models as regards validation loss.

GoogLeNet achieved better validation loss which goes to around 0.24 than AlexNet which achieved 0.3 at epoch 10. However, in all conducted experiments, it has been noticed that GoogLeNet is very steady and can be used effectively to tackle diagnosis issues of skin diseases. To investigate the allegation, this model has been compared with other current models and the outcome is illustrated in the following table. Table 1. Comparison between this work and the works done in this field of study The author

Dataset used

Classifiers

Results

Lobez et al. [26]

ISBI 2016 Challenge [27]

Modified VGG16

Acc = 81.33%

Abbas et al. [9]

2800 images from: Skin-EDRA ISIC DermNet Ph2-dataset

SVM

AUC = 0.880 Sensitivity = 88.2% Specificity = 91.3%

Mukherjee et al. [7]

Dermofit/MEDNODE

CNN malignant lesion detection (CMLD)

Acc = 90.58 and 90.14%

Prathiba et al. [28]

Harvard Dataset

CNN

NA

Matsunaga et al. [29]

ISIC 2016

CNN

Acc = 83.09%

Adjad et al. [6]

ISIC

SVM

Acc = 86.07%

Yu et al. [30]

ISIC 2016

CNN

Acc = 85%

Our proposed model

10275 images from ISIC 2019

AlexNet, GoogLeNet

Acc = 90.2%

A CNN-Based Model for Early Melanoma Detection

49

As shown in Table 1, the obtained results of this model are promising while comparing to the results of other models. The proposed model has been tested with a large dataset to evaluate its performance. Figure 8 shows the results as testing melanoma samples take place.

Fig. 8. Test samples have been selected randomly.

As shown in Fig. 8, the system is able to recognize melanoma lesion efficiently. Also, it is shown that the number located above each image and next to the recognized class name represents the confidence percentage of the obtained results.

5 Conclusion The designed model has been proposed to classify two types of skin diseases using deep learning pre-trained models. The proposed method can be applied in the field of health informatics to facilitate the diagnosis of melanoma process. In addition to this, it can provide dermatologists with a clear picture of skin diseases in order to help them to decide the treatment. The system can also provide an effective determination and robust solution about early melanoma skin cancer prediction. Expanding the dataset with more classes of diseases and the challenging task of bringing to more robust learning of the network parameters has been left for future works. Besides, the possibility to use the advantage of some pre-processing steps for input images (e.g. color constancy) could be accounted. Finally, the use of a segmentation phase could be considered to obtain registered images for a common reference.

References 1. Barata, C., Ruela, M., Francisco, M., Mendonça, T., Marques, J.S.: Two systems for the detection of melanomas in dermoscopy images using texture and color features. IEEE Syst. J. 8(3), 965–979 (2013) 2. Silveira, M., Nascimento, J.C., Marques, J.S., Marcal, A.R.S., Mendonca, T., Yamauchi, S., Maeda, J., Rozeira, J.: Comparison of segmentation methods for melanoma diagnosis in dermoscopy images. IEEE J. Sel. Top. Sign. Process. 3(1), 35–45 (2009) 3. Ahn, E., Kim, J., Bi, L., Kumar, A., Li, C., Fulham, M., Feng, D.D.: Saliency-based lesion segmentation via background detection in dermoscopic images. IEEE J. Biomed. Health Inf. 21(6), 1685–1693 (2017)

50

A. Sallam et al.

4. Codella, N.C.F., Nguyen, Q.-B., Pankanti, S., Gutman, D.A., Helba, B., Halpern, A.C., Smith, J.R.: Deep learning ensembles for melanoma recognition in dermoscopy images. IBM J. Res. Dev. 61(4/5), 5:1−5:15 (2017) 5. Li, Y., Shen, L.J.S.: Skin lesion analysis towards melanoma detection using deep learning network. Sensors 18(2), 556 (2018) 6. Adjed, F., Gardezi, S.J.S., Ababsa, F., Faye, I., Dass, S.C.: Fusion of structural and textural features for melanoma recognition. IET Comput. Vis. 12(2), 185–195 (2017) 7. Mukherjee, S., Adhikari, A., Roy, M.: Malignant melanoma classification using crossplatform dataset with deep learning CNN architecture. In: Bhattacharyya, S., Pal, S.K., Pan, I., Das, A. (eds.) Recent Trends in Signal and Image Processing: Proceedings of ISSIP 2018, pp. 31–41. Springer, Singapore (2019) 8. Mahbod, A., Schaefer, G., Wang, C., Ecker, R., Ellinge, I.: Skin lesion classification using hybrid deep neural networks. In: ICASSP 2019–2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 1229–1233. IEEE (2019) 9. Qaisar Abbas, M., Celebi, E.: DermoDeep-A classification of melanoma-nevus skin lesions using multi-feature fusion of visual features and deep neural network. Multimedia Tools Appl. 78(16), 23559–23580 (2019) 10. Pathan, S., Prabhu, K.G., Siddalingaswamy, P.C.: Control: techniques and algorithms for computer aided diagnosis of pigmented skin lesions—a review. Biomed. Sig. Process. Control 39, 237–262 (2018) 11. Stolz, W.: ABCD rule of dermatoscopy: a new practical method for early recognition of malignant melanoma. Eur. J. Dermatol 4, 521–527 (1994) 12. Menzies, S.W., Ingvar, C., Crotty, K.A., McCarthy, W.H.: Frequency and morphologic characteristics of invasive melanomas lacking specific surface microscopic features. Arch. Dermatol. 132(10), 1178–1182 (1996) 13. Argenziano, G., Fabbrocini, G., Carli, P., De Giorgi, V., Sammarco, E., Delfino, M.: Epiluminescence microscopy for the diagnosis of doubtful melanocytic skin lesions: comparison of the ABCD rule of dermatoscopy and a new 7-point checklist based on pattern analysis. Arch. Dermatol. 134(12), 1563–1570 (1998) 14. Henning, J.S., Dusza, S.W., Wang, S.Q., Marghoob, A.A., Rabinovitz, H.S., Polsky, D., Kopf, A.W.: The CASH (color, architecture, symmetry, and homogeneity) algorithm for dermoscopy. J. Am. Acad. Dermatol. 56(1), 45–52 (2007) 15. Mitchell, T.C., Karakousis, G., Schuchter, L.: Melanoma. In: Abeloff’s Clinical Oncology. pp. 1034–1051. e1032. Elsevier (2020) 16. What is Melanoma Skin Cancer ? https://www.cancer.org/cancer/melanoma-skin-cancer/ about/what-is-melanoma.html (2019). Accessed 16 May 2020 17. Massi, G., LeBoit, P.E.: Common nevus. In: Massi, G., LeBoit, P.E. (eds.) Histological Diagnosis of Nevi and Melanoma, pp. 29–46. Springer, Berlin (2014) 18. Massi, G., LeBoit, P.E.: Histological Diagnosis of Nevi and Melanoma. Springer, Berlin (2013) 19. Kittler, H., Pehamberger, H., Wolff, K., Binder, M.J.T.l.O.: Diagnostic accuracy of dermoscopy. Lancet Oncol. 3(3), 159–165 (2002) 20. ISIC Dataset. https://challenge2019.isic-archive.com/ (2019). Accessed 1 May 2020 21. Society, A.C.: Cancer Facts & Figures 2019. https://www.cancer.org/content/dam/cancerorg/research/cancer-facts-and-statistics/annual-cancer-facts-and-figures/2019/cancer-factsand-figures-2019.pdf (2019). Accessed 30 May 2019 22. P. Tschandl, C.R., Kittler, H.: The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions. arXiv:1710.05006. 23. Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems, pp. 1097–1105 (2012)

A CNN-Based Model for Early Melanoma Detection

51

24. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.: Going deeper with convolutions. arXiv 2014. 1409 (2014) 25. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., Rabinovich, A.J.: Going deeper with convolutions. CoRR. (2014) 26. Lopez, A.R., Giro-i-Nieto, X., Burdick, J., Marques, O.: Skin lesion classification from dermoscopic images using deep learning techniques. In: 2017 13th IASTED international conference on biomedical engineering (BioMed), pp. 49–54. IEEE (2017) 27. Gutman, D., Codella, N.C., Celebi, E., Helba, B., Marchetti, M., Mishra, N., Halpern, A.J.: Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (ISBI) 2016, hosted by the international skin imaging collaboration (ISIC) (2016) 28. Prathiba, M., Jose, D., Saranya, R.: Automated Melanoma Recognition in Dermoscopy Images via Very Deep Residual Networks. In: IOP Conference Series: Materials Science and Engineering 2019, vol. 1, p. 012107. IOP Publishing 29. Matsunaga, K., Hamada, A., Minagawa, A., Koga, H.: Image classification of melanoma, nevus and seborrheic keratosis by deep neural network ensemble (2017) 30. Yu, L., Chen, H., Dou, Q., Qin, J., Heng, P.-A.: Automated melanoma recognition in dermoscopy images via very deep residual networks. IEEE Trans. Med. Imaging 36(4), 994–1004 (2016)

SMARTS D4D Application Module for Dietary Adherence Self-monitoring Among Hemodialysis Patients Hafzan Yusoff1(B) , Nur Intan Raihana Ruhaiyem2 , and Mohd Hakim Zakaria1 1 School of Health Sciences, Universiti Sains Malaysia, 16150 Kota Bharu, Kelantan, Malaysia

[email protected] 2 School of Computer Sciences, Universiti Sains Malaysia, USM, 11800

Gelugor, Penang, Malaysia

Abstract. The mortality rate in hemodialysis patients is 6.3–8.2 times higher than the general population. Failure to adhere to dietary intake recommendation, was one of the most significant factors affecting patient survival. Technology-mediated approach such as web and mobile application could be the most desirable approach nowadays. This paper presents the SMARTS dual application modules development by using ADDIE model, beginning with the analysis of needs, followed by content and face validation in the design phase, and finally the development of application prototype. The application system was designed to enable seamless access, interaction, and monitoring between all the involved users; patient, caretaker, and Healthcare Provider (HCP). Twenty-five respondents involved in the need assessment and also face and validity testing, Most of them are dietitian from government hospital (n = 16, 64%), university medical centers (n = 6, 24%) and private hospital (n = 2, 8%), with ample experience managing hemodialysis patients. Majority of them rated the content (84%), and purpose of the app as a new nutrition education tool (84%) as the most appealing properties of the app, followed by the visual appealing (68%), and variety of topics offered (40%). Some improvisation was suggested on the comprehension and quality of the text, inclusion of nutrient tracker, presentation of education messages in video format, and adding more visuals rather than textual information to enhance understanding. The SMARTS D4D module was well-accepted and supportive of respondents’ needs. Appropriate modifications have been done based on the valuable respondents’ feedbacks. Keywords: Hemodialysis · Dietary plan · Kidney failure · Application module

1 Introduction 1.1 Study Background End-stage renal disease (ESRD), defined as chronic renal disease (CKD) stage five, is a permanent loss of renal function identified as a glomerular filtration rate of less than 15 ml/min requiring hemodialysis treatment. In Malaysia, the number of ESRD patients © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 52–60, 2021. https://doi.org/10.1007/978-3-030-70713-2_6

SMARTS D4D Application Module for Dietary Adherence Self-monitoring

53

has been on an upward track for the past 20 years [1] fueled by aging populations and a wide range of chronic non-communicable diseases, especially diabetes mellitus and hypertension. Survival wise, a previous study suggested that 9–13% of patients on hemodialysis die within one year [2]. Among the most important factors influencing patient survival was failure to conform to treatment regimens [3], including strict schedule of dialysis treatment, medicine, fluid, and dietary intake prescriptions. The standard dietary prescription for ESRD patients undergoing hemodialysis is 750 to 1L of fluid, not more than 2 g of sodium, 2 g of potassium and 1 g of phosphorus. As regards to protein and energy, an intake of 1.2 g of protein per kg body weight per day and 35 kcal energy per kg body weight per day were recommended [4]. Complex food and fluid prescriptions are, however, difficult for patients to adopt, without assistance and close monitoring by the HCP such as dietitian. In Malaysia, it is a common practice that dietitians facilitate dietary and fluid selfmanagement of patients through face-to - face consultation based on referral system, and the dietary record of the patients was done traditionally without proper monitoring system in hand. This might lead to low dietary adherence among the patients [5]. It is quite challenging too for the dietitian to make information available at a time convenient for patients; provide information tailored to individual needs, cultures, and food preferences; and complement patient decision making with useful feedback [6]. Thus, this application system was developed to facilitate self-monitoring among patients and assist healthcare providers especially dietitian to meet these challenges. 1.2 SMARTS D4D Module The idea to develop technology-mediated SMARTS D4D dietary module was originated from dietitian perspective based on years of experience in hemodialysis patients management. The acronym SMARTS D4D represents the specific dietary strategies (see Fig. 1). A dual-application system was then developed based on the module. The app was compatible for both iOs and Android platforms. It also enables for future updates by taking into account the possibility of new methods, treatment regime or parameters to be introduced in future.

Fig. 1. SMARTS D4D dietary strategy

54

H. Yusoff et al.

2 Methodology 2.1 SMARTS D4D Module Development The SMARTS D4d module was designed to empower the hemodialysis patients in controlling their nutrients intake, assist them with routine meal preparation, thus eliminates the unnecessary complication that might occur due to non-adherence. Besides, it also becomes an enabler and support system for the healthcare providers such as dietitian to give precise recommendations and provide personalized diet plan for patients based on anthropometric, physiological, clinical and nutritional evaluations. This module was designed by following the ADDIE instructional design model [7]. The acronym ADDIE represents five phases: Analysis, Design, Development, Implementation, and finally Evaluation (see Fig. 2). This model provides a systematic approach for designing and developing an effective application. We began by critically analyze the existing dietary apps, followed by designing the goals and objectives of our modules, and finally we proceed to the development phase by generating the food database (back-end) and interface design (front-end). Then, we conducted a pilot study to test the content and face validity of the modules. The evaluation phase is yet to be accomplished.

Fig. 2. ADDIE instructional design model

2.2 Analysis of Existing Application System Currently, there is no existing system that helps dialysis patients to keep track of their diet. Before the ideas of developing this system, Health Care Professional kept track manually of their patients’ diet by asking them every time during appointment. That previous method seems unreliable due to patients tend to not keep track of their daily meals. There are a few applications that kind of similar to proposed ideas, but they are not specifically developed for End Stage Renal Disease (ESRD) patient. Table 1 gives a summary of the existing applications targeting hemodialysis patients as potential user with feature comparisons.

SMARTS D4D Application Module for Dietary Adherence Self-monitoring

55

HealthifyMe Mobile App. This application enables the users to track their health, weight loss, and eat healthy food by using the guidance. It also tracks user’s calorie intake, based on the food content. The main purpose of this application is to recommend a custom diet plan for weight loss designed for men and women with specific health goals [8]. CKD Care Mobile App. This application allows medical professionals to estimate kidney function using the eGFR calculator and provides care guidelines. This application only estimates kidney function by calculating possibility of person having kidney problem or not [9].

Table 1. Comparison of existing system. Mobile application 1. Interaction between HCP and users 2. Nutrition and calorie calculator 3. Information about ESRD diet 4. HCP can monitor remotely the patient progress 5. Able to know the nutritional value in foods and drinks 6. Provide a Diet personalize diet plan for specific user

HealthifyMe App. CKD Care App. SMARTS D4D App. √ √

√ √

√

√

√

√

√

√

√

2.3 Modules of Dual-Application Systems The new system was a dual-systems consist of web-based and mobile application platforms as shown in Table 2. The system allows admin and HCP to login to the web application system, while allowing patients to login to mobile application. All patients’ data will be entered by HCP in web application. Patients are able to access the SMARTS dietary plan via mobile application, enables them to keep track of their daily nutrition intake by prompting the input of their daily food intakes in the food library. The HCP then be able to analyze the patient’s health condition based on the input and generate the report. The details on the module was illustrated in Table 3. 2.4 Content and Face Validation Testing A quantitative, online survey involving 25 dietitians was performed to assess their need for an app, their willingness to use an app during consultation with hemodialysis patient

56

H. Yusoff et al. Table 2. The two main platforms.

Mobile application

Web-based application

User: Patient User: Admin (Health-care provider, HCP) Act as information sharing platform Act as a platform to access the patient’s information, including personal data, dietary intakes, anthropometry, biochemical tests etc. Dietary plan Dietary tracker

Table 3. The modules Modules

Mobile and Web-based application

1. User (Patient) management

Allow admin to manage users account, view user information based on the user input Allow user to register and manage their own account or information

2. Food library management

Allow admin to create new food data through system Allow admin to update existing data through system Allow admin to view food data through the system Allow admin to delete food data through the system Allow users to access food list using application and get the food details

3. User report management

Allow admin to enter user assessment report data through application, when examining user/patient Allow admin to view lists of user reports through the system Allow admin and user to access the SMARTS dietary plan Allow admin and user to access food list to be used in the diet plan tracker

4. Diet plan management

Allow user to enter the intake of food based on the existed food library and send it to system Allow admin to monitor daily intake of user through system Allow admin to analyze the user intake based on SMARTS diet plan and monitor the user progress Allow user to view the daily intake of food and monitor the nutritional progress based on the food intake

and assess the content and face validity of the module. Each of them was equipped with a softcopy of the module and the proposed app visualization scheme sent via email. Only respondents who consented and fulfil the inclusion criteria were recruited into this study; a) Physicians, or dietitians, or dialysis nurse from health facilities in Malaysia, b) aged between 18 to 50 years old, and c) willing to complete the study through online survey. The respondents rated the drafted module and app visualization scheme (see

SMARTS D4D Application Module for Dietary Adherence Self-monitoring

57

Fig. 3) based on the appropriateness of the contents (selection of topics and variety), quality of graphics used, and the comprehension of the content (the scientific terms used, fonts and language). The survey assessment items were adapted from a previous study [10].

3 Module Design, Development and Findings 3.1 Technology Deployed and System Architecture The development of this complete system involves various technologies to make it compatible to many platforms. For hardware, processor 1.7 GHz and RAM sized 6 GB were used, developed using HTML, CSS, MySQL, Javascript, PHP and Angular 5 programming language. Other tools also used such as XAMPP server as the localhost server, VSCode as the code editor and Postman as API testing tool. The system architecture as follows (see Fig. 3).

Fig. 3. System architecture of SMARTS D4D application

For the system, it is accessed through the website and mobile application whereby internet connection is required to make use of the system. This system is mainly divided into two which are websites for admin, while the mobile application is for user/patient. This system stores all the data and information within MySQL database and can be accessed through phpMyAdmin. The connection between website, the apps and database were done through.php files and API created using PHP. 3.2 Dual-System Implementation The implementation strategy for developing this system took a bottom-up approach where the core systems were built first. Low level system, which are backend server, Application Programming Interface (API) and database were built first since they serve as the main components of the system. These low-level systems act as the main communication channel for the whole system.

58

H. Yusoff et al.

The backend mainly functions as a service provider that provides services in the form of formatted JSON data and handles all the back-end functionalities. Among those functionalities are communication with database, handling and routing requests from the client through HTTP protocols and protecting from unauthorized access. All requests will be handled by controller and middleware for their own actions such as authenticating users, data movement handling and data format handling to be sent out to the client. The whole backend system was built mostly using PHP and its framework. By using the bottom down approaches, we can build and test the group of subsystems that can be easily implemented with any top-level system (user interfaces). The top-level systems are built later after the bottom level systems already running in an acceptable manner.The top-level system (user interface) was also built using PHP and JavaScript. The top level and bottom level are separated but was communicated or connected by API. The usage of API as middleman can allows for more scalability and dynamic data processing. The API endpoint can be implemented by any other front-end systems such as a desktop application or a mobile application which serve data in an easy to manipulate form such as a JSON format. Using the bottom-up approaches will enable the system user interface to be independently designed and modified without having to make drastic changes to the main low-level system. This provides more efficiency in the development process by enabling multiple programmers to independently build the systems. 3.3 Needs, Content and Face Validity Testing Majority of the respondents are dietitians from government hospital (n = 16, 64%), university medical centers (n = 6, 24%) and private hospital (n = 2, 8%), with ample experience managing hemodialysis patients. Majority of them rated the appropriateness of the content (84%), and the purpose of the app as a new nutrition education tool (84%) as the most appealing properties of the app, followed by the graphic quality (68%), and variety of topics offered (40%). However, most of them recommend improvisation in terms of the comprehension and quality of the text (72%). Responding to the item “How would you judge the comprehension of the module?”, all respondents rated the module as good (52%) and very good (48%) respectively. The detailed item analysis on need, content, and face validity of the module was illustrated in the Table 4 and Table 5. Table 4. Need assessment Item

Yes (%) No (%)

1. Do you think this module is helpful for healthcare providers to monitor and facilitate dialysis patients?

100

0

2. Would you like to use this module as a nutrition education tool?

100

0

3. Would you like to use this module in a mobile application form?

100

0

SMARTS D4D Application Module for Dietary Adherence Self-monitoring

59

Table 5. Content and face validity assessment Assessment itema

Mean rating Standard deviation

1. Appropriateness of the content

4.28

0.66

2. Sufficiency of the content

4.16

0.67

3. Quality of text

4.08

0.56

4. Quality of graphics

3.80

0.94

4.16

0.88

5. Acceptability of: (a) Calorie management (b) Protein management

4.24

0.86

(c) Potassium management

4.24

0.81

(d) Phosphate management

4.20

0.94

(e) Sodium management

4.12

0.99

(f) Fluid management

4.24

0.99

4 Conclusion The SMARTS D4D application module was found to be greatly satisfactory and supportive of respondents’ needs. Appropriate modifications have been done to the module based on the valuable feedbacks given by the respondents. A dual-system comprising of a web service and a mobile application is currently under development. Once completed, this mobile application will enable the patient to keep track with their nutritional intake while HCP could monitor them remotely using one of the modules that features auto calculation and recording of the patient’s dietary intake via the web-based application.

References 1. Bujang, M.A., Adnan, T.H., Hashim, N.H., Mohan, K., Kim Liong, A., Ahmad, G., Haniff, J.: Forecasting the incidence and prevalence of patients with end-stage renal disease in Malaysia up to the year 2040. Int. J. Nephrol. 2(5), 24–34 (2017) 2. Chandrashekar, A., Ramakrishnan, S., Rangarajan, D.: Survival analysis of patients on maintenance hemodialysis. Indian J. Nephrol. 24(4), 206–213 (2014) 3. Collins, A.J., Foley, R.N., Herzog, C., Chavers, B.: US renal data system 2012 annual data report (cl-476). Am. J. Kidney Dis. 61, A7 (2013) 4. Fouque, D., Vennegoor, M., Ter Wee, P., Wanner, C., Basci, A., Canaud, B., VanHolder, R.: EBPG guideline on nutrition. Nephrol. Dial. Transplant. 22(Suppl 2), ii45–ii87 (2017). 5. Luis, D., Zlatkis, K., Comenge, B., García, Z., Navarro, J.F., Lorenzo, V., Carrero, J.J.: Dietary quality and adherence to dietary recommendations in patients undergoing hemodialysis. J. Ren. Nutr. 26(3), 190–195 (2016) 6. Welch, J.L., Astroth, K.S., Perkins, S.M., Johnson, C.S., Connelly, K., Siek, K.A., Scott, L.L.: Using a mobile application to self-monitor diet and fluid intake among adults receiving hemodialysis. Res. Nurs. Health 36(3), 284–298 (2013)

60

H. Yusoff et al.

7. Morrison, G.R.: Designing Effective Instruction, 6th edn. Wiley, UK (2010) 8. HealthifyMe [Mobile application]. https://www.healthifyme.com (2020) 9. CKD Care [Mobile application]. https://www.kidney.org/apps/professionals/ckd-care-intera ctive-guide-clinicians (2020) 10. Dali, W.P.E.W., Mohamed, H.J.J., Yusoff, H.: Development and evaluation of interactive multimedia-based nutrition education package IMNEP (researcher) to promote healthy diet for overweight and obese children. Health 8(1), 24–48 (2017)

Improved Multi-label Medical Text Classification Using Features Cooperation Rim Chaib1,2(B) , Nabiha Azizi1,2 , Nawel Zemmal1,3 , Didier Schwab4 , and Samir Brahim Belhaouari5 1 Labged Laboratory of Electronic Document Management, Badji Mokhtar University,

Annaba, Algeria [email protected] 2 Computer Science Department, Badji Mokhtar University, 23000 Annaba, Algeria 3 Department of Mathematics and Computer Science, Mohamed Cherif Messaadia University, 41000 Souk-Ahras, Algeria 4 LIG-GETALP Laboratory, Grenoble Alpes University, Grenoble, France 5 College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar

Abstract. Medical text categorization is a valuable area of text classification due to the massive growth in the amount of medical data, most of which is unstructured. Reading and understanding the information contained in millions of medical documents is a time-consuming process. Automatic text classification aims to automatically classify text documents into one or more predefined categories according to several criteria such as the type of output (multi-label or mono label). Feature extraction task plays an important role in text classification. Extracting informative features highly increases the performance of the classification models and reduces the computational complexity. Traditional feature extraction methods are based on handcrafted features which mainly depend on prior knowledge. The use of these features may involve an insignificant representation. Doc2vec is a way to generate a vector of informative and essential features that are specific to a document. In this paper, the impact of combining handcrafted and doc2vec features in the multi-label document classification scenario is analyzed by proposing a system named MUL-MEDTEC. The one-versus-all classification strategy based on logistic regression is adopted in this study to predict for each medical text it to one or several labels. Experimental results based on Ohsumed medical dataset are very encouraging with based classification accuracy equal to 0.92 as global precision. Keywords: Text categorization · Multi-label classification · Medical text · Handcrafted features · Doc2vec

1 Introduction With the rapid growth of the medical text datasets, most of which are unstructured, it is practical to use machine-based algorithms to extract useful knowledge from these data [1]. Unstructured medical documents are complicated and very hard to handle, but they © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 61–71, 2021. https://doi.org/10.1007/978-3-030-70713-2_7

62

R. Chaib et al.

commonly enclose detailed information on patients which is valuable [2]. The use of automatic tools that classify these documents can help alleviate the hard process of finding information. Text classification can be used to solve a variety of problems [3, 4] and has gained quite an importance in the classification of medical text documents [5, 6]. The main task of the automatic text classification approach is to classify the electronic documents into one or more predefined classes (multi-label or mono label) [7]. Different from the common text classification problems, medical data can be multilabeled: medical documents describing a single patient health may contain one or more illnesses [8]. For better mining knowledge from medical text, the classification algorithm requires an appropriate set of features. Extracting informative features highly increases the performance of the classification model and reduces the computational complexity [4]. Traditional feature extraction methods are based on handcrafted features which are mainly depends on the prior knowledge [9]. The use of those features may involve an insignificant representation. Also, it can generate redundant and irrelevant features in the description space of multi-label medical data, which could limit the obtained system performance. Recently, many studies reported the effects of the feature extraction based on document embedding using the doc2vec technique [10–12, 15]. Doc2vec is a way to generate a vector of informative and essential features that are specific to a document [11]. Doc2vec has shown a promote classification performance mainly in the classification of medical documents [4, 12]. In decision stage, many popular classification algorithms have been used for multilabel classification, among them logistic regression. Indeed, many works have adopted this classifier and they have obtained good results [12–14]. What motivates us to use this classifier to solicit our classification problem. The goal behind this work is to analyze and evaluate the impact of combining handcrafted features and doc2vec features in the multi-label document classification scenario. The proposed system MUL-MEDTEC (Multi-Label Medical Text Classification) has two main learning steps: Doc2vec technique was used as an automatic feature extractor from text documents to generate features vectors in the first stage. In order to reinforce the classification model, other handcrafted features were extracted and integrated as additional features. The second step is to classify the aggregated features using the one-versus-all strategy and logistic regression model as kernel classifier. The remainder of this paper is organized as follows: Sect. 2 describes the basic concepts used in this work. Section 3 exhibits the proposed system in detail. Section 4 deals with discussions of the obtained experimental results. Finally, Sect. 5 summarizes the proposed MUL-MEDTEC and gives further research direction.

2 Preliminaries In this section, we will describe the basic concepts adopted in our system. 2.1 Feature Extraction Feature extraction plays a major role in text classification as it has a direct impact on the classification accuracy [15]. It consists of extracting a list of words from a text data,

Improved Multi-label Medical Text Classification

63

then transforming them into a set of features usable by a classifier. The feature extraction algorithm computes the word’s weights in the text, then creates a numeric vector which represents the text’s feature vector [15, 16]. The techniques of vector representation of words can be divided into two categories: • Traditional approaches such as: bag of words and TF-IDF. • Word embedding based approaches such as: Glove, word2vec, doc2vec, Star-Space, and ELMO. 2.2 Doc2Vec Doc2vec technique proposed by Mikolov [17], can be considered as extension of the Word2vec model [18] is an unsupervised technique that uses a deep 3-layer neural network to create vector representations and facilitate similarity of content documents. The Doc2Vec model is based on the same word2vec concepts with only the addition of another vector (paragraph ID) unique to the document, when forming the word vectors W, the document vector ID is also formed. There are two main training methods for doc2vec, the distributed memory paragraph vector model (PV-DM), and the paragraph vector with a distributed word bag (PVDBOW) [14]. The architecture of the Doc2Vec model can be illustrated as follow (Fig. 1):

Fig. 1. The architecture of Doc2Vec model [19].

2.3 Multi-label Classification Multi-label classification is a challenging problem in the field of natural language processing. It is a variant of single-label classification where a set of labels is associated with a single instance. However, there are other classification issues where each instance can be associated with one or more labels. The traditional single-label classification associates the instance X with a single label L from a finite set of labels so the representation with a single label is (X, L). In multi-label problems, each instance X is associated with a subset of labels S or S ∈ L so the representation with multiple labels is (X, S) [20].

64

R. Chaib et al.

2.4 Multi-label Learning Approaches Many multi-label learning algorithms have been proposed in the literature [21, 22]. Multilabel learning approaches can be categorized into three main families: (i) transformation learning approaches which divide the multi-label problem into several mono-label problems, (ii) adaptation learning approaches which adapt mono-label algorithms so that they can process multi-label data, and (iii) ensemble learning approaches which use a set of classifiers from the first or second family of approaches. The used techniques in each paradigm are presented in Fig. 2.

Fig. 2. Multi-label learning approaches.

3 Methodology The objective of our approach is to generate a robust multi-label text classification system for medical text reports. The main steps of our approach are described in Fig. 3.

Fig. 3. Main steps of proposed MUL-MEDTEC system.

Improved Multi-label Medical Text Classification

65

3.1 Medical Text Preprocessing In order, to evaluate the efficiency of the automatic representation of the features for the multi-label classification, a set of basic preprocessing operations were applied which are: • Stop words: permit the elimination of stop words which are the common words in the language such as: “a, an, is, and… Etc.” because they are judged not representative and will not be able to give useful information for our system. Then, punctuation, special characters, hashtags, HTML, URLs, redundant phrases, and rarely used words were all removed from the dataset. • Lowercase: It aims to transform the input text data into lower case. This step allows us to avoid having a multitude of copies for the same word. For example, when calculating the number of words, “Diagnosis” and “diagnosis” will be considered as different words. • Spelling correction: Spell checking is a useful preprocessing step, as it will also reduce multiple copies of words. For example, “disease” and “desease” will be treated as different words even if they are used in the same sense. • Lemmatization: It consists of representing words in their canonical form: Verb will be replaced by its infinitive and name replaced by its singular masculine. • Tokenization: It transforms a text into a series of individual tokens. Each token represents a word for example [“cardiovascular”,“hypertension”,“diagnostic”, “physiology”, “pathology”,…] etc. Moreover, we have also decided to delete rarely used words because those last ones are considered as unusual data. Therefore, removing all these instances will help us reduce the size of the training data and keep that data informative. Figure 4 illustrates an example of textual data sample before and after applying preprocessing steps. 3.2 Feature Extraction As we describe below the importance of feature extraction stage in text representation and the improvement of the classification stage; a cooperation of two families of feature extraction paradigms is adopted in this work to analyze the impact of feature fusion. In fact, classical approaches in text classification are investigated using a different type of manual feature extraction like statistical features and bag of words. Recently, with the growth of deep learning and vector representation, a doc2vec strategy is considered among the effectiveness techniques to generate numerical vector represented the text. This feature vector has the advantage to take into account the statistical, syntactical, semantic relations between the text words. 3.2.1 Used Handcraft Features The handcrafted features in this study can be summarized by a set of statistical features which are:

66

– – – –

R. Chaib et al.

The number of words in each text document. The average length of the words of each text. The number of stop words And finally, the number of digits in a document text.

Fig. 4. An example of used preprocessing steps.

3.2.2 Doc2vec Based Feature Extraction As defined in Sect. 2.2, a doc2vec model is analyzed in our approach to extract the best numerical features. This is guaranteed by a series of empirical tests based on the many parameters such as the size of the generated vector, the window size, the epoch number, and the strategy choice (the distributed memory paragraph vector or the paragraph vector with a distributed word bag). 3.3 Multi-label Classification Stage Our Multi-label classification has the main objective the classification of the aggregated features using the one-versus-all strategy; this last one uses a basic classifier or kernel classifier as decision model; In our approach, the logistic regression model is adopted. This strategy consists of adjusting a classifier by a class which in our case is logistic regression model.

Improved Multi-label Medical Text Classification

67

In order to evaluate the performance of our system, we use the accuracy evaluation criterion defined in Eq. (1), 1 m I (Yi = h(xi )) CA = (1) i=1 m Y represents the true labels, h(xi ) representes the predicted labels, m denotes the number of instances of the test dataset, I(true) = 1, I(false) = 0.

4 Experimental Results and Discussion 4.1 Used Data In this work, the “Ohsumed” dataset is applied to validate our approach. It consists of a medical abstract concerning 23 categories of cardiovascular disease. The main task was to classify those categories where the documents can belong to several classes, for example, an abstract can belong to four types of diseases (Classes: C2, C4, C9, C23). This data collection is available at [23]. 4.2 Evaluation In this study, we carried out several empirical tests to be able to determine the impact of the different parameters of the Doc2Vec by changing the window size, the size of the feature vector, and the number of epochs necessary. Table 1. Obtained results using Doc2Vec based features. Strategy

Window

Size victor

DM

6

150 200

10

150 200

DBOW

6

150 200

10

150 200

Epoch

Precision

50

0.80

100

0.82

50

0.83

100

0.83

50

0.81

100

0.82

50

0.86

100

0.85

50

0.83

100

0.86

50

0.91

100

0.92

50

0.88

100

0.90

50

0.89

100

0.89

68

R. Chaib et al.

The features cooperation impact is analyzed in this study by comparing the multilabel classifier system with and without statistical feature. Table 1 and Table 2 illustrate the obtained results of some tests. Table 2. Obtained results using features cooperation. Strategy

Window

Size victor

DM

6

150 200

10

150 200

DBOW

6

150 200

10

150 200

Epoch

Precision

50

0.79

100

0.80

50

0.81

100

0.82

50

0.80

100

0.80

50

0.84

100

0.84

50

0.81

100

0.80

50

0.89

100

0.90

50

0.86

100

0.89

50

0.87

100

0.89

With the medical records dataset, the best result of our approach is 92% of accuracy with the following parameters: the DBOW method, window size = 6, vector size = 200, epoch = 100, with the cooperation of statistical and automatic features.

5 Discussion The results of our experiments have shown that the automatic representation of the features based on Doc2Vec has given encouraging results by changing the parameters of the Doc2Vec model. After the combination of the features generated by Doc2Vec and the statistical features, our approach gave better results compared to the model with doc2vec only (0.92 and 0.90 respectively). To better analyze the robustness of our system, we used another performance measure which is the MicroF1 score and we have compared our system with other works already existing in the literature which use the same database (Ohsumed) as shown in Table 3.

Improved Multi-label Medical Text Classification

69

Table 3. Comparison with some works that use Ohsumed dataset. Works

Feature extraction Used measure

[24]

BOW

MicroF1 = 59.91

[25]

BOW

MicroF1 = 73.97

[26]

BOW, TF, TF-IDF Accuracy = 0.72

[27]

TF-IDF

MUL-MEDTEC Doc2Vec

MicroF1 = 43.8 MicroF1 = 86.51 Accuracy = 0.92

From Table 3, we can notice that our system surpasses the others existing works. Based on our study, we can say that many approaches can generate too many features and the performance of the classification system depends on the choice of features vectors, which makes the task of learning even harder. To overcome this problem, feature selection proves to be a suitable solution.

6 Conclusion Classification of medical texts is a special case of text classification. In this paper, we have proposed a multi-label medical text classification system (MUL-MEDTEC) that can predict the different types of diseases in a medical record using the cooperation of two types of feature representation: generated features by doc2vec and statistical features to reinforce the learning of the prediction model. The logistic regression classifier is adopted as basic classifier of the one-versus-all strategy to classify the medical texts. Obtained results confirm the robustness of the proposed overall model (with the cooperation of the features) and gives better results with an accuracy of 0.92. In perspective, we want to use a meta-heuristic approach to find the optimal feature vector for better represent textual data.

References 1. Hughes, M., Li, I., Kotoulas, S., Suzumura, T.: Medical text classification using convolutional neural networks. Stud. Health Technol. Inf. 235, 246–250 (2017) 2. Lenivtceva, J., Slasten, E., Kashina, M., Kopanitsa, G..: Applicability of machine learning methods to multi-label medical text classification. In: Krzhizhanovskaya, V., et al. (eds.) Computational Science – ICCS, Springer, Cham, pp. 509–522 (2020) 3. Benzebouchi, N.E., Azizi, N., Hammami, N.E., Schwab, D., Khelaifia, M.C.E., Aldwairi, M.: Authors’ writing styles based authorship identification system using the text representation vector. In: 16th International Multi-Conference on Systems, Signals & Devices (SSD), IEEE, Istanbul, Turkey, pp. 371–376, 21–24 March (2019) 4. Benzebouchi, N.E., Azizi, N., Aldwairi, M., Farah, N.: Multi-classifier system for authorship verification task using word embeddings. In: 2018 2nd International Conference on Natural Language and Speech Processing (ICNLSP), IEEE, Algiers, Algeria, pp. 1–6, 25–26 April (2018)

70

R. Chaib et al.

5. Qing, L., Linhong, W., Xuehai, D.: A novel neural network-based method for medical text classification. Future Int. 11, 255–268 (2019). 6. Alkhatib, W., Rensing, C., Silberbauer, J.: Multi-label text classification using semantic features and dimensionality reduction with autoencoders. In: Gracia, J., Bond, F., McCrae, J., Buitelaar, P., Chiarcos, C., Hellmann, S. (eds.) Language, Data, and Knowledge. LDK 2017, Lecture Notes in Computer Science, vol. 10318. Springer, Cham, pp. 380–394 (2017) 7. Lenc, L., Kral, P.: Word Embeddings for multi-label document classification. In: Proceedings of Recent Advances in Natural Language Processing, Varna, Bulgaria, pp. 431–437, 4–6 September (2017) 8. Guo, Y., Chung, F., Li, G.: An ensemble embedded feature selection method for multilabel clinical text classification. In: IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 823–826 (2016) 9. Azizi, N., Farah, N., Sellami, M.: Ensemble classifier construction for Arabic handwritten recongnition. In: 7th International Workshop on Systems, Signal Processing and their Applications, WoSSPA, pp. 271–274 (2011) 10. Lee, H., Yoon, Y.: Engineering doc2vec for automatic classification of product descriptions on O2O applications. Electron. Commer. Res. 18(3), 433–456 (2017) 11. Kim, D., Seo, D., Cho, S., Kang, P.: Multi-co-training for document classification using various document representations: TF–IDF, LDA, and Doc2Vec. Inf. Sci. 477, 15–29 (2019) 12. Wan, S., Mak, M.-W., Kung, S.-Y.: mPLR-Loc: An adaptive decision multi-label classifier based on penalized logistic regression for protein subcellular localization prediction. Anal. Biochem. 473, 14–27 (2015) 13. Chalkidis, I., Fergadiotis, M., Malakasiotis, P., Androutsopoulos, I.: Large-scale multi-label text classification on eu legislation, arXiv preprint arXiv:1906.02192 (2019) 14. Hoque, M.T., Islam, A., Ahmed, E., Mamun, K.A., Huda, M.N.: Analyzing performance of different machine learning approaches with Doc2vec for classifying sentiment of Bengali natural language. In: 2019 International Conference on Electrical, Computer and Communication Engineering (ECCE). IEEE (2019) 15. Dzisevic, R., Sesok, D.: Text classification using different feature extraction approaches. In: 2019 Open Conference of Electrical, Electronic and Information Sciences (eStream), Vilnius, Lituanie, pp. 1–4 (2019) 16. Resham, N.W., Anuradha, D.: Thakare2.: a review of feature extraction methods for text, International Journal of Advance Engineering and Research Development (2018) 17. Le, Q., Mikolov, T.: Distributed representations of sentences and documents. In: Paper presented at the proceedings of the 31st international conference on international conference on machine learning (2014) 18. Mikolov, T., Sutskever, I., Chen, K., Corrado, G., Dean, J.: Distributed representations of words and phrases and their compositionality. In: Paper Presented at the Proceedings of the 26th International Conference on Neural Information Processing Systems (2013) 19. https://shuzhanfan.github.io/2018/08/understanding-word2vec-and-doc2vec/ Accessed 15 Aug 2020 20. You, X., Zhang, Y., Li, B., Lv, X., Han, J.: VDIF-M: Multi-label classification of vehicle defect information collection based on Seq2seq model. In: Yin, Y., Li, Y., Gao, H., Zhang, J. (eds.) Mobile Computing, Applications, and Services. MobiCASE 2019, Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, vol. 290. Springer, Cham (2019) 21. Pant, P., Sabitha, A.S., Choudhury, T., Dhingra, P.: Multi-label classification trending challenges and approaches. In: Rathore, V., Worring, M., Mishra, D., Joshi, A., Maheshwari, S., (eds.) Emerging Trends in Expert Applications and Security. Advances in Intelligent Systems and Computing, vol. 841, Springer, Singapore, pp. 433–444 (2019)

Improved Multi-label Medical Text Classification

71

22. Ganda, D., Buch, R.: A survey on multi label classification. Recent Trends Program. Lang. 5(1), 19–23 (2018) 23. https://disi.unitn.it/moschitti/corpora.htm Accessed 24 July 2020 24. Al-Salemi, B., Mohd Noah, S.A., Ab Aziz, M.J.: RFBoost: an improved multi-label boosting algorithm and its application to text categorization. Knowl.-Based Syst. 103, 104–117 (2016) 25. Al-Salemi, B., Masri, A., Noah, S.A.M: Feature ranking for enhancing boosting-based multilabel text categorization. Expert Syst. Appl. 113, 531−543 (2018) 26. Parlak, B., Alper, K.U.: The impact of feature selection on medical document classification. In: 2016 11th Iberian Conference on Information Systems and Technologies (CISTI). IEEE (2016) 27. Burkhardt, S., Stefan, K.: Online multi-label dependency topic models for text classification. Mach. Learn. 107(5), 859–886 (2018)

Image Modeling Through Augmented Reality for Skin Allergies Recognition Nur Intan Raihana Ruhaiyem(B) and Nur Amalina Mazlan School of Computer Sciences, Universiti Sains Malaysia, USM, 11800 Gelugor, Penang, Malaysia [email protected]

Abstract. Skin rashes and allergies are common on human body. To date, we could find many skin care products sold not only in pharmacy but also from individual business. However, not all products suitable for all skin types. As a normal human, we sometimes not know the type of rashes or allergies that we faced. Meeting dermatologist would not be the first choice for many patients – given that the fees are expensive especially. Skin rashes can occur to anybody and an early recognition could avoid the rash become worse. Seeking information online would be the first choice, however patients still in high possibilities in mistakenly buy skin care products. Therefore, the development of the augmented reality application for skin rashes and allergies detection is expected can solve the problem. With the help of dermatologist and healthcare people, the information in this application is established and trustable. Among the advantages of this application are the ability in detecting of different types of skin rashes, displaying informative details on the detected skin rashes to reduce wrong judgement on the allergies the patient faced, and reasonable processing speed on mobile screen. Keywords: Augmented reality · Skin rashes · Image processing · 3D modeling · Mobile application

1 Introduction Augmented reality is using technology to integrate the digital information from the user’s environment in real time. By using the augmented reality, the application will allow to overlay new information on top of the existing environment. 3-Dimensional (3D) modeling or 3D program is the main feature in augmented reality application as it will allow the developer to store the 3D animation or digital information in the computer program to an augmented reality marker in the real world. When the device of the augmented reality application receives digital information from a known marker, the application will execute the marker’s code and layer the correct 3D modeling or animation. In this research project, augmented reality will be used to solve the problem of skin rashes. This research work needs to get the data for each type of skin rashes and/or allergies, so that the application can differentiate the types of skin rashes, create 3D modeling for skin rashes and allergies and display the information on the mobile © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 72–79, 2021. https://doi.org/10.1007/978-3-030-70713-2_8

Image Modeling Through Augmented Reality

73

screen. To solve this problem, the augmented reality technology is used to process the patient’s skin problem image through scanning. At the system interface the target object will pop out on the focus area screen. The user needs to locate the camera at the targeted skin rashes within the target object focus area screen. After a complete scanning, the screen will display the expected 3D modeling of the skin rashes type to the user together with the detailed information. Skin rashes or allergies can make someone feel itchy and not comfortable. From baby to elderly are exposed to the skin rashes. There are about 50 types of rashes and in certain cases it looks exactly like the other type of rashes or allergy. Many of which would look like to each other. Although users prefer to seek information such as from Google image to search about the type of the skin rashes, but mistakes would happen perhaps because of user’s perception. Helping users to overcome the skin rashes problem in daily life has been the motivation for this research work as well as drive to the development of this application. This application is believed to help users who are very shy to ask about their skin problems, especially on their private part. The main objective of this system is to develop a complete augmented reality application for skin allergies that can be used by users or patients. The application system also equipped with special feature where the user is able to get to see the expected 3D modeling of the skin rashes and allergies from the scanning, which help the user to understand more what happen to their skin and the type of the skin rashes. Moreover, this application is a new development which is expected to be used by many types of users. Augmented reality (AR) is a technology that produces an information including processing 3D modeling from the user’s real time environment through Unity; a software which utilized for a good visualization and interaction of mobile AR (Kim et al. 2014), with large programming toolsets, (Eriksen et al. 2020) and able to improve the user interface (Kim et al. 2014; Nuryono and Iswanto 2020). Image processing, 3D model and database are the features in AR application. There are four special features of this application, firstly, the rotation speed of the 3D modeling – this will ease user to explore underneath the skin rashes (in video format). Secondly, information text display, for example the information such as symptoms to help users differentiate the type of skin rashes. Thirdly, the Vuforia target manager where the Vuforia SDK can detect and track from the image targets which represent images. Last one is the 3D modeling which will model 3D of that skin rashes to increase users’ understanding of the expected outcome from the image processing (i.e. early diagnosis). Four modules offered in the system development, tabulated in Table 1. The application would give benefits and impact to end users such as it helps the patient to get early diagnosis at home if the patient is too shy to go out, helps parents identify baby’s skin rashes, and it is believed could save more time for multiple users, such as the patients, parents as well as pharmacist. The uniqueness of the application is there is no skin rashes application that use augmented reality where this allows 3D model appear to give the probability or good percentage on early diagnosis using image processing compared to website which generally provide only photos.

74

N. I. R. Ruhaiyem and N. A. Mazlan Table 1. Description of system module.

Module

Description

Image database

The image will be store in database of the Vuforia Target Manager. In Unity, the image will be the AR marker which will trigger the 3D model to be display

3D model of type skin rashes

3D model of skin rashes is created and build by Blender as the platform. The complete 3D model is imported into the Unity. The 3D model will be setting with selected image from database according the types of the skin rashes

Real-time skin allergy tracking

Skin allergy is detected in real-time. The natural features found in the image itself is detected using SDK by comparing these natural features against a known target resource database

Real-time skin allergy result

The result of the skin allergy is in 3D model of the type skin rashes or allergy. The expected 3D model is added (layer) to the skin allergy of the user in real world through the AR application. The information in forms of text also displayed

1.1 Related Work As mentioned earlier, there is no similar mobile application for skin rashes detection through augmented reality technology. There is one mobile application which is very close to this project; called Doctor Mole Skin Cancer app, which is using the standard Asymmetry, Border, Color, Diameter and Risk (ABCDE) approach in order to determine and give instant risk feedback (Doctor Mole 2015). This app focusing on skin cancer and detecting the malignant lesions. Thus, the detection and recognition techniques are different. Doctor Mole is a medical app which used to detect skin cancer by using AR and camera to scan and analyze the suspicious mole in real time. The captured photo is saved and can be used again to see the evolution changes from time to time. Another similar approach for AR technology is tracking with fine object segmentation (TFOS) which originally proposed in year 1989, where it introduced the basic properties of three new variational problems which are suggested by applications to computer vision (Mamford and Shah 1989). In 2013, taking advantage of TFOS, a novel method for on-line, joint object tracking and segmentation was introduced (Konstantinos and Antonis 2013).

Image Modeling Through Augmented Reality

75

1.2 Proposed System Beside diagnosing the skin rashes or allergies in real time, the app also can be used to educate patients and users about skin rashes or allergies through 3D modeling. The main features in this app are; real time detection or scanning that allows the application being used directly to the human body and display 3D model as the outcome or the result from the diagnosis, the offline capabilities, and the application priority where the speed to display the result to user is one of the application priority and the application will not make the user to wait while the application loading.

2 System Analysis, Design, and Implementation This application system can be a service or a product to the client or user. The main features in this application; firstly, the real time detection – where the augmented reality allows the application being used on human body or skin directly and display 3D model as the outcome or result from the detection, and secondly the offline capability – where the apps surely can be used without internet connection. The system capabilities are including the detection speed and the 3D object modeling, where both will take advantage of the smartphone camera (at least 5 MP and above for better detection and results). Like any other application, this app also has its own limitation such as the application can’t be used in the dark or insufficient light place, there is also no social media sharing information and no sound integrated with this system. These limitations are something can be focused on future. The architecture diagram depicted in Fig. 1 shows the overview of the app on how it works from the detection phase until the production of results. For the image database – images of skin rashes are taken from the trusted website and from collection of private photos. The images are stored in database of the Vuforia Target Manager, where the features are tabulated in Table 2. Table 2. Features of the Vuforia target manager. Feature

Description

Rating

This rating is displayed in the Target Manager and the range rating from 0 to 5 for any given image. The higher the rating of an image target, the stronger the tracking and detection ability it contains. Zero rating indicates the image target is not tracked at all by the AR system. Rating at five indicates that the image target is easily tracked by the AR system

Add Target

Add the more the image target by uploading the image using the Add Target button to the Target Manager database

Download Database The downloaded database in form of unity package allow all the image target been import into the Unity

76

N. I. R. Ruhaiyem and N. A. Mazlan

Fig. 1. System architecture of the AR application on skin rashes and allergies.

Image Modeling Through Augmented Reality

77

3 System Testing and Evaluation In Unity system, there is a play button to render the scene. Once it is tuned on, it will show the result whether the app is working well or not. The system is considered successfully working when the 3D model appears after the camera target the image, 3D model can be rotated, and the text is displayed. Other scenario can occur such as with different target image and different setting of lighting such as different values of hue/saturation used. Generally, the test results are good as the speed of tracking and detection is fast.

4 System Interface Design Unity is the main software used for development of the app where the images were created to serve as AR marker (Fig. 2). Here, all settings including camera and image position will be fine-tuned. For 3D models of the skin rashes which used for providing extra information to users, Blender is applied. This software has the ability not only in creating a static 3D image, but also capable to generate motion 3D graphic (Fig. 3). As one of the objectives of this system is to produce an AR application with learning tool on skin rashes, 3D model is produced for learning purpose once the rashes detected and recognized (Fig. 4). Some information including the skin rashes, tips on how to recognize them and tips to heal them will pop out on the screen as well. The evidence that the application meets the requirement and work is in the application, the 3D model is working, the database can be stored and uploaded in a package into Unity, the SDK can detect and track the image and the 3D model rotation is working give the opportunity to user to explore 360° under the skin.

Fig. 2. Unity interface showing the settings of AR camera (left side) and image position (right side) and other settings important for AR marker.

78

N. I. R. Ruhaiyem and N. A. Mazlan

Fig. 3. Blender interface showing all available tools for 3D model development (left side) and render settings (right side).

What are the symptoms of Mosquito bite? 1. Puffy bump on the skin immediately after the bite 2. Reddish brown, itchy bumps 3. Dark spots or bruises caused by itching 4. Mild fever and body ache

Fig. 4. Interface of working AR app, 3D model together with the information will be pop up after the skin rashes successfully detected (which prove that AR marker is working well).

5 Conclusion Using AR as one of a new technology approach for medical field which easily can be used by a lot of people with interaction in real time is something interesting to be explored. Educating or displaying information through 3D model also something should be widely used as it can give a real scenario in real life and easy to understand it (Loke and Ruhaiyem 2020; Teh et al. 2020). Furthermore, this application also provides extra benefit in educating users to know more about skin problems. The important findings found is to know that Unity 3D can be used to create android application in cooperation with AR technology. To create the 3D model, many 3D modeler software (e.g. Maya, Blender, and 3D studio max) has been tested before Blender is chosen. In future however, there are rooms for other technologies could be explored for better AR findings as problem solver application.

Image Modeling Through Augmented Reality

79

References Doctor Mole app 2015 Homepage. https://apkpure.com/doctor-mole-skin-cancer-app/com.rev soft.doctormole. Accessed 02 Sep 2020 Eriksen, K., Nielsen, B.E., Pittelkow, M.: Visualizing 3D molecular structures using an augmented reality app. J. Chem. Educ. 97(5), 1487–1490 (2020) Kim, S.L., Suk, H.J., Kang, J.H., Jung„ J.M., Laine, T., Westlin, J.: Using Unity 3D to facilitate mobile augmented reality game development: In IEEE World Forum on Internet of Things (WF-IoT) (2014). Konstantinos, E.P., Antonis, A.A.: Integrating tracking with fine object segmentation. Image Vis. Comput. 31(10), 771–785 (2013) Loke, H.K., Ruhaiyem, N.I.R.: A conceptual perspective in mathematics through augmented reality and 3D image modeling application. In: 8th International Conference on Multidisciplinary Research on European Proceedings of Social and Behavioral Sciences, pp. 613–620. European Publisher (2020) Mumford, D., Shah, J.: Optimal approximations by piecewise smooth functions and associated variational problems. Commun. Pure Appl. Math. 42(5), 577–685 (1989) Nuryono, A.A., Iswanto, A.M.: Comparative analysis of path-finding algorithm unrestricted virtual object movable for augmented reality. Int. J. Sci. Technol. Res. 1(1), (2020) Teh, Y.X., Ruhaiyem, N.I.R., Syed-Mohamad, S.M.: MYTOXAPP: A mobile system toxicology emergencies through image processing. In: 8th International Conference on Multidisciplinary Research on European Proceedings of Social and Behavioral Sciences, pp. 694–700. European Publisher (2020)

Hybridisation of Optimised Support Vector Machine and Artificial Neural Network for Diabetic Retinopathy Classification Nur Izzati Ab Kader, Umi Kalsom Yusof(B) , and Maziani Sabudin School of Computer Sciences, Universiti Sains Malaysia, 11800 Gelugor, Pulau Pinang, Malaysia [email protected], {umiyusof,maziani}@usm.my

Abstract. Diabetic Retinopathy (DR) is a threatening disease which causes blindness in diabetic patients. With the increasing number of DR cases, diabetic eye screening is a challenging task for experts. Adopting machine learning to create a high accuracy classifier will be able to reduce the burden of diabetic eye screening. Therefore, this paper aims to propose a high accuracy DR classifier using clinical attributes. This study was conducted using nine clinical attributes of 385 diabetic patients, who were already labelled regarding DR, where 79 patients did not suffer from DR (NODR), 161 patients had nonproliferative DR (NPDR), and 145 patients had proliferative DR (PDR). The data was then used to develop a DR classifier through the hybrid of optimised Support Vector Machine (SVM) and Artificial Neural Network (ANN). The experiment results showed that the hybrid classifier had a high accuracy of 94.55. The accuracy yield was higher compared to single classifier. Keywords: Diabetic Retinopathy · Classification · Hybridisation · Support Vector Machine · Neural Network

1 Introduction Diabetic Retinopathy (DR) is one of the complications from Diabetes Mellitus (DM) that affects 1 in 3 persons with DM. It is caused by the damage of retinal blood vessels and light-sensitive tissue at the back of the eye. It may be asymptomatic at an early stage but eventually can cause permanent vision loss if not diagnosed and treated in time [1]. According to the WHO Global report, the number of adults living with diabetes is increasing year by year, which had quadrupled from 108 million in 1980 to 422 million adults in 2016. The rise in Type 2 diabetes and the factors driving it that include overweight and obesity have become the main factors that contribute to this drastic rise [2]. With the increasing prevalence of diabetes nowadays, classification of abnormal retina has become a challenging task as a large number of retinal images need to be diagnosed by ophthalmologists every day. Screening process and early detection of DR play a significant role in helping to reduce the incidence of visual morbidity and blindness. The screening processes are © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 80–90, 2021. https://doi.org/10.1007/978-3-030-70713-2_9

Hybridisation of Optimised Support Vector Machine

81

done manually in most countries [3]. Usually, normal healthy vessels and abnormal vessels are differentiated by ophthalmologists using relative characteristics based on their experience, which can lead to inconsistencies during the grading process [4, 5]. The process is carried out using an ophthalmoscope to inspect the fundus of the eye directly. The pupil will be dilated before it is examined [6]. The retinal qualitative scale, such as mild, moderate, severe, and extreme, is used to evaluate the retina. Occasionally it is useful, however, it is not that effective. It can cause issues of variability in grading as the boundaries between the grades may differ between observers [7] and may also be prone to errors [8]. The high prevalence of the disease is drawing attention from all parties to step up prevention and treatment of the disease. Possibility of collaboration between experts from different areas can be achieved using the technology nowadays [9]. Currently, the applications of computational technique have had a significant impact on the health sector. For instance, supervised machine learning is widely used to predict the presence and absence of the disease [10]. These methods play an essential role in improving the way for diagnosis and treatment of the disease. Amongst the solutions which have been proposed by previous researchers is a DR classification that can assist ophthalmologists in the grading process. Various methods had been done previously for DR classification such as retinal imaging which is a classification technique performed based on the abnormalities found on retinal fundus images [11]. Although it facilitates early detection of DR, additional equipment is required, which is quite cost-prohibitive or sometimes unavailable, especially in rural areas. On the other hand, several DR classifiers have been developed using clinical variables as an alternative to retinal imaging. However, there is still a need for improvement, especially in the accuracy of the classifiers. Therefore, this study proposed to classify DR with the objective to find DR classifiers with optimal or near-optimal performance matrices using a hybrid of optimised Support Vector Machine (SVM) and Artificial Neural Network (ANN), known as SVM-NN. The organisation of the paper starts with Sect. 2 that reviews the literature on the domain problem. Then, Sect. 3 discusses the proposed work. In Sect. 4, the results of the proposed work are evaluated. Finally, Sect. 5 provides a discussion on the results while Sect. 6 concludes the paper.

2 Related Works 2.1 Support Vector Machine Support Vector Machine (SVM) is an algorithm introduced by Corinna Cortes and Vladimir Vapik in 1995 [12]. It is used for classification of input data received by a computing system and also for regression tasks. SVM is categorised as a supervised machine learning method with the objective to classify data points by generating a hyperplane to discriminate between two classes after the input data has been transformed into high-dimensional space [13]. It works based on the concept of principle fitting boundary to the homogeneous region. When a boundary is fitted, the test sample has to be checked, either it lies inside the boundary or not. There is a core of set points that can help identify and fix the boundary. These set points are called support vectors as their function is to support the boundary. The name

82

N. I. Ab Kader et al.

of the vector is because each data point is a vector which is a row of data that contains values for a number of different attributes. The specialty of SVM is that it can efficiently perform a non-linear classification using the kernel trick. The kernel trick function allows the construction of the algorithm without a feature space [14]. 2.2 Artificial Neural Network Artificial Neural Network (ANN) is a computational technique inspired by the nature of neurophysiology. This is analogous to the human brain processing, where synapses are reinforced or weakened [15]. The human body has the capability to receive, process, and send signals through nerve pathways. With the presence of neural cells that are made up of nerve endings and nucleus of the axon, the signals can be transmitted across synapses through chemical treatment. ANN has been applied in various branches related to science and technology since the 1980s [16]. For instance, in the medical sector, ANN has proven useful in the analysis of blood and urine samples of diabetic patients, leukemia classification, diagnosis of tuberculosis, and complicated effusion samples with analysis [17]. There are three layers in the structure of ANN, namely input layer, hidden layer, and output layer. The number of neurons in each layer is determined by the complexity of the system studied. Figure 1 shows the architecture of an Artificial Neural Network. It shows the three layers of ANN, which are input layer, hidden layer, and output layer. The input layer is the information received and will be processed in the hidden layer in order to be an output.

Fig. 1. Architecture of Artificial Neural Network.

ANN is known for its ability to learn from examples, which is a significant trait of intelligence. Instead of following a particular rule specified by human experts, ANN appears to learn from examples (such as input-output relationship) which makes it attractive and exciting. The learning process of ANN can be understood as the problem of updating network architecture and connection weights that will enable the network to function efficiently. The network learns the connection weights from the available training patterns.

Hybridisation of Optimised Support Vector Machine

83

2.3 Hybrid Machine Learning Algorithm The hybrid algorithm is about combining two or more algorithms into a single hybrid algorithm [18]. It is inspired by the possibility of this new algorithm performing better than an individual algorithm. Hybrid is also known as poly-algorithm; when there is a choice at a high level between at least two distinct algorithms, each of which could solve the same problem. It is motivated by an increase in the performance of the execution, depending on both input/output data and computing resources. The hybrid algorithm implementation can be based on the divide, recursive, and conquer method. The divide means to split the step of the first algorithm into smaller steps to see the opportunity to create new parameters and fill in the feature from the subalgorithm. Recursive means repetition of the sub-algorithm, while conquer means the main algorithm controls the sub-algorithm. The first algorithm in this hybrid is denoted as A1 . The second algorithm is denoted as A2 . The hybrid algorithm, H, is developed by modifying the code of A1 by introducing a new parameter, n0 , where algorithm A2 fills in [19].

3 Proposed Work 3.1 Data from the Electronic Health Record The dataset used in this study was provided by the Eye Clinic of the Sakarya University Educational and Research Hospital. The dataset had previously been used in [20], which investigated the DR prediction using Naive Bayes. It contains 385 diabetic patients, who were already labelled regarding DR, where 79 patients were not suffering from DR (class NODR), 161 patients were presented with NPDR, and 145 patients were presented with PDR. NPDR is the moderate stage while PDR is the most severe stage in the DR classification. The attributes in this data are numerical (Haemoglobin, Glycated Haemoglobin, Low-Density Lipoprotein, High-Density Lipoprotein, Diabetes Duration, Creatinine, Triglyceride, Glucose, and URE). 3.2 Performance Evaluation The general purpose of performing classification is to predict the categorical class label for unknown data based on the classification model built by the training data. Confusion matrix is a table that consists of the performance of the classification model in which true values are known. It contains information regarding actual and predicted classification done by a classification system [21]. The accuracy of the algorithm is indicated by the percentage of the test dataset, which is correctly classified by the algorithm. It is used to measure the general performance of the algorithm using the confusion matrix. It is evaluated by calculating the correctly predicted True Positive (TP) and True Negative (TN) classifications based on Eqs. 1–3. Accuracy = (TP + TN ) = TP + FP + TN + FN

(1)

Precision = TP = (TP + FP)

(2)

84

N. I. Ab Kader et al.

Recall = TP = TP + FN

(3)

Apart from that, sensitivity and specificity are measured from the confusion matrix in order to get more specific information on the performance of the algorithm. Sensitivity measures the relevant instances selected while specificity measures the exactness of the algorithm. The words sensitivity and specificity had their origins in screening tests for diseases. Sensitivity is defined as the probability that the test says a person has the disease when, in fact, they do have the disease. In other words, it measures how likely it is for an algorithm to pick the presence of a disease in a person who has it. On the other hand, specificity is defined as the probability that the algorithm says a person does not have the disease when, in fact, they are disease-free. It is also an important measure to be considered. An ideal algorithm should have high sensitivity and high specificity values [22]. It is evaluated by calculating the correctly predicted True Positive and True Negative classifications based on Eqs. 4–5. Sensitivity = TP = TP + FN

(4)

Specificity = TN = TN + FP

(5)

F-measure is also used to measure the performance of the algorithms. F-measure is a harmonic mean of precision (positive predictive value) and recall (exactness of algorithm). According to Van Rijsbergen (1979), F-measure is defined as a combination of recall (R) and precision (P) with equal weight as in Eq. 6. F = 2PR/P + R

(6)

According to [23], precision can be understood as the probability that a randomly chosen predicted positive instance would be relevant while recall is how close we are to a specific target on average. 3.3 The Overall Flow Architecture of Hybrid Optimised Support Vector Machine and Artificial Neural Network SVM-NN is a combination of prediction output from optimised Support Vector Machine and Artificial Neural Network. In SVM-NN, the primary algorithm (A1 ) is SVM while the secondary algorithm (A2 ) is ANN. Figure 2 shows the overall flow architecture of SVM-NN. SVM-NN started with the input initialisation which was the clinical features from the dataset mentioned in Sect. 3.1 and SVM parameters. The kernel used was RBF kernel, because its efficiency is higher. Therefore, two hyperparameters of RBF, Cost, C, and Gamma, U were involved. Prior to the hybrid process, it underwent a phase called hyperparameter optimisation to ensure that the SVM was optimised. The goal is to ensure that the best hyperparameter runs on the SVM. Therefore, C was set to 64 from the optimised SVM and U was set to 0.03. Two essential hyperparameters which are Hidden Layer and Neuron were involved for ANN.

Hybridisation of Optimised Support Vector Machine

85

Fig. 2. The overall flow architecture of SVM-NN.

The input vectors were then propagated to the hidden layer for process (training) step development. It was equipped into an algorithm of backpropagation, also known as automatic differentiation. The backpropagation algorithm is a tool usually used to help ANN change neuron weights and biases if the result is unsatisfactory. Backpropagation aims to refine weights such that ANN can understand ways of mapping arbitrary input to outputs such that the target output can be similar/closer to the actual output. In ANN, this is what is known as “learning”. The initiation process, error estimation, and modified weight were continued until the full number of iterations was reached. Normally, the output is generated and sent to the output layer after the training phase has ended. For SVM-NN, the output layer was truncated and replaced with SVM’s RBF kernel. Thus, the output generated became an input for SVM operation. Next, SVM was trained using the RBF kernel for the processed data. Equation 7 describes the new RBF kernel formulation. This formula is nearly the same as RBF’s original formula. However, they are different in terms of input computation. RBF’s calculation uses the original dataset, while Eq. 7 uses SVM derived from the hidden layer. K(XP , X1 ) = C −Y ||XP −X 1||

2

(7)

86

N. I. Ab Kader et al.

Training was continued until the maximum iteration, maxit = 10, was reached. Upon reaching the maximum iteration, the model was tested using test data. Then, the results produced were analysed.

4 Results Table 1 shows the performance results of SVM-NN model. The first performance measure observed was the accuracy of the algorithm. With the nine inputs fed and processed in the hidden layer and also trained with the kernel function, SVM-NN showed a significant improvement in classification, which was 94.55% accuracy. In the other additional metrics, SVM-NN also obtained a considerable performance measure. The average value of sensitivity and specificity for all classes in SVM-NN were high at 0.9511 and 0.9704, respectively. SVM-NN showed a high F-measure with better precision and recall performance. F-measure conveys an average of precision and recall. The best performance value of F-measure is at 1. The F-measure value obtained by SVM-NN was almost 1 when calculated on average, which was at 0.9500. Table 1. Result of SVM-NN based on each class of Diabetic Retinopathy. Algorithm

Accuracy

Class

Sensitivity

Specificity

Precision

Recall

F-Measure

SVM-NN

94.55

NODR

0.9873

0.9902

0.9630

0.9873

0.9750

NPDR

0.9627

0.9375

0.9172

0.9627

0.9394

PDR

0.9034

0.9833

0.9704

0.9034

0.9357

4.1 Comparison of Results Between SVM-NN, Optimised SVM, and Non-optimised SVM algorithms The performance results’ comparison between SVM-NN algorithms with optimised SVM and non-optimised SVM are tabulated in Table 2 based on each class of DR. Optimised SVM is the SVM algorithm that runs with the optimal parameter setting. In contrast, non-optimised SVM is the SVM algorithm that runs with the default parameter setting. For NODR class, the performance of SVM-NN for each metric was more than 0.9630. The highest output metric obtained by SVM-NN in this class was specificity with 0.9902. High specificity means the algorithm can recognise a person as negative without the illness. It means the model has a good ability to correctly classify patients without DR and with low false positive outcomes. It was proven based on the confusion matrix that showed that SVM-NN correctly identified all NODR patients except one person who was wrongly labelled as NPDR patient. Compared to the optimised and non-optimised SVM, SVM-NN achieved the highest performance for each performance metric. SVM-NN also showed a strong result in the NPDR class, recording more than 0.9172. The highest output metric obtained by SVM-NN in this class was sensitivity with 0.9627.

Hybridisation of Optimised Support Vector Machine

87

High sensitivity means a person with the illness can be identified as positive by the algorithm. Hence, it means the model had successfully classified the patients who were positively at the NPDR level. It had shown progress by an additional 0.1677 relative to the non-optimised SVM. SVM-NN’s lowest output yield for this class was 0.9172 on precision. Nevertheless, when contrasted with optimised and non-optimised SVM, it was still considered high. Table 2. Result for Diabetic Retinopathy classification for SVM-NN compared with optimised SVM and non-optimised SVM. Class

Techniques

Accuracy

Sensitivity

Specificity

Precision

Recall

F-Measure

NODR

SVM-NN Optimised SVM Non-optimised SVM

94.55 85.45 76.62

0.9873 0.9494 0.8481

0.9902 0.9804 0.9771

0.9630 0.9260 0.9054

0.9873 0.9494 0.8481

0.9750 0.9375 0.8758

NPDR

SVM-NN Optimised SVM Non-optimised SVM

94.55 85.45 76.62

0.9627 0.8882 0.7950

0.9375 0.8438 0.7545

0.9172 0.8033 0.6995

0.9627 0.8882 0.7950

0.9394 0.8437 0.7442

PDR

SVM-NN Optimised SVM Non-optimised SVM

94.55 85.45 76.62

0.9034 0.7655 0.7813

0.9833 0.9376 0.7812

0.9704 0.8810 0.6897

0.9034 0.7655 0.7813

0.9357 0.8192 0.7327

SVM-NN sustained the performance quality in the PDR class, with more than 0.9034. The specificity of SVM-NN for this class yielded the highest value compared to the other measures. It means that SVM-NN can recognise people who have no PDR as really not having PDR (patients’ prompt negative outcome). SVM-NN’s precision was also high at 0.9704 for this class. Precision indicates the outcome accuracy on repeated tests. High precision means the evaluation of the result is highly consistent.

120 100 80 60 40 20 0 SVM-NN

OpƟmised SVM NODR

NPDR

SVM PDR

Fig. 3. Comparison of sensitivity between SVM-NN, Optimised SVM, SVM.

88

N. I. Ab Kader et al.

The sensitivity measurement of SVM-NN algorithm was also compared to optimised and non-optimised SVM. Figure 3 shows the comparison of sensitivity between SVMNN, Optimised SVM, and SVM. The SVM-NN sensitivity test was higher than the other two algorithms. This means that the addition of the hybridisation stage enhanced SVM’s ability to identify NODR, NPDR, and PDR patients.

5 Discussion In this research, the hybrid technique between optimised SVM and ANN, also called SVM-NN, which is a method of improving DR classification was carried out. SVM-NN provided substantial results with the right combination of these two algorithms, and it showed improvements compared to the optimised algorithm and the non-optimised algorithm. The hidden layer implemented in the optimised SVM was found to play a role in improving the performance. An experiment was performed where the hidden layer was removed from SVM-NN to test its result without the hidden layer. The result showed that the SVM-NN’s accuracy was lower. It thus proved that the hidden layer plays a significant part in SVM-NN. With the inclusion of the hidden layer in SVM-NN, the input vectors were processed prior to the SVM training. The process of measuring error occurred in the hidden layer, and the operation that occurred in the hidden layer produced an encoding of what the network considers to be the important input features. The number of neurons (nodes) in the hidden layer is an important factor that needs to be carefully determined because it will affect the model, either to be underfitting or overfitting. The number of neurons chosen in this study was an appropriate number for the SVM-NN model. Apart from that, another important factor that eased the cycle of learning mechanism in the hidden layer was the backpropagation algorithm that functions to change weight and biases for neurons. The weight and bias for each node cannot be optimised without the efficiency of the backpropagation algorithm, and a good output cannot be produced.

6 Conclusions In this paper, the study of a hybrid of optimised Support Vector Machine and Artificial Neural Network had been done in detail. From the study, the results showed that SVMNN gave the best performance with 94.55% accuracy and gave a better result compared to the current literature. The implementation of the proposed DR classification with excellent performance will be able to serve as an aid in assisting experts in the diagnosis of DR. It can help the experts in improving decision making and can become a standard guideline for the diagnosis. In addition, it is highly essential to classify and categorise the severity of DR to establish adequate therapy. With the healthcare industry continually looking to improve efficiency and throughput, this study seems to be a satisfactory solution that can provide fast results and timely management of eye screening. Further studies should be conducted to improve the performance of these classification techniques by using a larger dataset. Other performance measures, such as time complexity can also be included.

Hybridisation of Optimised Support Vector Machine

89

Acknowledgement. The authors would like to thank Universiti Sains Malaysia for the assistance it has provided through the Fundamental Research Grant Scheme (203/PKOMP/6711802) to complete the current work.

References 1. Sreekala, X.S., Piri, D., Delen, T., Liu, H.M.: Zolbanin: a data analytics approach to building a clinical decision support system for diabetic retinopathy: developing and deploying a model ensemble. Decis. Support Syst. 101, 12–27 (2017) 2. Zaki, W.M.D.W., et al.: Diabetic retinopathy assessment: towards an automated system. Biomed. Sig. Process. Control 24, 72–82 (2016) 3. Chen, W., Yang, B., Li, J., Wang, J.: An approach to detecting diabetic retinopathy based on integrated shallow convolutional neural networks. IEEE Access 8, 178552–178562 (2020) 4. Ramos, L., Novo, J., Rouco, J., Romeo, S., Álvarez, M.D., Ortega, M.: Retinal vascular tortuosity assessment: inter-intra expert analysis and correlation with computational measurements. BMC Med. Res. Methodol. 18(1), 1–11 (2018) 5. Gargeya, R., Leng, T.: Automated identification of diabetic retinopathy using deep learning. Ophthalmology 124(7), 962–969 (2017) 6. Mapayi, T., Tapamo, J.-R., Viriri, S., Adio, A.: Automatic retinal vessel detection and tortuosity measurement. Image Anal. Stereology. 35, 117–135 (2016) 7. Wu, B., Zhu, W., Shi, F., Zhu, S., Chen, X.: Automatic detection of microaneurysms in retinal fundus images. Comput. Med. Imaging Graph. 55, 106–112 (2017) 8. Qomariah, D.U.N., Tjandrasa, H., Fatichah, C.: Classification of Diabetic Retinopathy and normal retinal images using CNN and SVM. In: 2019 12th International Conference on Information & Communication Technology and System (ICTS), vol. 1, pp. 152–157 (2019) 9. Amin, J., Sharif, M., Yasmin, M.: A review on recent developments for detection of diabetic retinopathy. Scientifica 2016, 1–20 (2016) 10. Aher, J., Singh, P., Shah, M.: Diabetic Eye Disease Detection Using Machine Learning. Techniques 5, 725 (2020) 11. Ullah, H., Saba, T., Islam, N., Abbas, N., Rehman, A., Mehmood, Z., Anjum, A.: An ensemble classification of exudates in color fundus images using an evolutionary algorithm based optimal features selection. Microsc. Res. Tech. 82(4), 361–372 (2019) 12. Carrera, E.V., González, A., Carrera, R.: Automated detection of diabetic retinopathy using SVM. In: 2017 IEEE XXIV International Conference on Electronics, Electrical Engineering and Computing (INTERCON), vol.1, pp. 1–4 (2017) 13. Huang, H.-Y., Lin, C.-J.: Linear and kernel classification: When to use which? In: Proceedings of the 2016 SIAM International Conference on Data Mining, SIAM, vol.1, pp. 216–224 (2016) 14. Teixeira Jr, L.A., et al.: Artificial neural network and wavelet decomposition in the forecast of global horizontal solar radiation. Pesquisa Operacional. 35, 73–90 (2015) 15. Rigby, M., Anthonisen, M., Chua, X.Y., Kaplan, A., Fournier, A.E., Grütter, P.: Building an artificial neural network with neurons. AIP Adv. 9(7), 1–1 (2019) 16. Dubey, K.B., Shrivastava, D.: Forestalling growth rate in type ii diabetic patients using data mining and artificial neural networks: an intense survey. Int. J. Comput. Eng. Technol. 10(3), 31–38 (2019) 17. Rawat, A.S., Rana, A., Kumar, A., Bagwari, A.: Application of multi-layer artificial neural network in the diagnosis system: a systematic review. IAES Int. J. Artif. Intell. 7(3), 138–142 (2018) 18. Sayed, G.I., Hassanien, A.E.: A hybrid SA-MFO algorithm for function optimization and engineering design problems. Complex Intell. Syst. 4(3), 195–212 (2018)

90

N. I. Ab Kader et al.

19. Awad, M.: Enhanced hybrid method of divide-and-conquer and rbf neural networks for function approximation of complex problems. Turkish J. Electr. Eng. Comput. Sci. 25, 1095–1105 (2017) 20. Evirgen, H., Cerkezi, M.: Prediction and diagnosis of diabetic retinopathy using data mining technique. Turkish Online J. Sci. Technol. 4, 32–37 (2014) 21. Deng, X., Liu, Q., Deng, Y., Mahadevan, S.: An improved method to construct basic probability assignment based on the confusion matrix for classification problem. Inf. Sci. 340, 250–261 (2016) 22. Zalesky, A., Fornito, A., Cocchi, L., Gollo, L.L., van den Heuvel, M.P., Breakspear, M.: Connectome sensitivity or specificity: which is more important? Neuroimage 142, 407–420 (2016) 23. Sajjadi, M.S., Bachem, O., Lucic, M., Bousquet, O., Gelly, S.: Assessing generative models via precision and recall. Adv. Neural. Inf. Process. Syst. 1, 5228–5237 (2018)

A Habit-Change Support Web-Based System with Big Data Analytical Features for Hospitals (Doctive) Cheryll Anne Augustine and Pantea Keikhosrokiani(B) School of Computer Sciences, Universiti Sains Malaysia, 11800 Minden, Penang, Malaysia [email protected]

Abstract. Even with the advancement of medical services, we still see an increase in mortality rate around the world especially due to heart disease, where it constantly remains as the number one cause of death globally. In order for an individual to protect their health, they are required to adopt healthy eating and practice regular exercises which also means that they have to adapt to a habit change in their daily routine. This healthy habit change does not only protect against heart diseases but also other chronic diseases such as cancer and stroke. Therefore, a habit-change support web-based system with big data analytics and decision-making features called Doctive is developed in this study to lower the risks of heart diseases. Doctive is targeted for hospital authorities to monitor patients and their habits and to prescribe medication and advice based on patients’ habits and gathered information. Furthermore, this system also provides emergency assistance for patients based on their current location. This proposed system, would also be beneficial in collecting and organizing patients’ information to ease access and speed the process of data entry and retrieval. The system was tested and evaluated by 5 people who were medically qualified or with knowledge and expertise in the field of data analytics and visualization. After gathering their opinionated responses, the results were tabulated and analyzed to be taken into consideration for improvements and to garner ideas for the future development of the system. Doctive can be useful for healthcare providers and developers. Keywords: Habit-change · Medical information system · Web-based system · Big data analytics · Decision-making

1 Introduction As the use of technology spreads rapidly in various fields, the advancement in the medical field in the aspect of data gathering and processing has not bloomed to its fullness. This may be due to the immense number of information sent to hospitals where data gathered is not segregated and analyzed in order optimally. Besides, hospitals could also be lacking of useful input and real time data from patients. This rings a warning sign as the mortality rate sees an increase due to heart and chronic diseases, based on the survey done by World Health Organization [1], where, heart disease remains as the leading © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 91–101, 2021. https://doi.org/10.1007/978-3-030-70713-2_10

92

C. A. Augustine and P. Keikhosrokiani

cause of death around the world. The current systems that are available in hospitals lack the ability to monitor patients’ bad habits. Therefore, a Habit-Change Support WebBased System with Big Data Analytical Features for Hospitals (called “Doctive”) was developed on June 2020 in Penang, Malaysia for hospitals that supports patients based on their emotional and persuasive habit-change patterns. The proposed system can help to collect and store information related to patients’ bad habits and try to provide them with the prescriptions and supervision they need by qualified and authorized personals at the right time based on their habits and progression. Most of the diseases that attacks the community is largely based on the habits of individuals. Our habits vary from one person to another, therefore monitoring and promoting change in unhealthy habits will aid in creating a healthier generation. Patients who suffer from cardiovascular or other chronic diseases may be affected by their unhealthy habits such as lack of exercise, unhealthy eating patterns, lack of sleep, smoking and much more. There are recent studies that also shows the importance of early detection and diagnosis of diseases thorough healthcare systems using Big Data [2]. Therefore, a habit-change towards the proper and healthy direction is all that is needed to prevent detrimental health issues in the future. The developed web-based system works hand in hand with a mobile application that tracks patients’ daily activities, habits and stores it in a Firebase cloud platform, so that data can be collected, identified and analyzed by the hospital authorities for further supervision, medical treatments, specified prescriptions and advices. This collected data will be very helpful as Big Data analytics which is now widely used helps in providing precise and effective healthcare services by enabling data sharing and performing analytical calculations and analysis to provide strategic planning and better decision making [3, 4]. This new project would also help to improve the hospital’s patient management system and to be able to provide early care by reaching out to the society. This paper reviewed most of the existing healthcare systems and thus introduced the proposed habit-change support web-based system with big data analytical features for hospitals, called Doctive. The system requirement, design, development methodology, big data analytics, decision-making features, tests and evaluation are included in this paper. The paper is finally wrapped up with concluding remarks and future directions.

2 Related Work 2.1 Existing Systems There are existing system which are somewhat similar to the proposed system but lacks some functionalities. Two existing machine that closely works with the cardiology department to track patient’s heart rhythms are introduced in this paper. However, they do not constantly record the rhythms for over a month, nor do they monitor patients’ habits which actually plays a vital role. Holter Monitor The Holter Monitor [5] is a device that constantly records the heart’s rhythms. This monitoring device is worn by patients for about 24 to 48 h as they perform normal daily activities. This test is performed by sticking electrodes which are small conducting

A Habit-Change Support Web-Based System

93

patches onto patient’s chest. These electrodes are attached by wired connectors to a small recording monitor where it can be carried in a pouch or pocket. This monitor runs on batteries. After 24 to 48 h, this monitor should be returned to the doctor or the health care provider, to enable them to take a look at the collected records to see if there is any abnormal heart rhythms. Abnormal heart rhythms may include various arrhythmias, certain changes may also indicate that the heart is not getting enough oxygen supply. Adding to that, this system stresses on the importance of patients’ to record their symptoms and activities separately and manually in order for the provider to match them with the findings of the Holter Monitor when they attend the follow up sessions. Cardiac Event Recorder A Cardiac event recorder [6] is a battery-powered portable device that patients’ can control to tape-record their heart’s electrical activity, electrocardiogram (ECG) when they have symptoms such as fast or slow heartbeats, dizziness, or the feeling of fainting. It can also be used to monitor how patients respond to a certain type of medication. Some of these ‘cardiac event monitors’ are able to store patients’ ECG in the memory monitor. There are two types of ‘event recorders’ such as the loop memory monitor and the symptom event monitor. • Loop Memory Monitor A small device that can be programmed to record your ECG for a certain period of time, such as 5 min. A button has to be pushed to activate it and the ECG will be stored for a period of time before and during the symptoms. • Symptom Event Monitor A hand-held device or to be worn on the wrist. When patients’ feel an irregular heartbeat, they should place the monitor on their chest and activate a recording button. This device only records the ECG reading when it is activated. These devices are able to send the ECG recording by telephone to a specific unit in the hospital for the doctor to review. However, based on investigations, there are also some existing healthcare management systems in Malaysia such as Med-Pro Care, H-MagSys and My1HealthCare Solution [7] that includes clinical and administrative functions but lacks health monitoring and big data analytics features which provides decision support for medical professionals to make better judgements in improving the habits of patients, which is used in the Doctive system.

3 System Requirements and Analysis 3.1 Proposed Solutions The proposed habit-change support web-based system aims to collect, organize and maintain patients’ information to ease access and retrieval of data. In addition, it monitors patients and their habits through the proposed system. Habit change performance will be analyzed using big data analytics feature and the decision tree will be created for further prescription by the medical experts. The main motivation of this system is to build a

94

C. A. Augustine and P. Keikhosrokiani

healthier community by lowering the increasing percentage of chronic diseases due to the practices of unhealthy habits among people. Habit affects an individual’s mood and habits can be changed and improved over time. This has led to the idea of a HabitChange Support Web-Based System for Hospitals [8–12]. The main idea life-cycle of the proposed system (Fig. 1) is to recognize and improve the bad habits and the health status of individuals.

Fig. 1. System idea life-cycle

Their data is obtained from the Firebase cloud storage platform, where it will be classified, organized and analyzed by the system to be turned into useful information in order to help patients with their health status. Professional hospital authorities will then analyze patients’ habit change performance through machine learning and data visualization tools and provide prescription or advices. Patients’ habits will then be remotely monitored closely using Internet of Things (IoT) just as co-implemented in this system [13], in order to provide new prescription or suggest other medical alternatives based on the progress and changes in habits. 3.2 Development Methodology The development of the Doctive system is based on the System Development Life Cycle. This development cycle mainly consists of four main phases which are the modelling, assessment, design and prototype phase. The modelling phase involves the gathering of system requirements which includes hardware and software. The second phase, assessment, is carried out to understand and gather user requirements for this new system. This is a crucial phase as, user assurance is needed in order to be confident in completing the project, to avoid rejection and dissatisfaction in later stages. Therefore, an interview was carried out to gather users’ opinions. Next, the design phase includes requirement to design, feasibility study, analysis and design of the Doctive system. An architectural design specifies the hardware, software and environment of the new project. The fourth and last phase is the prototyping phase which is completed with constructing, coding or testing and evaluating the new system.

A Habit-Change Support Web-Based System

95

3.3 System Architecture The system architecture of the “Doctive” system is described in detail in the Fig. 2. There are two parts of the Doctive system, it partly works together with the Behabit mobile application that is connected to a smart watch. This Behabit mobile application collects and tracks end user’s demographic details which also includes, heart rate, habits and tries to change patients’ bad habits using emotional-persuasive features. These collection of data will then be stored to the Firebase cloud platform for the Doctive system to obtain. The Doctive system plays a huge role in collecting, organizing and analyzing these data so that the users will be able to maintain a fit and healthy lifestyle, far from potential chronic diseases. Doctive system collects data from the cloud platform and applies machine learning tools and technique to help doctors monitor progress and changes in user or patients’ habits. Collected data is stored in the form of the log file in WEKA’s Attribute-Relation File Format (ARFF). This file is used as input to train the classifier. Apart from that, the data is also used to create data visualizations in Tableau to assist doctors’ analysis.

Fig. 2. Doctive system architecture

The algorithm used in helping doctor or health care providers predict the right and suitable prescription and advices for specific users or patients is the J48 predictive decision tree. For example, ArffViewer tool is provided by Weka, which imports data from a comma-separated values (CSV) file and saves it in ARFF format. This file is then fed to Weka to train the classifier. The classifier then produces the J48 algorithm confusion matrix, along with the decision tree.

96

C. A. Augustine and P. Keikhosrokiani

4 Test and Evaluation of the System In order to test and evaluate Doctive system, testing strategy is utilized as shown in Fig. 3 which includes unit testing, integration testing, system testing and acceptance testing. Unit testing is a level of software testing where individual units/components of a software or system is tested. It is aimed to affirm that each unit of the system performs its functions as designed. A unit is the smallest testable part of any software. It usually has one or more input with a single output. In SDLC, unit testing is the first level of testing done before integration testing. PHPUnit was created by Sebastian Bergmann and it is the most popular unit testing framework independent library for testing PHP codes but it involves writing tests manually and running them, which takes more time.

Fig. 3. Proposed testing strategy for doctive system

Integration Testing is a software testing level where individual units are combined and tested to verify whether they are functioning as they are intend to when integrated. The main purpose is to check the interface between modules and to identify defects in the interaction between these software modules when they are integrated. Integration testing is a systematic methodology for installing a software system when conducting checks to discover interfacing-related errors. It can ensure the exception for parameter, function, run-time and incompatibility between object interactions. Integration testing performed from time to time, starting from project development is advisable, to ensure that everything runs smoothly. System Testing is a software test level which tests a complete and integrated software. This test was performed at every phase to determine the system’s compliance with the stated requirements. System testing helps reduce troubleshooting after delivery, and service calls. When there is a chance of any error occurrence, urgent action must be taken to fine-tune the system. System testing was carried out for the habit change and prescription module of the Doctive system to ensure, the habit-change can be observed from week to week and for the latter, to check if doctor’s prescriptions are able to reach the designated patients. User Acceptance testing of the Doctive system was carried out among a limited number of people as it required, individuals with some medical expertise to provide their

A Habit-Change Support Web-Based System

97

feedback and opinions on the system. Therefore, a questionnaire was created and distributed to 5 people who were either medically qualified or has knowledge or expertise in the field of data analytics and visualization. After gathering their opinionated responses, the results were tabulated and analyzed to be taken into consideration for improvements and ideas for the future development of the system. 4.1 Result The system shows the results of the big data analysis based on patient’s habit in two forms, the decision tree and visualizations from Tableau. The decision tree as shown in the Fig. 4 below will then help doctors or health care providers to predict user’s mood based on the level of exercise performed. For example, when user performs light exercise, and burns calories less than or equals to 36 kcal, their mood is dull. But users performs medium exercise, and burns more than 43 kcal with walking for more than 4266 steps a day, their mood is excited. Therefore, doctors or health care providers can advise users or patients to constantly perform medium level exercise and walk for more than 4200 steps a day to maintain a happy and cheerful mood.

Fig. 4. J48 decision tree model for doctive system

Besides, Doctive system also aids in the hospital administration activities such as registering and managing patients, storing and keeping track of patients’ medical history, scheduling appointments and follow-up sessions and most importantly, reaching out to patients in times of emergency. These hospital administrational data of patients will be stored in a centralized local database (phpMyAdmin) of the hospital. Apart from analyzing data in Weka, Doctive system also includes big data analytics and visualizations from Tableau that helps in simplifying raw data into very easily understandable formats. Figure 5 portrays data of an individual for a period of one month with a Body Mass Index (BMI) of 22.0 in the Normal category and Basal Metabolic Rate (BMR) of 1252 calories per day. Therefore this individual can maintain their healthy weight by making sure that their calorie burn is consistent with their calorie intake. The definition and calculation of BMR is as follows:

98

C. A. Augustine and P. Keikhosrokiani

Basal Metabolic Rate Equation by Mifflin-St Jeor. Basal metabolic rate (BMR) is the total number of calories that an individual needs to perform basic, life-sustaining functions. These basal functions include circulation, breathing, cell production, nutrient processing, protein synthesis, and ion transport [14]. (M ) = (10 × weight in kg) + (6.25 × height in cm) − (5 × age in years) + 5 (F) = (10 × weight in kg) + (6.25 × height in cm) − (5 × age in years) + 161 BMR multiplied by the activity factor of an individual based on the activity level, determines the amount of calories needed by that particular individual. The graph in Fig. 5 includes data such as date by weeks for a month, average number of steps, activity level (Sedentary, Low Active, Moderately Active, Active, Highly Active) which is the standardization provided by the World Health Organization (WHO), average calories, Behabit points (labelled on the graph as Very Bad, Bad, Average, Good, Excellent) and mood (Dull, Excited, Happy, Normal). As it is said that, there is a strong connection between good mental health and good physical health, and vice versa by the Harvard Medical School [15], there are also studies on the influence and relationship of mood on health [16, 17]. Based on the analysis, a clear visualization can be made about the mood of the individual based on their steps and calories weekly for a month. The graph also helps doctors filter the moods based on its colored categories.

Fig. 5. Habit Change analysis and visualization in Tableau

A Habit-Change Support Web-Based System

99

When the steps level of an individual is at a Sedentary level which is below 5000 steps and with low calorie burnt, the mood of an individual falls between Dull, Normal and Happy, but has a wider mood area of normal which also includes dull. This also show that their Behabit point within this period falls under Very Bad, Bad and Average. Comparatively, when an individual is at the Low Active level, which is between 5000 to 7500 steps, the mood of an individual shows more excitement and happiness with only a little Normal mood. At this point, their Behabit points is also labelled as Good. And topping it all, when an individual achieves steps count between 7500 to 9999 steps (Moderate Active), and burns a higher number of calories with an average of 125 cal, their mood is never dull nor normal. It only keeps an individual in a positive mood between happy and excited. This also helps them achieve an Excellent score in the Behabit points. Based on an example as shown in Table 1 a doctor can conclude, which steps level should be advised and is practical to be practiced by an individual to keep them in a positive mood with a healthy lifestyle. For example, concluding based on the data visualization, when the patient performs, Low Active level of steps, they are generally in a positive mood ranging from normal, happy and excited. Referring to that, the patient can be advised to continuously perform Low Active steps level between 5000 to 7499 steps a day with medium level exercise, for a month to take the patient off the Sedentary lifestyle and make changes to their habits progressively. Table 1. Activity classification based on WHO

5 Conclusion and Future Work Doctive system which is a habit-change support web-based system with big data analytical features for hospitals was developed, tested, and evaluated successfully in June 2020 in Penang, Malaysia. Doctive has been built to enforce healthy habit change and reduce the risk of heart diseases in the community which has been the number one killer globally. Based on the evaluation results from medical experts, the feedbacks were positive and constructive. This system has been developed after lots of consideration on the requirements of the end user and suggesting ways on how to further ease their responsibilities. Therefore, the feedbacks and ideas obtained from the user acceptance evaluation will also be taken into consideration for the further improvement and betterment of the system. Some important feature that can be added to the system in the

100

C. A. Augustine and P. Keikhosrokiani

future are the features that will help the community in leading a healthy lifestyle for the body and mind such as predictive habit change. This is becoming more and more important as the number of people especially teenagers, falling into depression is on the rise. Furthermore, ECG monitoring features can be added to the system that stores the readings into the system rather than printing out the long list of ECG reading to be filed physically. This can then be stored as reference for upcoming doctor visits and can be retrieved with ease. ECG is vital as it detects and provides a more accurate result and reading of the heart’s rhythm. More features will also be added in order to provide a complete patient management system for large scale hospitals in the very near future. Acknowledgment. The authors are thankful to School of Computer Sciences, and Division of Research & Innovation, USM for providing financial support from Short Term Grant (304/PKOMP/6315435) granted to Dr Pantea Keikhosrokiani.

References 1. Who.int. The top 10 causes of death. https://www.who.int/news-room/fact-sheets/detail/thetop-10-causes-of-death. Accessed 9 Oct 2019 2. Jagadeeswari, V., Subramaniyaswamy, V., Logesh, R., Vijayakumar, V.: A study on medical internet of things and big data in personalized healthcare system. Health Inf. Sci. Syst. 6(1), 14 (2018) 3. Madanian, S., Parry, D.: IoT, cloud computing and big data: integrated framework for healthcare in disasters. Stud. Health Technol. Inform. 264, 998–1002 (2019) 4. Kolasa, K., Goettsch, W., Petrova, G., Berler, A.: ‘Without data, you’re just another person with an opinion’. Expert Rev. Pharmacoecon Outcomes Res. 20(2), 147–154 (2020) 5. Holter monitor (24 h): MedlinePlus Medical Encyclopedia. https://medlineplus.gov/ency/art icle/003877.htm. Accessed 3 Nov 2019 6. www.heart.org. Cardiac Event Recorder, https://www.heart.org/en/health-topics/arrhythmia/ prevention–treatment-of-arrhythmia/cardiac-event-recorder. Accessed 3 Nov 2019 7. Hospital and Healthcare Management Systems for Hospitals, Medical Centres, Specialist Clinics and General Practitioners. http://www.my1healthcare.com/modules/web/index.php. Accessed 9 Oct 2020 8. Keikhosrokiani, P.: Emotional-persuasive and habit-change assessment of mobile medical information Systems (mMIS). In: Keikhosrokiani, P. (ed.) Perspectives in the Development of Mobile Medical Information Systems, Academic Press, pp. 101–109 (2020) 9. Keikhosrokiani, P., Mustaffa, N., Zakaria, N., Baharudin, A.S.: User behavioral intention toward using mobile healthcare system. In: Consumer-driven technologies in healthcare: breakthroughs in research and practice: IGI Global, pp. 429–444 (2019) 10. Keikhosrokiani, P.: Behavioral intention to use of Mobile Medical Information System (mMIS). In: Keikhosrokiani, P. (ed.) Perspectives in the Development of Mobile Medical Information Systems, Academic Press, pp. 57–73 (2020) 11. Keikhosrokiani, P., Mustaffa, N., Zakaria, N.: Success factors in developing iHeart as a patientcentric healthcare system: a multi-group analysis. Telematics Inf. 35(4), 753–775 (2018) 12. Keikhosrokiani, P., et al.: Assessment of a medical information system: the mediating role of use and user satisfaction on the success of human interaction with the mobile healthcare system (iHeart). Cogn. Technol. Work 22(2), 281–305 (2020)

A Habit-Change Support Web-Based System

101

13. Fernández-Caramés, T.M., Froiz-Míguez, I., Blanco-Novoa, O., Fraga-Lamas, P.: enabling the internet of mobile crowdsourcing health things: a mobile fog computing, blockchain and iot based continuous glucose monitoring system for diabetes mellitus research and care. Sensors (Basel) 19(15), 3319 (2019) 14. Frey, M.: How to Change Your Basal Metabolic Rate for Weight Loss. https://www.verywe llfit.com/what-is-bmr-or-basal-metabolic-rate-3495380. Accessed 5 Jun 2020 15. Publishing, H.H.: Mind & Mood. https://www.health.harvard.edu/topics/mind-and-mood. Accessed 11 Oct 2020 16. Salovey, P., Birnbaum, D.: Influence of mood on health-relevant cognitions. J. Pers. Soc. Psychol. 57(3), 539–551 (1989) 17. Yates, J.A., Clare, L., Woods, R.T.: What is the Relationship between Health, Mood, and Mild Cognitive Impairment? J. Alzheimers Dis. 2017 55(3), 1183–1193 (2016)

An Architecture for Intelligent Diagnosing Diabetic Types and Complications Based on Symptoms Gunasekar Thangarasu1(B) , P. D. D. Dominic2 , and Kayalvizhi Subramanian3 1 Department of Professional Industry Driven Education, MAHSA University,

Jenjarom, Malaysia 2 Department of Computer and Information Science, University Technology PETRONAS,

Seri Iskandar, Malaysia [email protected] 3 Department of Fundamental and Applied Sciences, University Technology PETRONAS, Seri Iskandar, Malaysia

Abstract. Information and communication technology can play a vital role in improving healthcare services by providing new and efficient ways of diagnosing diseases. Diabetic is recognized as the fastest-growing disease in the world. Due to insufficient diagnostic mechanisms, the number of undiagnosed diabetes has been increasing day by day. And it leads to creating long term complications such as neuropathy, nephropathy, foot gangrene and so on. The objective of this study is to design an intelligent architecture for diagnosing diabetes effectively based on the individual physical symptoms. The architecture has been designed by utilizing the combination of neural networks, data clustering algorithms and fuzzy logic techniques. Subsequently, a prototype system has been developed to validate against the diagnostic architecture on the aspect of efficiency and accuracy of diagnosing diabetes, and its types and complications. The overall qualitative findings from this study scored very high, which is 94.50% accurate. Keywords: Diabetes · Complications · Neural networks · Fuzzy logics and clustering

1 Introduction Information and Communication Technology (ICT) can help in coping with the information explosion. Information refers to any communication or representation of knowledge such as facts, data or opinions in any medium including textual, numerical, graphic Cartographic, narrative or audiovisual forms. Technology is the practical form of scientific knowledge or the science of the application of knowledge to practice. Today, ICT is used in a wide range of fields especially in healthcare. The use of information technology is to improve the healthcare system by saving cost, increasing patient safety and improvising the quality of healthcare. Digital technology will continue to be the catalyst for innovative initiatives in the healthcare sector. The computerized decision support systems sued © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 102–110, 2021. https://doi.org/10.1007/978-3-030-70713-2_11

An Architecture for Intelligent Diagnosing Diabetic Types

103

in a hospital are very useful for complete, accurate decision making. Among the benefits that information technology renders in the healthcare industry is the record-keeping of patients and their information such as symptoms, diagnosis treatment and etc. [1]. Diagnosis of diseases is an important and difficult task in medicine that primarily focuses on the various causes and functions that are interrelated. In general, a variety of factors and observations are taken into account for the diagnosis of real diseases [2, 3]. Therefore, doctors typically apply their specialized expertise, experience and hypothesis about illness and eventually assess the patient as the best scientific way to solve the issue or formulate the treatment. The clinical database also aids in identifying patterns and collecting the know-how to improve the diagnosis and treatment of patients. Many health professionals often use medical records in major medical evaluations. In the last 40 years, several statistical instruments have conducted a little research in the field of medicine. Nevertheless, over the past 20 years, the use of neural networks, fusion logic and other methods has steadily grown [4]. In the last five decades, developments in computer technology have been promoted through the development of artificial intelligence, expert systems and decision-making system which enables individuals to carry out specific tasks. Since its inception, the health industry has been using Artificial Intelligence (AI), Expert Systems (ES), Decision Support Systems (DSS) and in medical diagnostic applications, fuzzy reasoning, case-based thinking and the neural network have gradually increased [5]. In this research, the problem of study originated from the real diagnosis of diabetes. Few diabetes studies were performed and their trends and methods were not sufficient at the present time. According to the World Health Organization (WHO), more than 50% of diabetes is still undiagnosed worldwide [6]. The use of diabetes diagnosis included in the clinical details that are blood test results measures glychohemogloin, fast plasma glucose and oral glucose tolerance. Both diabetes-related blood tests include the blood drawing and transfer of the sample to a healthcare professional office in a clinical laboratory. To ensure correct test results, a blood laboratory analysis is necessary. Random plasma glucose monitoring in a number of diabetes diagnostic cases is also used during daily monitoring. For all diabetes diagnostic examinations, a second calculation [7, 8] is needed. So it’s the requirement to do the research for diagnosing the diabetes types and its complications in a faster manner without going through the series of medical tests and waiting for the physician’s appointment. The principle objective of this study is to design and improve the diagnostic framework for intelligent diagnostic architecture to evaluate the types and complications of diabetes mellitus. Health professionals can use this groundbreaking approach of improving the diagnosis of diabetes.

2 Review of Literature Saiful Rahman founded the Decision supporting diabetes diagnosis [9]. In his study, sign, symptoms, risk factors for diabetic disease and the results of a physical examination were used for data collection. The physician asked about the patient’s symptoms. The risk of disease was measured and evaluated on the basis of system response information obtained from 55 Types-I diabetes patients for instances, low probability, medium chance and

104

G. Thangarasu et al.

moderate risk diabetic or not. The system has shown an incredible performance, which has an accuracy of 98%. Innovation by intelligent agents has been used to design and improve the system. The study of Matsumoto et al. [10] have acknowledged in its report the information given to the doctor’s clinically definitive signs, symptoms, comments and research facilities. It proposed the idea of patient models and he patterns of disease, a structured algorithm for the diagnosis and a real context that was built with neural systems for medicinal research. As Gultepe et al., pointed out [11], it is necessary to recognize early sepsis in order to keep traveling in the more severe phases of illness, with one in four consequences. The Bayesian method is developed using systemic triggering disorder reaction criteria, mean vessel weight and lactate levels of sepsis patients. The following system reveals an appropriate conation between the levels of lactate and sepsis. Dakua, et al. [12] confirmed that one of the prevalent and debilitating diseases in the adult population worldwide is a cerebrovascular disorder. It causes cerebral vessels to break down within the brain that leads to a haemorrhage of the subarchnoids. A clinical workflow model was introduced into their work to assist the endovascular in selecting the type of stent-related treatment for cerebral aneurysms. The findings suggest the clinical potential benefit of the proposed computational workflow. Raiter, et al. [13], on the other hand, used inconspicuous and basic sensor devices to track the well-being of individuals fascinated by a healthy lifestyle. The information gathered from body sensor and survey data is analyzed by allowing improvement plans and suggestions to be extracted with the aid of feedback, training and motivations and techniques are recognized as ideally helping individuals to achieve their goals in daily life. They have a couple of weeks of customized take care frameworks that can be used as part of their own home. According to Vyssoulis et al.’s study [14], in non-diabetic Greek adults, hypertensive men and women have developed glycaemic profiles according to diabetes mellitus and obesity history in their families. Family history of diabetes, obesity markers, the criteria of glycaemia, insulin protection and IGH penetration have been resolved in an important collaborator of its organization.

3 Research Methods This research, proposes an architecture for intelligent diagnosing diabetes types and its complications based on the physical symptoms. The methodology contains two research phases: The first phase of designing the architecture for intelligent diagnosing diabetes types and its complications using back-propagation neural networks, fuzzy logics and K-Means clustering and the second phase of developing the experimental prototype system using Visual Studio 2017 and SQL Server 2016 with Windows 10 operating system in order to test and validate the proposed architecture. Back propagation neural network is a mathematical model widely used for classification and diagnosis in various thirst areas like effective decision making in medical fields, signal processing and so on. Fuzzy logic expert systems used in medical examinations are of great importance providing an exact evaluation report of medical data provided to the systems. These types of the system provide an instant and simple method of medical examination. Cluster analysis has been applied for such varied objectives as finding a true topology, model fitting, prediction based on groups, hypothesis generation, hypothesis testing,

An Architecture for Intelligent Diagnosing Diabetic Types

105

data exploration, data reduction and grouping similar entities into homogeneous classes. The corresponding results are presented in a series of different studies. This method is intended to further verify the validity of the proposed diagnostic system in order to classify diabetes and its complications. The experimental prototype has been adopted, since Karen has demonstrated only partial functions on some aspect of the system [15]. The study’s overall design is shown in Fig. 1.

Type-I Diabetes

Data Clustering

Type-II GestaƟonal

Neural Networks Cardiovascular Feature SelecƟon Data Preprocess

Clinical Database

Nephropathy Fuzzy Logics

Neuropathy ReƟnopathy Gangrene

Result

Fig. 1. An architecture for diagnosing diabetes types and complications

3.1 Back-Propagation Neural Networks In the recent year, several researchers have proposed that a back-propagation neural networks [16] is an effective method for data prevision in medical science. Numerous studies on training data, the computational design and the creation of real-time applications, such as robotic control have now been provided to the neural network of the national and international research groups every day. Mcculloch-Pitts Neuron on developed by McCulloch and Pitts [17] is the first representation of neural networks in the mammalian brain. Backpropagation neural network key characters are: (a) black box build, (b) each link-based node, (c) a single network structure is capable of performing a variety of tasks via data training e no data noise and (f) no interference [17]. A single neuron is n with inputs x 1 , x 2 ……. x n with real and y output values. Moreover, the inputs of the neuron are connected to actual numbers, w1 , w2 ….. wn weights. The output depends on the number of inputs weighted. n  i=1

Wi Xi

106

G. Thangarasu et al.

The non-linear function is known as activation or threshold function. Heaviside function for all aerials R, and Sβ functions described by a formula are the most common activation features. Sββ(a = (1 + e−βa )−1

(1)

When β is a positive constant (called the steepness parameters), the value of which specifies a specific sigmoid function, considering the neuron output to be specified as,   n  y = sβ Wi Xi − θ (2) i=1

βεR+

For some bias, the β is the so-called neuronal bias. The partition determines the weighted number of inputs in which the neuron output changes the most sensitive in the volume. For the convenience of the consumer, the bias θ and the corresponding weight W 0 = θ respectively, are normally expressed by an extra input. So the Eq. (2), as in Eq. (3) will be replaced by a simple formula.  n   y = sβ Wi Xi (3) i=1

3.2 Data Clustering Algorithms The data cluster is a popular technology for data mining. Relevant domains were applied successfully to identify trends. Data clustering algorithms are generally quick and simple. (i) Hierarchical clusters, (ii) clusters of means k, (iii) clusters of medium K and (iv) clustering of fuzzy C-means (FCM) are the four common methods used for clustering results. The fuzzy C-means the clustering methods are used to identify diabetes forms from clinical databases since data points can be assigned to more than one cluster according to this approach. It is an addition of algorithms of K-means. Each data point can only be allocated a single cluster for the K-means clustering. The clustering is one of the initial data analysis components. It stipulates that the transfer of data to different classes is important from measurements [18]. 3.3 Fuzzy Logic Techniques Based on the professional knowledge of the medical profession and clinical evaluation, the furious reasoning relationship between symptoms and risk factors for diabetes is established to classify multiple complications caused by diabetes [19]. Incomprehensible systems are controlled by the importance of the fudging logic techniques. In order to concentrate on effective decision-making, the fluffy reasoning strategies do not require extracting quantifications. More etymologic variables are used for the use of fuzzy logic. It makes the production and operation of systems quicker and easier. It can be a good data management decision-making tool. Fuzzy logic will require many repetitions in order to discover a variety of guidelines that provide a consistent solution in complicated

An Architecture for Intelligent Diagnosing Diabetic Types

107

systems. During the fusion of fuzzy logic techniques with the neural network, the study of data clusters will cut time to establish rules [20]. The fuzzy logic is one of the most common and widely applied artificial intelligence techniques, including medical diagnostics, assessment, image care, control systems and model recognition. Figure 2 shows the overall functions of fuzzy inference systems.

Fuzzy rule

Fuzzy rule

Fuzzy rule

Decision Making Unit (Inference Engine)

Fuzzification

Defuzzification

Output

Fig. 2. Fuzzy inference system

The first step in fuzzy logic is to take the measured data and determine the membership degree of these inputs to associated fuzzy sets. It is done by giving the value of each variable to a membership function set. Membership functions take different shapes. The two most common functions are triangular and trapexoidal.

4 Result Analysis and Discussion First of all, the questionnaire has been designed with associated 30 questions for collecting data for testing and proposed system. The questionnaire has been verified and approved by three Professional Diabetic Experts. The questionnaire was distributed to 235 individuals. Out of them, 200 respondent’s data were confirmed for experimental purposes. Table 1 shows details of the diagnosis of diabetes from 200 respondents. The respondents are listed as 98 men and 102 women. The results of the prototype device diagnosis are 62 respondents (27 males and 35 female interviewees) and 138 non-diabetic interviewees (71 males and 67 females). Table 2, shows the detailed prototype findings diagnosed for 200 respondents, including diabetic non-diabetes respondents put age group wise. Table 1. Summary of diabetes diagnosing results Respondents Prototype diagnosed results Diabetes Yes

Total Diabetes No

Male

27

71

098

Female

35

67

102

Total

62

138

200

108

G. Thangarasu et al.

The findings of the prototype diabetic device respondents are discussed in detail in Table 2. A total of 62 diabetes respondents was found. Of the 58 people with diabetes, one was affected by type-1, fifty-seven were affected by Type-II and 4 by gestational diabetes. Long term complications can grow gradually over a decide for people with type-1 diabetes. Patients should then adopt the daily medication and diet that will help to decrease the risk of complications. Table 2 shows how the disease works and how to deal with emotional difficulties and to make improvements in the required lifestyle. Table 2. Respondent’s diabetes diagnosing results Diabetes result

Age category

Nos.

Types

Nos.

Complications

Nos.

Yes

Below 20 years

0

Type-I

1

Cardiovascular

0

21–30 years

1

31–40 years

2

41–50 years

6

51–60 years

Type-II

25

Above 60 years

28 Gestational

62

57

62

4

Retinopathy

1

Neuropathy

0

Nephropathy

0

Gangrene Foot

0

Cardiovascular

23

Retinopathy

37

Neuropathy

13

Nephropathy

16

Gangrene Foot

15

Not Applicable

62

The author used predictive analysis a table confusion is a table with two rows and two columns that reports the number of false positive, false negatives, true positive and true negative. This allows more detailed analysis than mere proportion of correct classification. Table 3 shows the diabetes diagnosing types and its complications based on the proposed results. The Table 3 conduced that 94.50% accuracy of the models. Table 3. Table of confusion TN FN TP

FP Accuracy

200 11 189 0

94.5

5 Conclusion In this research, the design of intelligent diagnostics for diabetes, the development of experimental technologies and the improvement solution for the health sector in particular for the diagnosis of diabetes, types and its symptomatic complications are involved.

An Architecture for Intelligent Diagnosing Diabetic Types

109

The synthesis of neural network approaches with the back-propagation algorithm, clustering algorithm, and fluffy architectural logic techniques. The individual capabilities of each technology are unique. The neural network technology has been described in recent years, in terms of the literature review, as a significantly powerful and incredibly accurate decision-making method. As a second technology for the classification of diabetes types, the clustering algorithm is employed. The final soft logic approach is used to classify different diabetes complications. It can be a good tool for decision making, easier, simpler and versatile with high precision. The best technological concept will be the three best innovations integrated into a single system. The findings were reported with a precision of 94.50%. This proposed system will help the people to diagnose their diabetes disease types and its complications in the very early stage and get medication on time to live a longer life.

References 1. Usama, M., Ahmad, B., Xiao, W., Hossain, S., Muhammad, G.: Self-attention based recurrent convolutional neural networks for disease prediction using healthcare data. Comput. Methods Programs Biomed. 190, 105–122 (2020) 2. Liu, L, Wang, l., Huang, Q., Zhou, L., Fu, X., Liu, L.: An efficient architecture for medical high-resolution images transmission in mobile telemedicine system. Comput. Methods Programs Biomed. 187, 88–101 (2020). 3. Sandhu, K.J., Verma, A., Rana, P.: An Expert Approach for data Flow Prediction: Case Study of Wireless Sensor Networks 112(325–352), 73–91 (2020) 4. Kamdar, J.H., Jeba Praba, J., John, J.: Artificial intelligence in medical diagnosis: methods, algorithms and applications. Learning and Analytics in Intelligent Systems book series LAIS 13, 27–37 (2020) 5. Uzoka, F.M.E., Osuji, J., Obot, O.: Clinical decision support system (DSS) in the diagnosis of malaria: a case comparison of two soft computing methodologies. Expert Syst. Appl. 38(1), 1537–1553 (2018) 6. IDF Diabetes Atlas, International Diabetes Federation: 9th Ed. (2019) 7. Medical News Today. https://www.medicalnewstoday.com/info/diabetes. Accessed 10 Feb 2020 8. Michael, B.: Inadequacies of current approaches to pre-diabetes and diabetes prevention. J. Endocrine 44(3), 623–633 (2018) 9. Rahaman, S.: Diabetes diagnosis decision support system based on symptoms, signs and risk factors using special computation algorithm by rule base. In: 15th International Conference on Computer and Information Technology, pp. 65–71, Chittagong (2016) 10. Matsumoto, T., Shimada, Y., Kawaji, S.: Clinical diagnosis support system based on symptoms and remarks by neural networks. In: IEEE Conference on Cybernetics and Intelligent Systems, pp. 1304–1307, Singapore (2018) 11. Gultepe, E., Hien, N., Albertson, T., Tagkopoulos, I.: A bayesian network for early diagnosis of sepsis patients: a basis for a clinical decision support system. In: IEEE 2nd International Conference on Computational Advances in Bio and Medical Sciences, pp. 1–5, Las Vegas (2016) 12. Dakua, S.P., Navkar, N.V., Abi-Nahed, J., Groen, D., Bernabeu, M.O., Saghir, M.A.R., Kamel, H., Al-Ansari, A., Coveney, P.V.: Towards a computational system to support clinical treatment decisions for diagnosed cerebral aneurysms. In: Middle East Conference on Biomedical Engineering, pp. 281–284, Doha (2018)

110

G. Thangarasu et al.

13. Reiter, H., Naujokat, E., Pinter, R., Devot, S.: Take Care: a home-based sensor system for the management of cardiovascular risk factors primary prevention by monitoring vital body signs, analysing the data and closing the loop by feedback, coaching and motivation. In: 5th International Summer school and Symposium on Medical Devices and Biosensors, Hong Kong, pp. 186–189 (2018) 14. Vyssoulis, G.P., Liakos, C.I., Karpanou, E.A., Triantafyllou, A.I., Michaelides, A.P., Tzamou, V.E., Markou, M.I., Stefanadis, C.I.: Impaired glucose homeostasis in non-diabetic greek hypertensives with diabetes family history, effect of the obesity status. J. Am. Soc. Hypertens. 7(4), 294–304 (2018) 15. Inke, M., Guy, C.: The role of institutional design and organizational practice for health financing performance and universal coverage. Healthy Policy 99(3), 183–192 (2016) 16. Carrin, G., Mathauer, I., Xu, K.: Universal coverage of Health Services: tailoring its Implementation. Bull. World Health Organ. 86(1), 09–24 (2018) 17. Sherrod, P.H.: DTREG Predictive Modeling Software Manual (2019) 18. Nathan, D.M.: Advances in diagnosis and treatment. Int. J. Med. 314(10), 1052–1062 (2019) 19. Suyash, S., Lokesh, S., Vijeta, S., Ajai, K., Hemant, D.: Prediction of diabetes using artificial neural network approach. Eng. Vibration, Comm. Inf. Process. 478(1), 679–687 (2018) 20. Mohamed, S., Baskar, S., Sarma, V.R., Mustafa, M.J.: Cloud-based framework for diagnosis of diabetes mellitus using K-means clustering. Health Inf. Sci. Syst. 16(6), 321–232 (2018)

An Advanced Encryption Cryptographically-Based Securing Applicative Protocols MQTT and CoAP to Optimize Medical-IOT Supervising Platforms Sanaa El Aidi1(B) , Abderrahim Bajit1 , Anass Barodi1 , Habiba Chaoui1 , and Ahmed Tamtaoui2 1 Laboratory of Advanced Systems Engineering (ISA), National School of Applied Sciences,

Ibn Tofail University, Kenitra, Morocco {barodi.anass,habiba.chaoui}@uit.ac.ma 2 National Institute of Posts and Telecommunications (INPT-Rabat), SC Department, Mohammed V University, Rabat, Morocco [email protected]

Abstract. Our proposed Platform is to detect and to measure the temperature of persons with PIR Node IOT, and then verify his identity through the combination of an RFID Node IOT and facial recognition with Cam Node IOT and if these tests are valid, the persons can then access the public area. With the security layer of the CoAP (Constrained Application Protocol) and MQTT (Message Queuing Telemetry Transport) communication protocols to compare these 2 protocols in terms of the execution time, the RAM memory space occupation, and the execution CPU consumptions. Then we have able to create an intelligent and secure medical IoT Platform this has been designed to monitor citizens to access this vast area in a more organized and secure manner to reduce the severity of this pandemic. Keywords: MQTT · MQTT Client IOT · CoAP Client IOT · Broker · CoAP SERVER IOT · AES encryption · CoAP · IoT · Artificial intelligence · Microcontroller · OpenCV

1 Introduction Given the spread of the coronavirus pandemic, we thought to create an intelligent and secure medical IoT Platform and make it work in an existing environment (such as in a hospital, company, establishment….) without the complexity of integration with the network or other existing infrastructures. Our objective of this platform is to improve and optimize health precautions to have no integration between citizens who have covid-19 or who had contact with an infected person with citizens who are never infected with covid-19. The Internet of Things allows an interaction between the physical and digital worlds. The digital world interacts with the physical world through sensors and actuators. These © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 111–121, 2021. https://doi.org/10.1007/978-3-030-70713-2_12

112

S. El Aidi et al.

sensors collect information that must be stored and processed. Data processing can take place at the edge of the network or on a remote server or in the cloud. The storage and processing capacities of an IoT object are limited by the available resources, which are restricted due to limitations in size, energy, power, and computing capacity [1]. In our platform, we have also used artificial intelligence, specifically facial recognition for image detection and processing, the human face provides several social signals essential for our public life [2]. The face mediates person identification, attractiveness, and facial communicative. A.I. is a science that dates back some thirty years. Its purpose is to reconstruct intelligent reasoning and action using artificial means - almost always computers. The difficulties are a priori of two types: • Most of our activities we don’t know ourselves how we do it. We don’t know a precise method - no algorithms today say computer scientists - to understand a written text or to recognize a face, to demonstrate a theorem, to establish a plan of action, to solve a problem, to learn… • Computers are a priori very far from such a level of competence. They have to be programmed from the very beginning. Indeed, programming languages only allow to express very ’elementary’ notions. A.I. is, from this double point of view, an experimental science: experiments on computers that allow to test and refine the models expressed in the programs on many examples; observations on humans (generally the researcher himself) to discover these models and better understand the functioning of human intelligence [3, 4]. IoT needs artificial intelligence AI and vice versa, and the use of AI is beneficial for real-time processing [5, 6]. In our platform IoT has been used to connect medical devices to the internet, to collect several data about the citizen to process it’s analyze and to act appropriately the citizen’s access right. Our Platform is used to use 4 tests for monitoring citizens to access the surface of the body in a more organized and secure manner, the first we did is to check the temperature of citizens using the PIR Node IOT and if the test is less than or equal to 37 °C we move to the 2nd test for the detection of the identity of citizens by the RFID tag, and we verify if the person has already negative test of PCR/Serological, then the person presented need to confirm his information by using artificial intelligence recognition system. The system is capable to analyze the facial structure by comparing it to the information in the database and then identifying the detected person. We devoted ourselves to the implementation of a MQTT/CoAP protocols with an authentication server in the IOT. The intervention of encryption algorithms to solve the above-mentioned problems by encrypting/decrypting messages. We have therefore implemented a secure version of the CoAP and MQTT protocols using the AES encryption algorithm [7].

2 General Architecture Our intelligent and secure medical IOT Platform is used to identify citizen to have access at the public area public area, so we have 4 tests to do to appropriate and enable citizens. we used 3 nodes, the first one -PIR and temperature Client Node- is applied to detect the

An Advanced Encryption Cryptographically-Based Securing

113

presence of a citizen and samples its temperature, the second one -RFID Identification IOT Client Node- is deployed to identify the citizen and sanitary information, and the third one –Image Recognition Camera IOT Client Node– is implemented to recognize the citizen face and his identity. In this proposed medical IOT platform, we employed 2 application IOT communication protocols, the Message Queuing Telemetry Transport -MQTT- and the Constrained Application Protocol –CoAP- in order to choose the best one in terms of the executing time, the RAM memory space occupation and the execution CPU consumptions (Fig. 1).

Fig. 1. Intelligent and secure medical IOT Platform

The illustrated figure presents the medical IOT Platform which is used to identify cases of coronavirus infection and to control access to a public place using an RFID card by performing 4 tests the 1st is used to test if the citizen’s temperature is below 37°C. Then we verify if the citizen has already presented a PCR/serological test thanks to the data connected to the RFID Identification tag, then we verify that a citizen has already been in contact with a positive case and if we find that the citizen’s temperature is normal and RFID information and identification has shown the condition the citizen has negative test and not contact with an infected person, then we move to the citizen identity by using the facial recognition based on artificial intelligence, if this test is verified valid we give the authorized access, and if detected the non-conformity of the identity with the data associated to the RFID tag we give the access denied in this case, even if the person has a normal temperature, has been tested negatively and has not had contact with an infected person. For the MQTT protocol implementation, we employed 3 topics: 3 publishers - IOT MQTT Client Nodes (PIR, Temperature Sensor, RFID identifier and CAM recognizer)-,

114

S. El Aidi et al.

and one IOT MQTT Client subscriber -Web Server IOT Client Node-, a platform key is also set to encrypt and decrypt transmitted data. And for the CoAP implementation, we used 3 topics: IOT CoAP Server, IOT CoAP Client, we also used AESLIB and PyCryptodome for encrypting and decrypting transmission DATA.

3 Methodology 3.1 Protocols Application IOT CoAP is an application, IOT communication and web transfer protocol based on Representational State Transfer (REST) that is used for resource-constrained devices operating in an IP network, resource-constrained devices can be numerous, but they are often linked to each other by function or location, group communication mechanisms can improve the efficiency and latency of communications and reduce the bandwidth for a given request [8]. CoAP is primarily designed for constrained devices. Clients may send GET, PUT, POST and DELETE resource requests to the server. CoAP messages are encoded in a simple binary format. Packets are simple to generate and can be parsed in place without consuming energy in constrained devices [9]. The MQTT uses a publish/subscribe model, has low network overhead and can be implemented on low-power devices such as IOT node microcontrollers that could be used in remote sensors in the Internet of Things. As such, Mosquitto is destined to be employed in all cases where there is a need for light messaging, especially on constrained devices with limited resources [10]. The primary difference between CoAP and MQTT is that the former works over the user’s datagram (UDP), while the latter works in addition to TCP. Since UDP is inherently unreliable, CoAP provides its own reliability mechanism, so it has two modes: reliable and unreliable. In reliable mode it is the use of confirmable messages that require an ACK, while in non-reliable mode it is the use of non-confirmable messages that do not require recognition. Another difference between CoAP and MQTT is the availability of different QoS levels. The MQTT defines 3 levels of QoS while the CoAP does not offer a differentiated quality of service [11]. AES is a symmetric key system in which the sender and recipient of a message share a unique common key, which is used to encrypt and decrypt the message. AES supports key sizes of 128, 192, and 256 bits, and consists of 10, 12, and 14 encryption repetition (also known as rounds), respectively. Each round mixes the data with a round-key derived from encryption key. Except last round, each round comprises four processing steps, including SubBytes, ShiftRows, MixColumns, and AddRoundKey [12]. • SubBytes is an invertible and nonlinear transformation, which adopts 16 identical 256byte substitution tables (i.e., S-box) for individually mapping bytes of the data block into other bytes. S-box entries are produced by calculating multiplicative inverses in Galois Field GF(28) and applying an affine transformation. • ShiftRows performs a byte transposition by cyclically shifting rows of the data block according to predefined offsets, i.e., left shift of the second, third, and fourth row by one, two, and three bytes, respectively.

An Advanced Encryption Cryptographically-Based Securing

115

• MixColumns multiplies each column of the data block with a modular polynomial in GF(28). Instead of computing separately, SubBytes and MixColumns can also be combined into large Look-Up-Tables (LUT). • AddRoundKey transformation adds the data block with round-key derived from initial secret key in the key schedule unit. This function XORed each byte of the block with the corresponding bye in the round-key [12]. The operations in decryption are basically the inverse of the operations in encryption. Besides, the number of rounds of the looping is set to Nr-1 in which Nr is specified according to the AES specification [13].

4 Related Works 4.1 Comparative Study for the Proposed Protocols IOT In an CoAP environment, a solution that consists in integrating DTLS and CoAP protocols for IoT [14] through CoAP-DTLS integration has been developed to allow the application to automatically access CoAP. The results of the evaluation show a significant gain in terms of power consumption, network response time and pro-cessing time. Research is currently oriented towards securing the IoT, several aspects by highlighting the security and proposing several solutions by establishing the specif-ic characteristics of the protocols of the application layer proposed by the RSA-based security solution [15]. The most used asymmetric cryptography, being to achieve low overload and high interoperability, because the overload of the DTLS handshake process that consumes a large amount of power not supported by IoT devices. An-other analysis of the two known security protocols that can be used to secure CoAP networks: DTLS and Internet Protocol Security (IP-sec) [16]. They concluded that these protocols are not the most optimized solutions for CoAP security by citing the drawbacks of these security protocols. In an MQTT environment, the demand for a new approach to secure the MQTTbased platform in order to guarantee the confidentiality and integrity of transmitted data [17]. According to the name “secure-MQTT” which is standardized by IANA and port 8883 is exclusively reserved for MQTT over TLS [18], security between the MQTT broker and users can be provided by SSL and TLS [19], but the TLS protocol is not cost-effective for optimal security at MQTT. While the additional use of the CPU is generally negligible for the broker, it can be a problem for highly constrained devices that are not designed for computationally intensive tasks [18]. The CA-based solution to generate a private key and a certificate, which will be published manually for certified customers [19]. This approach is not applied to an IoT environment that may contain a wide range of nodes, so manual configuration is so difficult to achieve, and security has a cost in terms of CPU usage and communication costs.

5 The Proposed Approach Encryption is the process of converting the original plain text into non-readable format. There are various encryption techniques that exist in cryptography such as DES, Triple

116

S. El Aidi et al.

DES, AES, RSA, etc. AES has been widely used in many devices, especially in resourceconstrained environments, due to their efficient, secure, and high-performance use in these resource-limited environments. And the major point of the AES algorithm is that the AES key has the smallest dimensions which are much less important than the others. Reducing the size of the key will reduce computing resources, Conserve more energy from all the nodes IOT and extend the life of the network. Symmetric Encryption uses the same key concept to encrypt as well as decrypt. There are a number of benefits to this approach. The performance is relatively high. There are two aspects of this algorithm. The first is the encryption algorithm and the other is the key. The encryption algorithm is a process of transformations that take place on the plain text with the key itself. At the time of decryption, the same process of encryption is followed in a reverse manner with the same key. A strong algorithm should depend on its key entirely [20]. Our goal is to apply an encryption layer based on AES. To achieve this, a security layer is added for both secured protocols. In this proposal, we will ensure that only authorized citizens can access the information, and we will provide a more secure text to those who do not have permission or right to access the data. According to Fig. 2, there is a security layer added to the CoAP protocol, and the data is encrypted from end to end. Encryption Data is done at the CoAP Client IOT level, and the Decryption at the webserver IOT level, we can conclude that only authorized persons have the permission to access the data. We have proposed the algorithm design described as follow:

Fig. 2. Algorithm transmission data

We have used in our platform IOT-Medical precisely in the CoAP protocol: • Coapthon Server [21]: CoAP Server implementation in python. • ESP Nodes: CoAP Server and Client implementation in MicroController WiFi Module. • AESLIB: AES implementation. • PyCryptodome: AES implementation in python

An Advanced Encryption Cryptographically-Based Securing

117

According to Fig. 3, there is a security layer added to the MQTT protocol, the diagram has 3 principal elements: the MQTT broker and 2 MQTT customers: a subscriber and a publisher. The execution environment for a subscriber can be a microcontroller node, for example, an ESP8266 card and ESP32CAM, so it includes physical sensors and a subcomponent which is the MicroPython, the latter contains first of all two parts: the “AESCipher Encrypt” and the encryption key. The “AESCipher Encrypt” is the piece of code responsible for encrypting the collected data into a ciphertext which is then sent to the broker. In our approach, the subscriber is a Linux web server whose role is to perform several tasks such as managing the graphical user interface, processing decrypted data via “AESCipher Decrypt”, transmitting encrypted orders to other platform nodes via the MQTT broker, reading and/or writing to the database (Fig. 4).

Fig. 3. Diagram of deployment

Our project has been evaluated in a pure IoT environment, to view the exchange of messages in an encrypted format via an MQTT broker between all the microcontroller nodes of the platform (Fig. 5).

118

S. El Aidi et al.

Fig. 4. Encryption with AES

Fig. 5. Decryption with AES

6 Discussion The objective of this work is to apply the AES algorithm in our platform for both secured and unsecured protocols because it allows to use small keys for encryption and decryption. Table 1 shows a comparative results of the two versions of the platform: secured and unsecured for the 2 protocols in terms of executing time, occupation RAM space and execution CPU consumptions, for PIR and temperature node (node1), RFID Identification IOT Node (node2), and Image Recognition Camera IOT (node 3), according to the Table 1, we find that the secure MQTT has a very high consumption by secure CoAP. And in the table, we have shown the 2 protocols in unsecure mode, also after the addition of the security layer on the 2 protocols IOT in order to choose the best secure protocol and on the other hand to show that the security layer did not influence our Platform IOT in terms of time and power consumption.

An Advanced Encryption Cryptographically-Based Securing

119

According to the analyses made in [22], it was found that encryption and decryption with ECC (Elliptical Curve Cryptography) is better than RSA, and in this article it was concluded that encryption and decryption with AES is better than RSA, because the principle of AES is to ensure effective communication between nodes in IOT, ensuring confidentiality, integrity and authentication exceptional and represent the best options for resource-constrained environments. In future work we will apply encryption with ECC on our Medical-IOT platform with both secure protocols to make a comparative study between AES and ECC. Table 1. The results of the implementation of IOT protocols Nodes

Protocols

TIME (ms)

CPU (µs)

Node1

MQTT

2320

534323.67

CoAP

3020

431575.3

584

Node2

MQTT

2165

15259729.7

328

CoAP

2785

5343170.7

504

MQTT

1606.29

520469

335872

CoAP

1510

435351

50217

Node1

S-MQTT

1390

404615.5

19820

S-CoAP

2163

326815.7

Node2

S-MQTT

1593.33

6043050.5

18620

S-CoAP

1876

4597250.5

376

S-MQTT

1270

508456

S-CoAP

50217

1220

Node3

Node3

RAM (bytes) 404

502

68000 402267

7 Conclusion and Perspectives In this given work we have able to create an intelligent and secure medical IOT Platform and making it more efficient and secure by using our proposed protocols with AES algorithm for encryption and decryption transmission DATA. Our proposed approach has been tested in a real-time environment, to illustrate the exchange of DATA in an encrypted format via an IOT Server protocol between all the nodes IOT of the platform. In future work, we will implement ECC encryption on our platform to compare AES with ECC and choose the best encryption algorithm on our proposed protocol.

References 1. Garg, H., Dave, M.: Securing IoT devices and securelyconnecting the dots using REST API and middleware 978–1–7281–1253–4/19/$31.00 © 2019 IEEE https://doi.org/https://doi.org/ 10.1109/IoT-SIU.2019.8777334

120

S. El Aidi et al.

2. Pantic, M., Patras, I.: Dynamics of facial expression: recognition of facial actions and their temporal segments from face profile image sequences. IEEE Trans. Syst. Man Cybern. https:// doi.org/10.1109/TSMCB.2005.859075 3. Laurière, J.-L.: Intelligence artificielle : résolution de problèmes par l’homme et la machine. https://ulysse.univ-lorraine.fr/discovery/fulldiplay?vid=33UDL_INST:UDL&docid=alm a991003670679705596&lang=fr&context=L&adaptor=Local%20Search%20Engine 4. Barodi, A., Bajit, A., Benbrahim, M., Tamtaoui, A.: Improving the transfer learning performances in the classification of the automotive traffic roads signs. In: E3S Web Conf., -Proceeding (2020) 5. Zhou, J., Wang, Y., Ota, K., Dong, M.: AAIoT: accelerating artificial intelligence in IoT systems. IEEE Wirel. Commun. Lett. 8(3), 825–828 (2019) 6. Barodi, A., Bajit, A., El aidi, S., Benbrahim, M., Tamtaoui, A.: Applying real-time object shapes detection to automotive traffic roads signs. In: Proceeding of the 2020 International Symposium on Advanced Electrical and Communication Technologies (ISAECT), Morocco, Kenitra, 2020, pp. 1–6 (2020) 7. Bajit, K.A., Nahid, M., Tamtaoui, A., Benbrahim, M.: A Psychovisual Optimization of wavelet foveation-based image coding and quality assessment based on human quality criterions. Adv. Sci. Technol. eng. Syst. J. 5(2), 225–234 (2020). https://doi.org/10.25046/aj050229 8. Rahman, A., Dijk, E.: Group Communication for the Constrained Application Protocol (CoAP) (2014). https://www.hjp.at/doc/rfc/rfc7390.html 9. Kayal, P., Perros, H.: A comparison of IoT application layer protocols through a smart parking implementation. 2017 20th Conference on Innovations in Clouds, Internet and Networks (ICIN), Paris, pp. 331–336 (2017). https://doi.org/10.1109/ICIN.2017.7899436. 10. Light: Mosquitto: server and client implementation of the MQTT protocol. J. Open Source Softw. 2(13), 265 (2017). https://doi.org/10.21105/joss.00265 https://doi.org/10.1109/ACC ESS.2018.2852563 11. Thangavel, D., Ma, X., Valera, A., Tan, H., Tan, C.K.: Performance evaluation of MQTT and CoAP via a common middleware. In: 2014 IEEE Ninth International Conference on Intelligent Sensors, Sensor Networks and Information Processing (ISSNIP), Singapore, pp. 1–6 (2014). https://doi.org/10.1109/ISSNIP.2014.6827678. 12. Tsai, K., Huang, Y., Leu, F., You, I., Huang, Y., Tsai, C.: AES-128 based secure low power communication for LoRaWAN IoT environments. IEEE Access 6, 45325–45334 (2018). https://doi.org/10.1109/ACCESS.2018.2852563 13. Lu, C.-C., Tseng, S.-Y.: Integrated design of AES (Advanced Encryption Standard) encrypter and decrypter. In: Proceedings IEEE International Conference on Application- Specific Systems, Architectures, and Processors, San Jose, CA, USA, 2002, pp. 277–285. https://doi.org/ 10.1109/ASAP.2002.1030726. 14. Raza, S., Shafagh, H., Hewage, K., et al.: Lithe: lightweight secure CoAP for the internet of things. IEEE Sensors J. 13(10), 3711–3720 (2013) 15. Kothmayr, T.: A security architecture for wireless sensor networks based on DTLS. Master’s thesis in the Software Engineering Elite Graduate Program at the University of Augsburg (2011) 16. Alghamdi, T.A., Lasebae, A., Aiash, M.: Security analysis of the constrained application protocol in the Internet of Things. In: Second International Conference on Future Generation Communication Technologies (FGCT 2013). IEEE, pp. 163–168 (2013) 17. Silva, C., Toasa, R., Martinez, H.D., Veloz, J., Gallardo, C.: Secure Push ‘notification service based on MQTT Protocol for mobile platforms’, Conference: XII Jornadas Iberoamericanas de Ingenieria de Software e Ingenieria del Conocimiento 2017, JIISIC 2017 - Held Jointly with the Ecuadorian Conference on Software Engineering, CEIS 2017 and the Conference on Software Engineering Applied to Control and Automation Systems, ISASCA 2017At: Latacunga, Ecuador

An Advanced Encryption Cryptographically-Based Securing

121

18. Mektoubi, A., Hassani, H.L., Belhadaoui, H., Rifi, M., Zakari, A.: New approach for securing communication over MQTT protocol A comparaison between RSA and Elliptic Curve. In: 2016 Third International Conference on Systems of Collaboration (SysCo) (2016). https:// doi.org/10.1109/sysco.2016.7831326 19. Khamphroo, M., Kwankeo, N., Kaemarungsi, K., Fukawa, K.: MicroPython-based educational mobile robot for computer coding learning. In: 2017 8th International Conference of Information and Communication Technology for Embedded Systems (IC-ICTES) (2017). https://doi.org/10.1109/ictemsys.2017.7958781 20. Mishra, P., Agrawal, M.: A Comparative Survey on Symmetric Key Encryption Techniques, Monika Agrawal et al. / International Journal on Computer Science and Engineering (IJCSE) ISSN: 0975–3397 Vol. 4 No. 05 May 2012 877. ’https://citeseerx.ist.psu.edu/viewdoc/dow nload?doi=10.1.1.433.2037&rep=rep1&type=pdf 21. Tanganelli, G., Vallati, C., Mingozzi, E.: CoAPthon: easy development of CoAP-based IoT applications with Python. In: 2015 IEEE 2nd World Forum on Internet of Things (WF-IoT). IEEE, 2015, pp. 63–68 (2015) 22. EL aidi, S., Bajit, A., Barodi, A., Chaoui, H.: An elliptic-curve based cryptographically optimized vehicular protocols applied to secured applicative protocols MQTT and CoAP. In: 2020 International Conference on Electronics, Control, Optimization and Computer Science (ICECOCS), Kenitra, 2020, pp. 1–6, Proceeding (2020)

Pulmonary Nodule Classification Based on Three Convolutional Neural Networks Models Enoumayri Elhoussaine and Belaqziz Salwa(B) LabSIV Laboratory, Department of Computer Science, Faculty of Science, Ibn Zohr University, BP 8106, 80000 Agadir, Morocco [email protected]

Abstract. The leading reason of death linked to cancer worldwide is lung cancer. To plan effective treatment, create monetary and care plans, early diagnosing of lung nodules in computed tomography (CT) chest scans must be performed. In this context, the purpose of this paper is to take into account the problem of classification between malignant and benign pulmonary nodules in CT scans, which aims to automatically map 3D nodules to category labels. Thus, we propose an ensemble learning approach based on three Convolutional Neural Networks including a basic 3D CNN, a 3D model inspired by AlexNet, and another 3D mod-el inspired by ResNet. The result from these CNNs is combined to estimate one result, using a fully-connected layer with a softmax activation. These CNNs are trained and evaluated on the LIDC-IDRI public dataset. The best result is obtained by the ensemble model, providing a larger AUC (84.66%); “area under the receiver operating characteristic curve” and 94.44% for TPR (sensitivity), with a data augmentation technique. Keywords: Pulmonary nodule classification · LIDC-IDRI · Deep neural networks · 3D AlexNet · 3D ResNet

1 Introduction Lung cancer is the pathology that has more mortality globally, accounting for more deaths than cancers of the prostate, breast, colon, and pancreas combined [1], and its mortality rate can be reduced utilizing Low-Dose Lung CT screening [2]. However, the subtle differences between benign and malignant pulmonary nodules make lung cancer diagnosis a difficult task even for human experts. Moreover, the evaluation of radiologic diagnosis is very subjective, it induces much variety in the radiologist’s opinions. Computer-aided diagnosis (CAD) provides an objective prediction and a non-invasive solution for the problem of classification between malignant and benign pulmonary nodules in CT scans, CAD can be used to increase the radiologist’s confidence in the diagnosis of a pulmonary nodule. For current CAD systems, there are two categories: the first one measure radiological traits (e.g. shape, nodule size, texture, location), in this approach, the feature selection is © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 122–128, 2021. https://doi.org/10.1007/978-3-030-70713-2_13

Pulmonary Nodule Classification

123

extracted by hand or with predefined filters [3, 4], then a classifier is adapted to determinate the malignancy status. The second category is based on the automatic extraction of features via deep neural networks [5–7]. To handle the pulmonary nodule classification problem, these networks learn from the data through a general learning process without needing to extract features as the traditional way. In [8], the author proposes four Convolutionary Neural Networks evaluated on the public LIDC-IDRI dataset, the best classification performance (0.9010 for AUC and 86.84% for accuracy) was achieved using 3D multi-output DenseNet. In [9] the author uses a boosting classifier to obtain the best result based on a voting idea with several CNNs, and the classification method is carried out by modifying the weight of the training sample, the classifiers are combined linearly to increase their performance. The model suggested in [10] uses a Gradient Boosting Machine (GBM). First, using a convolutional layer, the features are extracted, then a set of 3D dual-path blocks is employed to learn higher-level features. Finally, for malignant or benign classification the author applies a 3D average pooling and binary logistic regression. We propose in this paper a new classification approach composed of an ensemble classifier using three 3D models: one with a set of convolution and fully-connected layers, the others are inspired by AlexNet [11] and ResNet [12] architectures respectively. The proposed ensemble classifier uses the outputs from the three 3D models, and then a fully-connected-layer with softmax activation is trained to get the results. The public LIDC-IDRI dataset is used to train and evaluate the proposed approach. This paper is structured as follows. First, in Sect. 2, the details of the proposed method for pulmonary nodule classification are given. Then Sect. 3 dives into the experiments and their results, this section is split into three major subsections, the Dataset is introduced in the first one, the second and the third gives the details about experiment settings and results respectively. Finally, in Sect. 4 a conclusion is held to summary the proposed method, and discuss the perspectives.

2 Proposed Method In this section, we describe our models for classifying lung nodules in CT scans using deep neural networks. In practice, radiologists check several slices of a lung nodule and consider the 3D information of the nodule to make a diagnosis. Most of the previous approaches do not include full 3D information for a pulmonary nodule, they simply use single or multi-view 2D images. Therefore, the proposed method discriminates malignant lung nodules from benign ones using as input a 3D CT chest scan with the location of the nodules. A typical CT scan consists of hundreds of 2D gray images with a dimension of 512 × 512. The design of the proposed networks is as below: Basic 3D CNN: Consists of four convolutional layers, each with a sequence of 32, 32, 64, 64 feature maps, and a filter of size 3 × 3 × 3, respectively. Batch normalization and max-pooling layers are applied after every convolution layer, the filter size of the last max-pooling is 1 × 1 × 1, and all others have a filter of size 2 × 2 × 2. The input of this model is a 32 × 32 × 32 volume where a pulmonary nodule dominates, the result

124

E. Elhoussaine and B. Salwa

feature map from the set of CNN is sent to a classifier e.g., a fully-connected layer with a softmax activation to distingue benign and malign nodules. 3D AlexNet: This architecture is more profound than the previous CNN, inspired by 2D AlexNet architecture. It composed of 6 convolutional layers, each one with a sequence of 32, 32, 64, 64, 128, 128 feature maps, and a filter of size 3 × 3 × 3 respectively. Then, batch normalization and max-pooling layers are applied after each convolution layer, with max-pooling filters of size 2 × 2 × 2, except the last two ones that have a filter of size 1 × 1 × 1. To distingue benign and malign nodules, the result from the last max-pooling layer is connected to a classifier e.g. fully-connected layer with a softmax activation. 3D ResNet: Consists of different stages, with a convolution and identity block at each stage. The identity block (Fig. 1) is used in the case where the output and the input activation have the same dimension, otherwise the convolution block is used; in this case a convolution layer is added to the shortcut path. The implemented model consists of two stages: the first with a 3D convolution layer and the second with a convolution block and two identity blocks. Each convolution block and identity block has three convolution layers. Then average pooling layer and dense layer with softmax are used to perform classification output.

Fig. 1. ResNet identity block

Ensemble Model: Create a new model to better combine the predictions from the models above. First, each of the previous models classifies the input nodule individually, then their output results are combined in one vector and sent to a new fully-connected layer to perform new classification results (Fig. 2).

Pulmonary Nodule Classification

125

Fig. 2. The proposed method

3 Experiments 3.1 Dataset and Preprocessing The LIDC-IDRI dataset offers 1010 different DICOM-format CT scans with a uniform size of 512 × 512. The thickness of the image varies from 0.5 to 5 mm, where 1, 1.25, and 2.5 mm are the recurrent image thicknesses. Each LIDC-IDRI dataset case contains hundreds of images and an XML file providing the details of the lung lesions found. The diameter of each of the observed lesions was measured using electronic calipers, based on their classification there are three main groups of lesions including nodules (with a diameter of size 3–30 mm), non-nodules (with a diameter of size ≥30 mm), and micro-nodules (with a diameter of size ≤3 mm). One to four radiologists annotate each nodule and assign a score of 1 to 5, with 1 and 5 being the extremes of benign and malignancy, respectively. We ignored the zero score, which means that there is no diagnosis available [13]. 3.2 Experiment Settings The proposed approach is implemented based on the Keras framework [14] with TensorFlow as a backend [15]. We choose the binary cross-entropy as the loss function since the classification problem is of a binary nature. To prevent over-fitting, the models are trained using data augmentation technique; horizontal flip, vertical flip, z-axis flip, and random orientation. We consider two sets for training; DS3 and DS4, which are the lung nodules diagnosed by at least three and four radiologists respectively. Then, we compute the median value of annotated scores for a nodule. A median value greater than three is taken as malignant and less than three as benign and a median value equal to three is excluded. The proposed models were trained on both DS3 and DS4 (training set: 70%, test set: 30%) using Adam optimizer, L2 regularizer, and Xiaver initialization method to initialize models weights. 3.3 Experiment Results The experimental results of the proposed networks on the DS3 and DS4 datasets are presented in Table 1 and Table 2. Their ROC curves are shown in Fig. 3 and Fig. 4.

126

E. Elhoussaine and B. Salwa

Since our problem is imbalanced and we care for negative and positive classes equally, we have used the AUC metric rather than accuracy, as the models can easily get a high ac-curacy value by simply labeling all observations as the majority class. The ensemble model obtains the highest AUC, TNR on both DS3 and DS4, and the highest TPR, TNR, and AUC are obtained on DS4, resulting in a TPR of 94%, TNR of 93%, and AUC of 84%. These results indicate that the fully-connected layer learns to perfectly weight the results of the three models in order to obtain optimum performance. The advantage of a fully-connected layer is that it gets a weighted average instead of a standard one. A problem present in the LIDC-IDRI dataset is the ambiguity that exists in defining the malignancy score of a nodule, being the evaluation very subjective. It results in a disparity in the ratings given in the evaluations since radiologists have different opinions when evaluating the same nodule. A model that has been trained with nodules diagnosed by fewer radiologists has a higher chance of being biased than the one trained with nodules diagnosed by more radiologists. This explains why the results from DS4 are good than the results from DS3. Regarding the model size and the number of parameters for the networks (basic 3D CNN with 1 060 862 parameters, 3D AlexNet with 1 271 554 parameters, and 3D ResNet with 539 266 parameters). Although ResNet has complicated and deeper architecture and more layers compared to the basic 3D CNN and AlexNet, the optimization takes advantage of the shortcut connection approach to help achieve better optimal results. Table 1. Performance on DS3 test set. 3D network

TPR% TNR% PPV% AUC%

AlexNet

0.9816 0.2822 0.5776 0.6319

ResNet

0.6835 0.3805 0.5246 0.5320

Basic model

0.9691 0.4546 0.6399 0.7119

Ensemble model 0.7268 0.8026 0.7864 0.7647

Table 2. Performance on DS4 test set. 3D Network

TPR% TNR% PPV% AUC%

AlexNet

0.6695 0.9337 0.9579 0.8200

ResNet

0.9352 0.9167 0.7721 0.8296

Basic model

0.7512 0.7891 0.9157 0.8410

Ensemble model 0.9444 0.9391 0.7925 0.8466

Pulmonary Nodule Classification

127

Fig. 3. ROC curves using DS3 test set.

Fig. 4. Testing ROC curves using DS4 test set.

4 Conclusion In this paper, we proposed an ensemble model using three 3D networks to classify pulmonary nodules in a CT image into benign or malignant classes. Working on 3D images provides better results for the classification of lung nodules compared to the use of approximate 3D images with multi-view or 2D images. One limitation of the proposed

128

E. Elhoussaine and B. Salwa

networks is that they did not take into account the thickness of the CT scans during training, which could affect performance. For future work, we aim at improving performance by using normalized CT scans to avoid the thickness problem, one other future work is automatic pulmonary nodule detection and segmentation; instance segmentation, which will relax the requirement of manual annotations for nodule locations.

References 1. Luís, G., Jorge, N., António, C., Aurélio, C.: Evaluation of the degree of malignancy of lung nodules in computed tomography images (2017) 2. National Lung Screening Trial Research Team et al.: Reduced lung-cancer mortality with low-dose computed tomographic screening. Natl. Engl. J. Med. 2011(365), 395–409 (2011) 3. Senthilkumar, K., Ganesh, N., Umamaheswari, R.: Three-dimensional lung nodule segmentation and shape variance analysis to detect lung cancer with reduced false positives. In: Proceedings of the Institution of Mechanical Engineers, vol. 230, no. 1, pp. 58–70, Journal of Engineering in Medicine (2016) 4. Ying, L., Yoganand, B., Thomas, A., Sanja, A., Qian, L., Ronald, C.W., Gary, S., Pierre, P.M., Matthew, B.S., Robert, J.G.: Radiological image traits predictive of cancer status in pulmonary nodules. In: Clinical Cancer Research, clincanres–3102 (2016). 5. Wei, S., Mu, Z., Feng, Y., Caiyun, Y., Jie, T.: Multi-scale convolutional neural networks for lung nodule classification. In: International Conference on Information Processing in Medical Imaging, pp. 588–599. Springer (2015) 6. Aiden, N., Zhen, H., Dennis, W.: Pulmonary nodule classification with deep residual networks. In: International Journal of Computer Assisted Radiology and Surgery, p. 10 (2017) 7. Kui, L., Guixia, K.: Multiview convolutional neural networks for lung nodule classification. Int. J. Imaging Syst. Technol. 27(1), 12–22 (2017) 8. Sarfaraz, H., Kunlin, C., Qi, S., Ulas, B.: Risk stratification of lung nodules using 3D CNNbased multi-task learning. In: International Conference on Information Processing in Medical Imaging, pp. 249–260. Springer (2017) 9. Hongtao, X., Dongbao, Y., Nannan, S., Zhineng, C., Yongdong, Z.: Automated pulm nary nodule detection in CT images using deep convolutional neural networks (2018). 10. Wentao, Z., Chaochun, L., Wei, F., Xiaohui, X.: DeepLung: deep 3D Dual Path Nets for Automated Pulmonary Nodule Detection and Classification. In: arXiv preprint arXiv:1709. 05538 (2017) 11. Alex, K., Ilya, S., Geoffrey, E.H.: ImageNet classification with deep convolutional neural networks 12. Kaiming, H., Xiangyu, Z., Shaoqing, R., Jian, S.: Deep Residual Learning for Image Recognition 13. Anthony, P.R., Alberto, M.B.: The lung image database consortium (lidc) nodule size report, October. https://www.via.cornell.edu/lidc/ 14. Chollet, F., et al.: Keras (2015.) https://github.com/keras-team/keras 15. Abadi, M., et al.: Large-scale machine learning on heterogeneous systems, 2015. Software available from https://www.tensorow.org/

A Comparative Study on Liver Tumor Detection Using CT Images Abdulfattah E. Ba Alawi(B) , Ahmed Y. A. Saeed, Borhan M. N. Radman, and Burhan T. Alzekri Software Engineering Department, Taiz University, Taiz, Yemen

Abstract. Liver cancer (LC) is a globally known issue. It is one of the most common cancers that can cause human beings. It is a fatal disease spreading especially in developing countries. Many algorithms have been used to perform the detection of liver cancer with the help of both traditional machine learning classifiers and deep learning classifiers. To analyze the performance of commonly used algorithms, this paper attempts a comparative study on LC detection. It includes both machine learning and deep learning techniques; and several methods for liver and tumor detection from CT images are used. With the advances in Artificial Intelligence (AI) and convolution neural networks algorithms, the methods included in this comparative study achieved great results. The best accuracy among traditional machine learning classifiers reaches 90.46% using Support Vector Machine (RBF). Inception V4 pre-trained model obtained 93.15% in terms of testing accuracy, and it is the best classifier among deep learning models. The performance of deep learning models is very promising to take place in medical decisions. Keywords: Liver tumor · CT scan · CNN · Pre-trained model · Deep learning

1 Introduction Liver cancer is the type of cancer that occurs in the liver, which organ is one of the major parts of the human body, which requires our care and caution to keep it sound and healthy. The liver is situated below the right lung and under the ribcage. People who suffer from liver tumors usually died due to inaccurate or late detection. There are several important diagnostic tests for liver cancer such as CT scans and MRIs. In general, every doctor asks the patient to obtain a CT scan to make sure whether a liver tumor exists or not. If doctors find damages in the liver are old, they ask for taking MRI to obtain detailed knowledge of the liver tumor since MRI provides a better view of tumor location. Liver cancer is the common cause of death throughout the world using computed tomography (CT) images; the cancerous tissue can be precisely identified [1]. Because many methods are used for detecting liver cancer, this paper investigates the performance of machine learning and deep learning models commonly used in this respect. The common classifiers have been used to successfully classify abnormal liver cancer features. In this, the effectiveness of liver cancer prediction models is inspected using the assistance of precision, recall, and accuracy. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 129–137, 2021. https://doi.org/10.1007/978-3-030-70713-2_14

130

A. E. Ba Alawi et al.

The rest of this paper is organized as follows: In Sect. 2, the common works in the area of liver tumor detection are reviewed. Section 3 briefly describes the methods that have been applied in this study. The obtained results are analyzed and discussed in Sect. 4. Section 5 is given in the form of a conclusion and recommended further works.

2 Related Works Computer-aided diagnosis systems are commonly used with image processing techniques to identify liver cancer and assist the clinician in decision-making [2]. Many algorithms are applied for the identification of liver tumors, which includes regional approaches, watershed transformation, and machine learning approaches. An automated method was documented using GLCM-based features within a CAD framework to successfully classify liver tumors [3]. Huang et al. [4, 5] designed a Computer Aided Diagnosis procedure for the segmentation and classification of liver tumors using CT images. Their work had been extended to use the auto-covariance texture features for classifying the tumor with an accuracy of 81.7% [4]. Ji et al. [6] suggested an efficient computational model for the clinical diagnosis of hepatocellular carcinoma based on a framework for optimizing particle swarm. A novel, effective, and optimized approach based on Instance Optimization (IO) and SVM have been documented to more accurately identify liver cancer [7]. Li et al. [8] used a regularized level-set assessment approach based on edge distance that was effectively segmented into the cyst, tumor, calculi, and normal liver in CT images. A complete Convolutionary multi-channel network (MC-FCN) model that provides greater precision in CT images of liver tumors is proposed in [9]. The Gray Level Co-occurrence Matrix (GLCM) is used to effectively extract liver tumor features. The extracted statistical features are commonly used in machine learning approaches [10–13] such as support vector machine [15], and back propagation [14], or fuzzy clustering approach with a multi-SVM classifier [16]. Recent works have successfully applied deep learning techniques using DNN to solve a wide range of issues especially in liver tumor detection [9, 17]. The Convolutionary Neural Networks (CNNs) are effectively used in an automated system to segment affected lesions in CT images. The coefficient of dice similarity, it has achieved, is 80.06% [18]. Lu et al. [19] developed a deep learning algorithm with a cut refinement of the graph to segment the CT scans automatically and effectively. Kaizhi et al. [20] designed a system using deep learning for the classification of liver diseases. Hu et al. [14] addressed deep learning strategies such as Convolutionary neural networks in a recent survey study Completely Convolutionary network, auto-encoders, and deep conviction networks for cancer detection and diagnosis. In the work [21], the liver was initially isolated by marker-controlled watershed segmentation method and the lesion caused by cancer was eventually split into the Gaussian model mixture protocol. The deep neural classifiers are used for efficient recognition of three types of liver cancer; they are hemangioma, hepatocellular carcinoma, and metastatic. In the paper [22], Liver Function Tests (LFT) evidence is used in the diagnosis of computer-assisted Liver disease screening. The authors suggested a tightly related deep neural network with 13 LFT markers and population knowledge of liver disease screening subjects. A data set

A Comparative Study on Liver Tumor Detection Using CT Images

131

with 76,914 was used and the under curve area of DenseDNN reaches 0.8919; the under curve area of DNN is 0.8867; the under-forest of random forest is 0.8790; and the rational regression reaches 0.7974. DenseDNN demonstrates higher results than DNN in comparison with the deep learning methods. This paper presents a comparative study on deep learning classification and the detection of the region of liver tumors. Nineteen classifiers are used for recognizing liver tumors in CT images. In deep learning approach, about thirteen classifiers are used including (ResNet-50, DenseNet121, DenseNet201, GoogLeNet, InceptionV4, AlexNet. SequeezNet1.0, Se-queezeNet1.1, VGG11, VGG13, VGG16, VGG19, and Xception). This is n addition to the implementation of six traditional machine learning classifiers which include Support Vector Machine (SVM), Radial Base Gaussian Function (RBF), and K-Nearest Neighbors (KNN), to name just a few.

3 Methodology To analyze commonly used classifiers for clinical diagnosis and computer-aided decision systems, two approaches of artificial intelligence are investigated in this comparative study. These approaches are given in the following figure.

Fig. 1. Steps commonly followed for detecting liver tumors.

As shown in the above figure (Fig. 1), for machine learning algorithms, feature extraction is done then training the classifiers on the extracted features (e.g. SVM, KNN, etc.). The processioning operation is resizing the images to 128 × 128 before extracting features with HOG descriptor. However, for deep learning algorithms, the images are processed to have a size of 224 × 224 to fit the dimensions of the first layer of the pre-trained models. Only Inception V4 needs input images with a size of 299 × 299 because it has an input layer with 299 × 299 dimensions. Then, the pre-trained models are retrained in the liver tumor dataset.

132

A. E. Ba Alawi et al.

3.1 Data Collection The used dataset was downloaded from TCGA [23, 24], 3D-IRBADb 01 [25], and the Data of CHAOS Challenge - Combined (CT-MR) Healthy Abdominal Organ Segmentation [26]. After the removal of bad images and anomalies of the dataset, we get 735 images divided into 350 images as a normal class and 385 images as an abnormal class. The dataset was ready to be preprocessed to analyze the performance of machine learning and deep learning classifiers and classify liver CT scans as normal and abnormal. The dataset divided into 3 partitions; about 515 images for training, 147 images for validation, and 73 images for testing. 3.2 Data Augmentation To prevent the model from overfitting, and to ensure a balanced classification, the data required an augmentation process. Various augmentation operations such as Salt and paper noise and Gaussian noise are used. Also, all normal images rotated from angle 1 to angle 20. 3.3 Machine Learning Techniques The classification of liver images is performed using traditional machine learning classifiers. In this approach, the images are preprocessed and extracted using Histogram Oriented Gradients (HOG) descriptor. Then, the extracted features are classified using skin machine learning classifiers to analyze the performance of each one. 3.4 Deep Learning Techniques Convolution Neural Networks (CNN) algorithm is used here for its vital role in image classification tasks. The power of CNN is in a hidden area between input and output layers. The classification tasks by CNN show high-performance findings. Figure 2 illustrates the architecture of CNN.

Fig. 2. Convolution neural networks.

A Comparative Study on Liver Tumor Detection Using CT Images

133

3.5 Transfer Learning Transfer learning is a new technique that has been used recently. The most important advantage of this technique is that it reduces the required time and resources for training. Instead of training from scratch that takes more time and GPU resource and a large dataset of images, the pre-trained model (e.g. ResNet50, AlexNet, and GoogLeNet) is used to transfer the knowledge and perform the task. 3.6 Deep Learning Pre-trained Model In this study, thirteen pre-trained models are used; they are ResNet -50, DenseNet121, DenseNet201, GoogLeNet, InceptionV4, AlexNet, SequeezNet1.0, SequeezeNet1.1, VGG11, VGG13, VGG16, VGG19, and Xception. These models are fine-tuned by replacing the last layer of the pre-trained models with suitable layers according to the number of classes in the fully connected layers. The training phase steps are summarized in Fig. 3 below.

Fig. 3. Training and testing steps in a deep learning approach.

To evaluate the performance of the deep learning pre-trained models, the test images are inputted. Then input images are processed and tested on the obtained expertise model to recognize whether or not the liver CT scan image contains a tumor.

4 Results and Discussion The experiment was carried out using an HP laptop having 4 GB RAM, and Core i5 Microprocessor. During the training phase using the collected dataset, the loss related to

134

A. E. Ba Alawi et al.

each phase is used as a performance metric; and besides accuracy, Precision, and recall, 32-batch size and 25 epochs are used for training deep learning pre-trained models. The performance of these models in terms of training loss is depicted in Fig. 4.

Fig. 4. Training loss of deep learning models.

The training loss of Inception-V4 and Xception was the best among the pre-training models achieving 0.04 and 0.035, respectively. AlexNet pre-trained model achieved a loss of 0.5. Therefore, it can be regarded as the worst performance of all pre-trained models during training phase. The performance of pre-trained models in terms of validation loss is represented in the following diagram (Fig. 5). In terms of validation loss, DenseNet-201 and SqueezeNet-1.0 pre-trained models achieved the best performance. The following table summarizes the performance of all deep learning pre-trained models during testing phase: The above table (Table 1) illustrates the performance of the pre-trained models in detecting liver tumors during testing phase. The performance of Inception-V4 is the best reaching an accuracy of 93.15%. This indicates that the pre-trained models with more layers achieved better results than others can do. However, some deep learning models that have more layers do not perform well such as GoogLeNet that reaches the best training loss and the worst testing performance. In machine learning approach, the following table (Table 2) shows the performance of machine learning classifiers with a cross-validation of k = 7. Vividly, Kernel SVM (Radial Basis Function RBF) achieved the best performance with an accuracy reached 90.64%. In terms of accuracy and recall, both decision tree, random forest, and Naïve Bayes performed poorly during testing phase. The experimental findings show the feasibility of using ML and DL techniques in diagnosing liver tumor.

A Comparative Study on Liver Tumor Detection Using CT Images

Fig. 5. The obtained training loss for the pre-trained models.

Table 1. Performance Metrics Results using ResNet-50 Pre-trained model. Pre-trained model

Testing accuracy

Recall

Precision

F1_score

AlexNet

90.41%

90.86%

90.49%

90.41%

DenesNet-121

86.30%

86.57%

85.83%

84.73%

DenesNet201

87.67%

88.87%

88.57%

87.00%

GoogLeNet

86.30%

87.17%

88.11%

87.21%

Inception-V4

93.15%

91.84%

92.26%

91.70%

Vgg-11

83.56%

84.44%

84.50%

83.34%

Vgg-13

89.04%

88.67%

89.91%

88.20%

Vgg-16

90.41%

91.43%

89.04%

89.85%

Vgg-19

90.41%

89.73%

91.73%

89.48%

ResNet50

84.93%

84.94%

84.85%

84.84%

SqueezeNet-V1.0

87.67%

88.52%

89.93%

87.14%

SqueezeNet-V1.1

89.04%

89.75%

89.99%

89.76%

Xception

89.25%

89.14%

88.86%

89.04%

135

136

A. E. Ba Alawi et al. Table 2. The performance of machine learning classifiers.

ML classifier

Accuracy

Precision

Recall

Value

Standard division

Value

Standard division

Value

Standard division

SVM (kernel RBF)

90.46%

± 2.83%

91.63%

± 6.95%

90.41%

± 5.77%

Linear Regression

78.88%

± 4.99%

77.36%

± 5.96%

84.66%

± 7.65%

KNN, k = 7

89.51%

± 1.98%

90.90%

± 5.23%

89.32%

± 4.52%

KNN, k = 5

90.19%

± 2.21%

92.51%

± 5.39%

88.56%

± 6.53%

Naïve Bayes

76.29%

± 4.94%

81.32%

± 9.39%

72.16%

± 7.17%

Random Forest 79.84%

± 3.84%

93.55%

± 4.06%

66.33%

± 9.26%

Decision Tree

± 6.67%

85.17%

±11.20%

62.91%

± 23.72%

73.30%

5 Conclusion and Future Work The present clinical results obtained by using the developed modalities of CT imaging techniques are excellent with an accuracy of around 90% in the validation process. The results are expected to show a high impact on the diagnostic process. By combining these techniques effectively, all targeting different properties of malignant tissue of the liver could be diagnosed. As already discussed, this comparative study analyzes thirteen pre-trained models and six traditional machine-learning classifiers. As findings of this study, the best performance was achieved by Inception V4 pre-trained model with accuracy, precision, recall, and F_1 measure of 93.15%, 91.84%, 92.26%, and 91.70%, respectively. Among traditional ML classifiers, SVM (RBF) achieved the best accuracy that reaches 90.46%. To use the ensemble learning technique and apply segmentation process with the studied models are left for future work.

References 1. Bartolozzi, C., Ciatti, S., Lucarelli, E., Villari, N., de Dominicis, R.: Ultrasound and computer tomography in the evaluation of focal liver disease. Acta Radiologica. Diagnosis 22(5), 545– 548 (1981) 2. Kononenko, I.: Machine learning for medical diagnosis: history, state of the art and perspective. Artif. Intell. Med. 23(1), 89–109 (2001) 3. Chen, E.-L., Chung, P.-C., Chen, C.-L., Tsai, H.-M., Chang, C.-I.: An automatic diagnostic system for CT liver image classification. IEEE Trans. Biomed. Eng. 45(6), 783–794 (1998) 4. Huang, Y.-L., Chen, J.-H., Shen, W.-C.: Diagnosis of hepatic tumors with texture analysis in nonenhanced computed tomography images. Academic Radiol. 13(6), 713–720 (2006) 5. Huang, Y.-L., Chen, J.-H., Shen, W.-C.: Computer-aided diagnosis of liver tumors in nonenhanced CT images. Comput. Biol. Med. 9, 141–150 (2004) 6. Ji, Z., Wang, B.: Identifying potential clinical syndromes of hepatocellular carcinoma using PSO-based hierarchical feature selection algorithm. BioMed Res. Int. 2014, 1–12 (2014).

A Comparative Study on Liver Tumor Detection Using CT Images

137

7. Jiang, H., Zheng, R., Yi, D., Zhao, D.J.: A novel multiinstance learning approach for liver cancer recognition on abdominal CT images based on CPSO-SVM and IO. Comput. Math. Methods Med. 2013, 1–10 (2013) 8. Li, C., Xu, C., Gui, C., Fox, M.D.: Distance regularized level set evolution and its application to image segmentation. IEEE Trans. Image Process. 19(12), 3243–3254 (2010). 9. Sun, C., Guo, S., Zhang, H., Li, J., Chen, M., Ma, S., Jin, L., Liu, X., Li, X., Qian, X.J.: Automatic segmentation of liver tumors from multiphase contrast-enhanced CT images based on FCNs. Artif. Intell. Med. 83, 58–66 (2017) 10. Haralick, R.M., Shanmugam, K.: Its’Hak Dinstein: Textural features for image classification. IEEE Trans. Syst. Man Cybern. SMC-3(6), 610–621 (1973) 11. Newell, D., Nie, K., Chen, J.-H., Hsu, C.-C., Hon, J.Y., Nalcioglu, O., Su, M.-Y.: Selection of diagnostic features on breast MRI to differentiate between malignant and benign lesions using computer-aided diagnosis: differences in lesions presenting as mass and non-mass-like enhancement. Eur. Radiol. 20(4), 771–781 (2010) 12. Nie, K., Chen, J.-H., Hon, J.Y., Chu, Y., Nalcioglu, O., Su, M.-Y.: Quantitative analysis of lesion morphology and texture features for diagnostic prediction in breast MRI. Acad. Radiol. 15(12), 1513–1525 (2008) 13. Moon, W.K., Shen, Y.-W., Huang, C.-S., Chiang, L.-R., Chang, R.-F.: Biology: Computeraided diagnosis for the classification of breast masses in automated whole breast ultrasound images. 37(4), 539–548 (2011) 14. Hu, Z., Tang, J., Wang, Z., Zhang, K., Zhang, L., Sun, Q.J.P.R.: Deep learning for image-based cancer detection and diagnosis− a survey. 83, 134–149 (2018) 15. Devi, P., Dabas, P.: Liver tumor detection using artificial neural networks for medical images. IJIRST 2(3), 34–38 (2015) 16. Sakr, A.A., Fares, M.E., Ramadan, M.: Automated focal liver lesion staging classification based on Haralick texture features and multi-SVM. Int. J. Comput. Appl. 91(8), 0975–8887 (2014) 17. Ben-Cohen, A., Klang, E., Kerpel, A., Konen, E., Amitai, M.M., Greenspan, H.J.N.: Fully convolutional network and sparsity-based dictionary learning for liver lesion detection in CT examinations. 275, 1585–1594 (2018) 18. Li, C., Wang, X., Eberl, S., Fulham, M., Yin, Y., Chen, J., Feng, D.: A likelihood and local constraint level set model for liver tumor segmentation from CT volumes. 60(10), 2967–2977 (2013) 19. Lu, F., Wu, F., Hu, P., Peng, Z., Kong, D.: Surgery: Automatic 3D liver location and segmentation via convolutional neural network and graph cut. 12(2), 171–182 (2017) 20. Wu, K., Chen, X., Ding, M.J.O.: Deep learning based classification of focal liver lesions with contrast-enhanced ultrasound. 125(15), 4057–4063 (2014) 21. Das, A., Acharya, U.R., Panda, S.S., Sabut, S.: Deep learning based liver cancer detection using watershed transform and Gaussian mixture model techniques. 54, 165–175 (2019) 22. Yao, Z., Li, J., Guan, Z., Ye, Y., Chen, Y.J.N.N.: Liver disease screening based on densely connected deep neural networks. 123, 299–304 (2020) 23. NBIA Dataset. https://nbia.cancerimagingarchive.net/.Accessed 25 Jan 2020 24. Erickson, B., Kirk, S., Lee, Y., Bathe, O., Kearns, M., Gerdes, C., Rieger-Christ, K., Lemmerman, J.: Radiology Data from The Cancer Genome Atlas Liver Hepatocellular Carcinoma [TCGA-LIHC] collectionThe. (2016). 25. IRCAD France Dataset. https://www.ircad.fr/research/3d-ircadb-01/ (2020). Accessed 25 Jan 2020 26. Kavur, A.E., Gezer, N.S., Barı¸s, M., Conze, P.-H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., Özkan, S., Baydar, B.: CHAOS Challenge--Combined (CT-MR) Healthy Abdominal Organ Segmentation (2020)

Brain Tumor Diagnosis System Based on RM Images: A Comparative Study Ahmed Y. A. Saeed(B) , Abdulfattah E. Ba Alawi, and Borhan M. N. Radman Software Engineering Department, Taiz University, Taiz, Yemen

Abstract. Cancers or tumors have their impact effects on humans, especially if the cancer is localized in an important organ such as the brain. It is important to detect cancer earlier so that many lives can be saved. As cancer diagnosis is highly time-consuming and needs expensive tools, there is an immediate requirement to develop non-invasive, cost-effective, and efficient tools for brain cancer staging and detection. Brain scans that are commonly used are magnetic resonance imaging (MRI) and computed tomography (CT). In this paper, we studied the common algorithms that are used for brain tumor detection using imaging modalities of brain cancer and automatic computer-assisted methods. The main objective of this paper is to make a comparative analysis of several methods of detecting tumors in the Central Nervous System (CNS). The results of the applied classifiers are compared and analyzed using different metrics including accuracy, precision, and recall. The best accuracy reached using machine learning algorithms is 85.56% accuracy with Random Forest, while the best classifier among applied deep learning algorithms is Inception V4 with 97.36%. Keywords: Brain cancer · Central nervous system tumor · Pathophysiology · Deep learning

1 Introduction The brain is the central nervous system control hub that helps the entire human body to carry out its operations. Tumors in the brain will directly threaten people’s lives. The patients would be more likely to live if the tumor is identified at an early stage. MR imagery is commonly used by doctors to assess if cancer defects are present or the tumor is determined [1]. MR is a form of resonance imaging which has become a hot field of research. Many researchers have sought to develop smart structures to classify brain cancer into various groups such as brain tissues for normal, pathological, biting, and malignant, low-grade and high-grade forms. Main carcinoma cells that affect the brain are considered the worse cancer not only due to the weak prognostics but also because of their strong effects on executive ability loss and reduced life expectancy. Lymphomas and gliomas, in the main central nervous system that are responsible for nearly 80% of malignant cancers [2], are the most prominent major brain tumors in adults. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 138–147, 2021. https://doi.org/10.1007/978-3-030-70713-2_15

Brain Tumor Diagnosis System Based on RM Images

139

Brain cancer has various degrees. Complex prognostic results are found for low gliomas (LGGs) with an average 10-year survival rate of approximately 57% [3, 4]. Past results indicate that brain tumors, which have been recently diagnosed, can be used to predict potential diagnosis and treatment strategies using MRI characteristics [5–7]. Feature selection is one of the most critical issues for brain tumor diagnosis and segmentation. Due to the importance of detecting brain tumors in early-stage and the plenty of methods that have been used by researchers in this area, we performed a comparative study to analyze the common methods to justify the performance of twenty-two classifiers. This paper is divided into five sections. The following section (Sect. 2) presents the related works on brain tumor detection. Section 3 illustrates the methods of brain tumor detection in addition to illustrating the techniques that are used such as CNN pre-trained models. Section 4 presents a full description of the obtained results of the proposed system. Section 5 is the concluded given in the form of a sum-up and recommendations.

2 Related Works There are various revolutionary processing methods, which have been demonstrated to improve the detailed diagnosis and segmentation of brain tumors at the same time. In tumor segmentation on average, Huang et al. [8] obtained an accuracy of 74.75% on a subspace mapping basis. Previous studies adopted various models that are commonly used. These models are Support Vector Machine (SVM) and the Neural Network (NN), which demonstrated strong tumor classification results. However, before grouping, the manual collection of features in common. Studies on brain tumor classification found that Soltaninejad et al. [9] have been utilizing 38 first-order or second-order statistical tests to rate tumors of various grades depending on SVM. More than 80% of their tests indicates the quality of 21 patients with various scoring combinations. The treatment involves segmented tumor slices as models and features that are chosen carefully before training. The current advances that have recently risen from deep learning approaches such as Convolutionary Neural Network have been shown to be good in classifying objects. Comparatively speaking, deeper learning models are typically unattended learning models, which randomly learn the characteristics of the entity from the data. Ethiopia [10] indicates that a high-performance feature detector with the Convolutional Neural Networks was used in a research that focused on the ImageNet dataset. The detector achieved 15.8 percent precision with ImageNet results, besides a quantitative improvement of 70% over other previous work. Generally speaking, in previous studies on brain segmentation there can be unsupervised learning methods [11–14] and supervised [15–19] learning strategies. The present study attempts to comparatively examine AI-based brain cancer diagnosis models while using deep learning and while using machine learning approaches. Eight traditional machine learning classifiers are applied (e.g. Naïve Bayes, Logistic Regression, Decision Tree, Support Vector Machine, K-Nearest Neighbors, Random Forest), besides thirteen pre-trained models (e.g. ResNet18, ResNet50, ResNet101, ResNet152, ResNext50, ResNext101, SqueezeNet1_0, SqueezeNet1_1, AlexNet, DenseNet121, DenseNet201,GoogLeNet, Inception V4).

140

A. Y. A. Saeed et al.

3 Methods This section discusses the basic methods for developing AI-based diagnosis systems to recognize brain cancer based on images. Figure 1 shows the basic methods for brain cancer classification.

Fig. 1. The basic methods for building brain tumor diagnosis models.

As shown in Fig. 1, machine learning and deep learning methods as commonly used to diagnose brain tumors. After working on image acquisition and preprocessing, the images are forwarded to deep neural networks in the deep learning approach. However, in the traditional machine learning approach, image features extraction and segmentation are applied. 3.1 Dataset Collection Phase The dataset of this model was collected from two public sources. Around 3000 images were collected from [20], and about 698 images were collected from [21]. All collected images dataset from the previous sources are MRI. These images were very precisely collected. From all the collected images, only 160 images for tumor samples were used, besides 216 images used for a non-tumor class. These images are divided into three partitions: 264 images for training, 75 images for validation, and 38 images for testing.

Brain Tumor Diagnosis System Based on RM Images

141

3.2 Data Selection The above-mentioned downloaded images are selected precisely after getting rid of images that have poor resolution. 3.3 Covert 3D MRI to JPEG Images The dataset needs to be processed in a 2D CNN model. Therefore, we used a script for the conversion process. The output of this phase is a 2-D image. 3.4 Deep Learning Method In this approach, fourteen pre-trained models have been implemented by the following steps such as: Preparation of the Dataset. This task aims at preparing the training images in a specific folder to start a training process. Image Pre-processing. During this phase, each image is resized or rescaled to 224 * 224 to fit the input layer of the pre-trained models (e.g. ResNet-50). Only, Inceptun-V4 requires images with 299 × 299 size, because the first layer of Inception pre-trained model has the size of 299 × 299. Retraining the Pre-trained Model. Transfer Learning is applied to transferred to the knowledge of the pre-trained model (ResNet) to perform new tasks of classification with the dataset of the brain tumor. This task requires using the model which trained on a large dataset to be retrained in the task of classifying brain tumor. Obtaining the Expertise Model. The output of the previous steps is a trained model that can recognize the brain tumor. These steps are performed using Python with Pycharm environment. The model has the extension (.pt) and saved in a specific file. The previous steps of the training stage are summarized in the following Fig. 2.

Input BT Dataset

Preprocessing

Retrain the Pre-trained model

Save Trained Model Weights

Fig. 2. Training the pre-trained models.

3.5 Machine Learning Approach To this approach, more than six classification algorithms have been applied using Histogram Oriented Gradients (HOG) for feature extraction. The Principle Component Analysis (PCA) algorithm has been implemented to reduce the feature vector. Eight traditional machine learning classifiers are applied (e.g. Naïve Bayes, Logistic Regression, Decision Tree, Support Vector Machine, K-Nearest Neighbors, Random Forest).

142

A. Y. A. Saeed et al.

3.6 Deep Learning Approach CNN is a popular technique in supervised learning methods that have greatly evolved at the end of the 20th century. It mimics the human brain’s function. Furthermore, it indicates strong success in the field of 2D data classification with CNN-based algorithms (e.g., LeNet-5, ResNet, and DenseNet). It shows a testing failure at a rate of less than 1%, based on a neural network and the innovative neural network model centered on a CNN system. CNN is now commonly used in the image processing area.

Fig. 3. Convolution Neural Networks.

3.7 Transfer Learning In the deep learning domain, it is a common method to start training a new model using an advanced pre-trained model, rather than arbitrarily initializing parameters of the current one. A model is pre-trained, but designed with distinct datasets for a specific or separate mission, with a similar or identical design to the new model. Transfer learning which is beneficial in a variety of ways acquires the visual representations of a pre-trained model from a massive dataset of millions of samples: shortened time for testing the new model, future advancement of the new model, and less training data from the new task area. In cases with different learning activities and data sets, the value of transfer learning is regularly used. For example, the recognition of objects, scene recognition [22], and object recognition by natural images, the classification of interstice lung diseases with CT images [23] and [24]. 3.8 Experimental Setup This experiment is performed using an HP laptop, with 8 GB of RAM. Colab environment and Rapid Miner Tools are used, with using the 32-batch size and 25 epochs for deep learning models. In deep learning classifiers, the dataset was divided into a training set, validation set, and testing set as 70%, 20%, and 10% respectively. However, the cross-validation was applied to machine learning classifiers with k = 10.

Brain Tumor Diagnosis System Based on RM Images

143

4 Results and Discussion In this study, 22 different classifiers have been used for cancer detection including nine traditional Machine Learning classifiers, and thirteen pre-trained models. For machine learning approach, KNN, Linear Regression, Decision tree, Logistic regression, SVM, Random Forest, and Gradient Boosted Trees, Multilayer Perceptron, K-Nearest Neighbored, and Radial Base Function (RBF) are used. The best accuracy of the traditional ML classifiers was Random Forest; it achieved a value of 85.65% in terms of testing accuracy. Whereas, K- Nearest Neighbors (KNN) had shown the worst results as 81.37%. The accuracy, precision, and Recall, of the used machine learning classifiers, are shown in Table 1. Table 1. The performance of machine learning models. The classifier\Metric

Accuracy Value

Standard deviation

Precision Value

Standard deviation

Recall Value

Standard deviation

Naive Bayes

83.50%

0.0563

83.44%

0.0584

83.55%

0.0601

Logistic Regression

81.36%

0.0348

81.42%

0.0396

81.61%

0.0437

Decision Tree

83.53%

0.0473

83.62%

0.0479

84.10%

0.0491

Random Forest

85.65%

0.073

85.70%

0.0479

86.15%

0.0364

Gradient Boosted 82.39% Trees

0.0614

82.44%

0.0592

81.95%

0.0609

Support Vector Machine SVM

82.43%

0.0294

82.84%

0.0315

82.43%

0.0345

K- Nearest Neighbors KNN, K=5

81.37%

0.028

81.50%

0.0479

81.43

0.0467

Mlti-layer Perceptron MLP

82.45%

0.0489

82.25%

0.0485

82.45%

0.0486

Radial Base Function RBF

82.72%

0.055

90.62%

0.0508

66.24%

0.1312

In deep learning, thirteen pre-trained models are used with the help of AlexNet, DenesNet-121, DenesNet-201,’GoogleNet, Inception-V4, ResNet-18, ResNet50, ResNet-101, ResNet-152, ResNext-50, ResNext-101, SqueezeNet-1_0, and SqueezeNet-1_1. Training and Validation loss of all pre-trained models are shown in Fig. 4. Figure 4 shows the obtained loss during training deep learning models for 25 epochs. The best training loss obtained is by using Inception V4. Figure 5 shows the validation loss for CNN-based models.

144

A. Y. A. Saeed et al.

Fig. 4. The training loss of the pre-trained models.

Fig. 5. The validation loss of the pre-trained models.

Brain Tumor Diagnosis System Based on RM Images

145

As Fig. 5 shows, the validation loss of the pre-trained models. AlexNet reached the best validation loss at around 0.2 in epoch 11 and 15, but with inconsistent performance. Table 2 shows a comparison between the classifiers used in the pre-trained phase in terms of test accuracy, precision, and recall. Table 2. The performance of deep learning models. The classifier\Metric Test accuracy Precision Recall DenesNet121

89.47%

78.47%

78.12%

DeneseNet201

86.84%

92.26%

92.05%

ResNet50

86.84%

76.64%

76.56%

ResNet101

86.84%

81.22%

88.72%

ResNet18

94.73%

97.49%

96.42%

ResNet152

86.84%

88.65%

81.15%

ResNext50

84.21%

83.92%

85.93%

ResNext101

78.94%

80.63%

82.44%

AlexNet

84.21%

89.04%

79.68%

GoogleNet

89.47%

86.86%

89.18%

SqueezeNet-v1.0

81.57%

83.33%

84.37%

SqueezeNet-v1.1

92.10%

95.03%

95.44%

Inception-V4

97.36%

98.61%

98.33%

As Table 2 shows, the model that achieves the best result in term of testing accuracy was Inception-V4 with 97%: whereas, ResNext-101 was the worst model showing a poor accuracy of 78.94%. Vividly, the experimental findings show the outperformance of Random Forest among traditional machine learning classifiers, besides Inception-V4 that achieved the best performance among deep learning models.

5 Conclusion and Recommendations This study focused on identifying the common deep learning and machine learning algorithms in diagnosing the brain tumor. The findings indicate that the best machine learning classifier achieved the best results was Random Forest with 85.6% accuracy. Among the deep learning pre-trained models, Inception-V4 is found to best perform the task achieving 97.36%. The pre-trained models that have more deeply layers achieved better results than others did. This study recommends the application of these classifiers for the different brain tumor datasets. For future works, using a large dataset to make the results more generalized.

146

A. Y. A. Saeed et al.

References 1. Pereira, S., Pinto, A., Alves, V., Silva, C.A.: Brain tumor segmentation using convolutional neural networks in MRI images. IEEE Trans. Med. Imaging 35(5), 1240–1251 (2016) 2. Schwartzbaum, J.A., Fisher, J.L., Aldape, K.D., Wrensch, M.: Epidemiology and molecular pathology of glioma. Nature Clin. Practice Neurol. 2(9), 494–503 (2006) 3. Ramakrishna, R., Hebb, A., Barber, J., Rostomily, R., Silbergeld, D.J.N.: Outcomes in reoperated low-grade gliomas. 77(2), 175–184 (2015) 4. Jacob, C.P.: Post prandial hypertriglyceridemia in patients with CAD and without CAD: A comparative study. Sree Mookambika Institute of Medical Sciences, Kulasekharam (2018) 5. Mazzara, G.P., Velthuizen, R.P., Pearlman, J.L., Greenberg, H.M., Wagner, H.: Brain tumor target volume determination for radiation treatment planning through automated MRI segmentation. Int. J. Radiat. Oncol. Biol. Phys. 59(1), 300–312 (2004). 6. Yamahara, T., Numa, Y., Oishi, T., Kawaguchi, T., Seno, T., Asai, A., Kawamoto, K.: Morphological and flow cytometric analysis of cell infiltration in glioblastoma: a comparison of autopsy brain and neuroimaging. Brain Tumor Pathol. 27(2), 81–87 (2010) 7. Bauer, S., Wiest, R., Nolte, L.-P., Reyes, M.: Biology: a survey of MRI-based medical image analysis for brain tumor studies. Phys Med Biol. 58(13), R97 (2013) 8. Huang, W., Yang, Y., Lin, Z., Huang, G.-B., Zhou, J., Duan, Y., Xiong, W.: Random feature subspace ensemble based extreme learning machine for liver tumor detection and segmentation. In: 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2014, pp. 4675–4678. IEEE (2014) 9. Soltaninejad, M., Ye, X., Yang, G., Allinson, N., Lambrou, T.: Brain tumour grading in different MRI protocols using SVM on statistical features (2014) 10. Le, Q.V.: Building high-level features using large scale unsupervised learning. In: 2013 IEEE international conference on acoustics, speech and signal processing 2013, pp. 8595–8598. IEEE (2013) 11. Szilagyi, L., Lefkovits, L., Benyo, B.: Automatic brain tumor segmentation in multispectral MRI volumes using a fuzzy c-means cascade algorithm. In: 2015 12th International Conference on Fuzzy Systems and Knowledge Discovery (FSKD) 2015, pp. 285–291. IEEE (2015) 12. Mei, P.A., de Carvalho Carneiro, C., Fraser, S.J., Min, L.L., Reis, F.: Analysis of neoplastic lesions in magnetic resonance imaging using self-organizing maps. J. Neurol. Sci. 359(1–2), 78–83 (2015). 13. Juan-Albarracin, J., Fuster-Garcia, E., Manjon, J.V., Robles, M., Aparici, F., Martí-Bonmatí, L., Garcia-Gomez, J.M.: Automated glioblastoma segmentation based on a multiparametric structured unsupervised classification. Plos One 10(5) (2015) 14. Rajendran, A., Dhanasekaran, R.J.P.E.: Fuzzy clustering and deformable model for tumor segmentation on MRI brain image: a combined approach. 30, 327–333 (2012) 15. Wu, W., Chen, A.Y., Zhao, L., Corso, J.J.: Surgery: brain tumor detection and segmentation in a CRF (conditional random fields) framework with pixel-pairwise affinity and superpixel-level features. 9(2), 241–253 (2014) 16. Pinto, A., Pereira, S., Correia, H., Oliveira, J., Rasteiro, D.M., Silva, C.A.: Brain tumour segmentation based on extremely randomized forest with high-level features. In: 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 2015, pp. 3037–3040. IEEE (2015) 17. Soltaninejad, M., Yang, G., Lambrou, T., Allinson, N., Jones, T.L., Barrick, T.R., Howe, F.A., Ye, X.: Automated brain tumour detection and segmentation using superpixel-based extremely randomized trees in FLAIR MRI. Int. J. Comput. Assist. Radiol. Surg. 12(2), 183–203 (2017).

Brain Tumor Diagnosis System Based on RM Images

147

18. Jafari, M., Kasaei, S.J.: Automatic brain tissue detection in MRI images using seeded region growing segmentation and neural network classification. Australian J. Basic Appl. Sci. 5(8), 1066–1079 (2011) 19. Subbanna, N., Precup, D., Arbel, T.: Iterative multilevel MRF leveraging context and voxel information for brain tumour segmentation in MRI. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2014, pp. 400–405 (2014) 20. Navoneel, C.: Brain MRI Images for Brain Tumor Detection (2020). https://www.kaggle. com/navoneel/brain-mri-images-for-brain-tumor-detection. Accessed 16 May 2020 21. Nbia Cancer Imaging Archive (2020). https://nbia.cancerimagingarchive.net/nbia-search/ 22. Yu, W., Yang, K., Bai, Y., Xiao, T., Yao, H., Rui, Y.: Visualizing and comparing AlexNet and VGG using deconvolutional layers. In: Proceedings of the 33rd International Conference on Machine Learning (2016) 23. Agostinelli, F., Hoffman, M., Sadowski, P., Baldi, P.: Learning activation functions to improve deep neural networks (2014) 24. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M.: Imagenet large scale visual recognition challenge. Int. J. Comput. Vis. 115(3), 211–252 (2015)

Diagnosis of COVID-19 Disease Using Convolutional Neural Network Models Based Transfer Learning Hicham Moujahid1(B) , Bouchaib Cherradi1,2 , Mohammed Al-Sarem3 , and Lhoussain Bahatti1 1 SSDIA Laboratory, ENSET of Mohammedia, Hassan II University of Casablanca,

28820 Mohammedia, Morocco 2 STIE Team, CRMEF Casablanca-Settat, Provincial Section of El Jadida,

24000 El Jadida, Morocco 3 Information Systems Department, Taibah University, Al-Madinah Al-Monawarah,

Kingdom of Saudi Arabia

Abstract. COVID-19 disease is similar to normal pneumonia caused by bacteria or other viruses. Therefore, the manual classification of lung diseases is very hard to discover, particularly the distinction between COVID-19 and NON-COVID-19 disease. COVID-19 causes infections on one or both lungs which appear as inflammations across lung cells. This can lead to dangerous complications that might cause death in the case of gaining or having an immune disease. The problem of COVID-19 is that its symptoms are similar to conventional chest respiratory diseases like flu disease and chest pain while breathing or coughing produces mucus, high fever, absence of appetite, abdominal pain, vomiting, and diarrhea. In most cases, a deep manual analysis of the chest’s X-ray or computed tomography (CT) image can lead to an authentic diagnosis of COVID-19. Otherwise, manual analysis is not sufficient to distinguish between pneumonia and COVID-19 disease. Thus, specialists need additional expensive tools to confirm their initial hypothesis or diagnosis using real-time polymerase chain reaction (RT-PCR) test or MRI imaging. However, a traditional diagnosis of COVID-19 or other pneumonia takes a lot of time from specialists, which is so significant parameter in the case of a pandemic, whereas, a lot of patients are surcharging hospital services. In such a case, an automatic method for analyzing x-ray chest images is needed. In this regard, the research work has taken advantage of proposing a convolutional neural network method for COVID-19 and pneumonia classification. The X-ray processing have been chosen as a diagnosis way because of its availability in hospitals as a cheap imaging tool compared to other technologies. In this work, three CNN models based on VGG-16, VGG19, and MobileNet were trained using the zeroshot transfer learning technique. The best results are obtained on VGG-19 based model: 96.97% accuracy, 100% precision, 100% F1-score, and 99% recall. Keywords: Convolutional neural network · Transfer learning · COVID-19 · Pneumonia · X-ray images

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 148–159, 2021. https://doi.org/10.1007/978-3-030-70713-2_16

Diagnosis of COVID-19 Disease

149

1 Introduction The novel coronavirus (COVID-19) also called SARS-CoV-2 appeared the first time in Wuhan city, China. This virus propagated from animals to humans in December 2019. It has been spread mainly among people through the respiratory system. The droplets of the infected person most likely contain COVID-19 viruses and spread to others in many ways like coughing, sneezing, or even talking to someone else. The spreading factors continue to increase until the World Health Organization (WHO) declared coronavirus as a pandemic threatening every human’s life on 11 February 2020. By August 9, 2020, the spread of virus reaches 20 million cases all over the world and 730 000 deaths1 . Pneumonia has almost the same symptoms as COVID-19 disease which makes it hard to diagnose and differentiate between them. Pneumonia is generally an inflammation in the small air sacs in lung. It can be caused by many germs like viruses, bacteria, and fungi. Therefore, specialists need additional tools like blood tests and deep x-ray imaging. Because of the remarkable increase of COVID-19 infected patients, the manual diagnosis is not sufficient. Therefore an automatic and rapid way of diagnosis is needed. X-rays are radiation waves or electromagnetic waves that help to create images of the internal body and organs with different shades and levels of a gray color. The level of gray color results from the corresponding absorption of radiations by molecules, for example, calcium in bones, absorbs x-ray waves the most, which makes the color more near to the white color than to the black or gray. The air has very leak absorption, so lungs for example look more near to the black color. According to the recent works, many diseases could be diagnosed automatically. For example, the brain tumor segmentation task can be achieved by convolutional neural networks applied for 3D MRI images [1]. For predicting patients with type 2 diabetes mellitus with the help of machine learning algorithms, in [2] four algorithms: decision tree, K-nearest neighbors, artificial neural network, and Deep Neural Network were applied and evaluated. Heart diseases such as atherosclerosis [3] could be early diagnosed to prevent many health complications, using artificial Neural Network (ANN) and KNearest Neighbor (KNN). Convolutional Neural Networks based algorithms could help us detect abnormal lungs and diagnose pneumonia [4] based on processing and analyzing thoracic X-ray images. The X-ray radiation can be used to diagnose chest diseases like pneumonia and recently COVID-19 disease, by exposing the patients to a normal amount of radiation which does not put the patients at a big risk of radiation. At this end, most parts of a normal lung look black because of the excessive presence of the air, whiles the infected parts of lung are shown more near to the white color because of the leak of the air particles and the presence of other types of tissues and pus. CT images can be exploited for COVID-19 diagnosis [5, 6], but the existence of this technology is limited in hospitals. In the proposed method, the diagnosis is automatic using the power of convolutional neural networks on image processing. This method is very useful which presents good results and the probability of predictions compared to the manual way [7]. It consists of deep analyzing chest x-ray images of patients to classify diseases and differentiate normal lungs from abnormal lung [8, 9]. For the abnormal lungs, our proposed method 1 https://www.worldometers.info/coronavirus.

150

H. Moujahid et al.

can distinguish and classify traditional pneumonia and COVID-19 disease. However, the convolution aspect of CNN networks makes this methodology efficient in dealing with image processing. Thus, to accelerate and minimize time and resources complexity for tasks with huge amount of data, the use of parallel architecture is recommended [10]. The rest of this paper is organized as follows: In Sect. 2, a review of some potential and recent related works is presented. Section 3 presents all sources used to build the used dataset and describe the proposed methods. In Sect. 4, the finding results and discussion in terms of model performances are included. Section 5 concludes the paper and gives some future perspectives.

2 Related Works The actual plague of the COVID-19 pandemic obliged most researchers to focus their efforts on finding besides a medical treatment a rapid and appropriate way for diagnosing the disease at its early stage. Many attempts have been done lately concerning the exploitation of machine learning algorithms in order to help to diagnose coronavirus disease by analyzing clinical resources. The current section presents some interesting publications related to our work concerning the application of machine learning in the field of COVID-19 problematic. In [11], the authors proposed a convolutional neural network model by combining Xception and ResNet50V2 networks. The proposed model was trained on a dataset of 11302 X-ray images to classify images into three classes: pneumonia, normal, and COVID-19 cases. The used dataset was unbalanced were only 180 samples of COVID-19 against 6054 of pneumonia and 8851 of normal cases. Although the average accuracy of the detecting COVID-19 cases on the validation dataset was 99.51% and 91.4% accuracy for other classes, this work did not test the model on the testing set to get accurate model performances. Another work in [12] concerns a new automatic system of diagnosing COVID19 disease from x-ray findings. The proposed system employs hybrid deep learning techniques in which long short-term memory (LSTM) is concatenated with CNN. The CNN part is used for feature extraction, whilst LSTM is used at the detection phase. Same as the work of [11], the dataset used for training the model was very small. The authors used a dataset of only 421 x-ray images and 141 of them were COVID-19 features, 140 images for normal cases, and 140 for pneumonia cases. The proposed methods achieved 97% accuracy. However, deeply analyzing the finding results, the proposed method suffers from, on one hand, the weakness of the used data augmentation technique which generated data with a big correlation. On the other hand, to evaluate the model, the authors reported the performance of the model only respects the validation set and omitted the testing set. A new CNN based model was designed and described in [13] for detecting COVID-19 in the human body. In that study, the model was trained on a dataset of three classes: normal x-ray images, pneumonia x-ray images, and COVID-19 x-ray images. The authors used a dataset from the Kaggle dataset repository2 . The dataset contains 234 normal 2 https://github.com/ieee8023/covid-chestxray-dataset.

Diagnosis of COVID-19 Disease

151

images, 390 images of pneumonia, and only 94 coronavirus images. The dataset is divided into a training set, a validation set, and a testing set. The best accuracy achieved by the model was 87.4%. However, analyzing deeply the finding results, the model gets a big over-fitting problem due to the weakness of used image preprocessing techniques. In [14], the authors proposed a CNN model as a combination of three models: autoregressive integrated moving average (ARIMA) model, the prophet algorithm, and LSTM. The proposed model used a dataset of only 128 x-ray images including 28 healthy cases and 70 of COVID-19 cases3 . Similarly, the authors used an augmentation technique to generate more data until 1000 x-ray images. Later, the model was trained and tested to predict the area that will be the most infected in the next seven coming days in terms of new cases, recovered cases, and reported deaths. In [15], the study shows clearly the difference between the use of a pre-trained model and the modified CNN model. In the experiment, the authors used two different datasets (CT) images and X-ray images. The results showed that using the pre-trained Alex-Net model yields better results comparing with the modified CNN model with 98% and 94.1% accuracy respectively. A summary of the aforementioned related work methodologies and validation accuracies obtained on training models over different datasets is presented in Table 1. Table 1. Overview of the related works results. Authors/Reference

Used method

Images type

COVID-19 images number

Accuracy

Rahimzadeh et al. [11]

Combined Xception & ResNet50V2

x-ray

180

99.51%

Islam et al. [12]

Concatenated LSTM & CNN

x-ray

141

97%

Gonesh et al. [13]

CNN model

x-ray

94

87.4%

Alazab et al. [14]

Combined LSTM & PA & ARIMA

x-ray

70

99.94%

3 Materials and Methods The next subsections introduce an artificial intelligence concept, the used methodology of convolutional neural networks, and the process of collecting sufficient features from public resources to build a valid dataset. 3.1 Dataset Assembling The lake of publicly available datasets that concerns the new pandemic of COVID19 makes the collection of sufficient data for our work hard and difficult. Especially, 3 https://www.kaggle.com/nabeelsajid917/covid-19-x-ray-10000-images.

152

H. Moujahid et al.

the collected and selected data from public resources need additional filtering and preprocessing. The final dataset contains three classes of images: normal cases, pneumonia cases, and COVID-19 cases. Figure 1 shows a sample of each class.

Fig. 1. Example of samples from the used classes in dataset: (a) Normal case, (b) Pneumonia case, (c) COVID-19 case.

Our work requires a dataset of chest x-ray findings. For normal and pneumonia images, a COVID-19 Radiography Database of chest x-ray images collected by a team of researchers from many universities was used. The database contains 1341 normal x-ray images and 1345 pneumonia x-ray images. Also, the dataset contains also 219 COVID-19 positive x-ray images that will be combined with other images to build the final dataset as described in Table 2. Table 2. Different sources of COVID-19 datasets. Dataset

COVID-19 images

Valid images

Joseph Paul Cohen (ieee8023)a

661

456

Covid-19 chest x-ray dataset initiativeb

56

35

ACTUALMED COVID-19 chest x-ray datasetc

239

58

COVID-19 Radiography Databased

224

224

76

68

Dataset-01 Chest x-rayse

a https://github.com/ieee8023/covid-chestxray-dataset. b https://github.com/ieee8023/covid-chestxray-dataset. c https://github.com/agchung/Actualmed-COVID-chestxray-dataset. d https://kaggle.com/tawsifurrahman/covid19-radiography-database. e https://github.com/zeeshannisar/COVID-19.

In terms of valid chest x-ray images, 840 images were collected from the literature. Then, the dataset is divided into three sub-datasets as follows: 70% for the training dataset, 15% for the validation dataset, and 15% for the testing dataset. A brief description of sample distribution is shown in Table 3.

Diagnosis of COVID-19 Disease

153

Table 3. Description of dataset distribution SUBSETS

NORMAL

PNEUMONIA

COVID-19

Training set

939

941

588

Validation set

201

202

126

201

202

126

1341

1345

840

Testing set Total

3.2 Proposed CNN Based Methodology CNN, as earlier stated, is a deep learning model composed of several artificial neurons related to each other mathematically by specific functions. CNN is usually used intensively for extracting distinctive features from visual images. A CNN network is based on mathematical convolution operation applied in at least one convolutional layer [16]. For complex-valued functions f, g defined on Z ensemble, the discrete convolution of f and g is given as follows: (f ∗ g)[n] =

m=+∞ 

f [m].g[m − n]

(1)

m=−∞

A traditional CNN network consists of an input layer related to multiple hidden layers and an output layer called a classifier. The hidden layers are a combination of convolutional layers, pooling layers, and fully connected layers. In a CNN architecture, the input is a tensor with shape depending on the input image dimension. After the tensor passes through a convolutional neural network, an abstraction of the image happens to generate a feature map with a defined shape. When creating a convolutional layer, some hyper-parameters are tuned, e.g., a kernel with a specific depth and height, the number of input channels and output channels, the depth of convolution filter, and finally the activation function. Each layer output is transformed into an input for the next layer which makes shapes of inputs and outputs are highly correlated. The general architecture for a convolutional neural network is presented in Fig. 2, showing the input layer, two hidden layers, and the output layer.

Fig. 2. Convolutional neural network architecture.

154

H. Moujahid et al.

Convolutional Layers Inside the convolutional layer, each neuron is connected to only a subset of neurons connected spatially in the layer before and the weights of these connections are shared with all other neurons in the convolution layer [17, 18]. The mean utility of the convolutional layer is detecting the local features at all positions in the input feature maps with learnable kernels (connection weights between the feature map i at the layer n − 1 and the feature map j at the layer n). Pooling Layers Pooling operation consists of sliding a 2-dimensional filter across the feature map resulted from the previous convolutional layer, then summarizing the output features. Generally, the mean goal of the pooling layer is to reduce the feature size and preserve only the important information and release the rest. There are several types of pooling operations among them the max pooling, average pooling, and global pooling are the most commonly used in CNN networks [17, 19]. The typically max-pooling operation is presented in Fig. 3. The output element y of a pooling layer is defined as follows: y = maxxij i,jR

(2)

where xij represents an element covered by the filter R.

Fig. 3. The max-pooling operation.

Fully Connected Layers (FC) Fully connected layers (are also called dense layers) are equivalent to the convolutional layers with the difference that all units in a fully connected layer are connected to every unit in the next layer as well as those neurons at the previous layer. Those layers are activated by an activation function generally a rectifier linear unit (ReLU) [18]. Transfer Learning (TL) The trained convolutional neural network model can be transferred to be exploited for another prediction task [20]. Such a process is known as the transfer learning process. There are many derivatives of transfer learning depending on the task to process. The used one in this paper is a zero-shot transfer learning, where all layers must be retrained

Diagnosis of COVID-19 Disease

155

except the output layer. Totally, the COVID-19 and pneumonia classification task is different from the original task of the pre-trained CNN model. In this case, a large dataset and computing power are needed as well. Also, some layers at the end must be added and specified the appropriate classifier. Every machine learning model must be tested and evaluated before exploitation in the real-world task. This can be done by calculating evaluation metrics. However, there are many metrics to evaluate a model. In the experimental part, three models are trained by using transfer learning approach using architectures of VGG-16, VGG-19, and MobileNetV2.

4 Results and Discussion A CNN model presents different results depending on the depth of the network, type of layers, hyper-parameters and its architecture. In this paper, the VGG-16, VGG-19, and MobileNetV2 architectures are used as a base of the proposed model. 4.1 Training Results As shown in Fig. 4, while training, both the VGG-16 model and the VGG-19 model continue to be improved in a linear manner in terms of validation accuracy and loss until the 10th epoch. After that, the training performance begins to stabilize and the improvement becomes slow. On the opposite, the MobileNetV2-based model achieves good training improvements until the 13th epoch, and then becomes slower. The main problem of this model is the divergence of the validation performances from the training. To avoid the over-fitting problem, a callback parameter is adopted and specified to optimize the number of epochs needed to obtain the maximum possible accuracy. After that, the models are trained with the dataset described in Table 3. The overall accuracy and loss results are presented in Fig. 4 and Fig. 5, showing the variation of those metrics across epochs. In addition to that, the three models showed different convergence levels. The VGG16 based model is the fastest in terms of convergence. It needs 25 epochs to get maximum accuracy value. Whilst VGG-19 based model needs only18 epochs, and MobileNetV2 based model needs 47 epochs. 4.2 Testing Results The training step generates a ready model for testing. In this experiment, the three models were tested on a totally an independent dataset from the ones used for training and validation. Then, according to the obtained testing results, the models are evaluated with specific metrics as shown in Table 4. VGG-19 and VGG-16 showed the best accuracy value of almost 97% against MobileNetV2 that achieved an accuracy of 95.84%. In terms of COVID-19 classification, VGG-19 and VGG-16 make a 100% prediction correctly.

156

H. Moujahid et al.

Fig. 4. Accuracy metric of trained models: (a) VGG-16, (b) VGG-19 and (c) for MobileNet.

Fig. 5. Loss metric of trained models: (a) VGG-16, (b) VGG-19 and (c) for MobileNet

Diagnosis of COVID-19 Disease

157

Table 4. Detailed metrics values for each trained model Metric Precision Recall F1-score

VGG-16 VGG-19 MobileNetV2 1.00 0.97 0.98

1.00

0.95

0.99

1.00

1.00

0.98

Accuracy 96.22%

96.97% 95.84%

Loss

17.40%

17.33%

14.66%

4.3 Comparison with Related Works Comparing the finding results of our proposed model with those cited in the existing methods listed previously in the related work section, our model achieved a 100% precision and 100% F1-score. In terms of accuracy, our proposed model shows a slightly good accuracy respecting different datasets. More details are shown in Table 5. Table 5. Comparison of our results with other similar works Method

Recall

Precision

F1-score

Accuracy

Combined Xception & ResNet50V2 [11]

80.53%

35.27%

NA

99.51%

Concatenated LSTM & CNN [12]

100%

NA

100%

97%

CNN model [13]

NA

NA

NA

87.4%

Combined LSTM & PA & ARIMA [14]

NA

NA

NA

99.94%

This work: VGG-19 based TL model

99%

100%

100%

96.97%

5 Conclusion and Perspectives Through this paper, a COVID-19 detection methodology was reported by personalizing and retraining three CNN models (VGG-19, VGG6, and MobileNetV2). A specific dataset of 3526 X-ray images was generated with the help of many sources. It contains 840 COVID-19 cases, 1345 images of pneumonia cases and 1341 images of normal cases. After testing the models on a test set containing 30% of the original dataset, a deep analysis of results and performance of the models was performed based on some essential metrics (Recall, Precision, F1-score, and Accuracy). The best result on VGG-19 shows 99% on recall, 100% for precision, and 100% for f1-score. The results prove also the importance of using CNN architecture for predicting COVID-19 disease based on X-ray images. The behavior of a convolutional neural network towards a classification task is unpredictable. This conducts us to think that there is no dedicated model for each field of study.

158

H. Moujahid et al.

Therefore, our vision in the future is to train and test as much as possible of models. However, there are so many tasks to process in the medical field using deep learning methodologies. The availability and accuracy of the dataset in the specific field is an important factor that must be taken into consideration. Acknowledgements. This work is a part of a project supported by co-financing from the CNRST (Centre National pour la Recherche Scientifique et Technique) and the Hassan II University of Casablanca, Morocco. The project is selected in the context of a call for projects entitled “Scientific and Technological Research Support Program in Link with COVID-19” launched in April 2020 (Reference: Letter to the Director of “Ecole Normale Supérieure de l’Enseignement Technique de Mohammedia” dated 10 June 2020).

References 1. Moujahid, H., Cherradi, B., Bahatti, L.: Convolutional neural networks for multimodal brain mri images segmentation: a comparative study, pp. 329–338 (2020) 2. Daanouni, O., Cherradi, B., Tmiri, A.: Type 2 diabetes mellitus prediction model based on machine learning approach. In: The Proceedings of the Third International Conference on Smart City Applications, pp. 454–469 (2019) 3. Terrada, O., Cherradi, B., Raihani, A., Bouattane, O.: Classification and Prediction of atherosclerosis diseases using machine learning algorithms. In: 2019 5th International Conference on Optimization and Applications (ICOA), pp. 1–5 (2019) 4. Moujahid, H., Cherradi, B., Gannour, L., Bahatti, O.T., Hamida, S.: Convolutional Neural Network Based Classification of Patients with Pneumonia using X-ray Lung Images, vol. 5, no. 5, p. 9 (2020) 5. Singh, D., Kumar, V., Vaishali, Kaur, M.: Classification of COVID-19 patients from chest CT images using multi-objective differential evolution–based convolutional neural networks. Eur. J. Clin. Microbiol. Infect. Dis. 39(7), 1379–1389 (2020). https://doi.org/10.1007/s10096020-03901-z. 6. Wang, S., et al.: A deep learning algorithm using CT images to screen for Corona Virus Disease (COVID-19), medRxiv, p. 2020.02.14.20023028, April 2020. https://doi.org/10.1101/2020. 02.14.20023028. 7. Zhang, Q., Liu, Y., Liu, G., Zhao, G., Qu, Z., Yang, W.: An automatic diagnostic system based on deep learning, to diagnose hyperlipidemia. Diabetes Metab. Syndr. Obes. Targets Ther. 12, 637–645 (2019). https://doi.org/10.2147/DMSO.S198547 8. Heidari, M., Mirniaharikandehei, S., Khuzani, A.Z., Danala, G., Qiu, Y., Zheng, B.: Improving the performance of CNN to predict the likelihood of COVID-19 using chest X-ray images with preprocessing algorithms. Int. J. Med. Inf. 144, 104284 (2020). https://doi.org/10.1016/ j.ijmedinf.2020.104284 9. Ozturk, T., Talo, M., Yildirim, E.A., Baloglu, U.B., Yildirim, O., Rajendra Acharya, U.: Automated detection of COVID-19 cases using deep neural networks with X-ray images. Comput. Biol. Med. 121, 103792 (2020). https://doi.org/10.1016/j.compbiomed.2020.103792 10. Bouattane, O., Cherradi, B., Youssfi, M., Bensalah, M.O.: Parallel c-means algorithm for image segmentation on a reconfigurable mesh computer. Parallel Comput. 37(4), 230–243 (2011). https://doi.org/10.1016/j.parco.2011.03.001 11. Rahimzadeh, M., Attar, A.: A modified deep convolutional neural network for detecting COVID-19 and pneumonia from chest X-ray images based on the concatenation of Xception and ResNet50V2. Inform. Med. Unlocked 19, 100360 (2020). https://doi.org/10.1016/j.imu. 2020.100360

Diagnosis of COVID-19 Disease

159

12. Islam, M., Islam, M., Asraf, A.: A Combined Deep CNN-LSTM Network for the Detection of Novel Coronavirus (COVID-19) Using X-ray Images (2020) 13. Gonesh, C., Ganie, I., Rajendran, G., Nathalia, D.: CNN Analysis for the detection of SARSCoV-2 in Human Body, pp. 2369–2374, June 2020 14. Alazab, M., Awajan, A., Mesleh, A., Abraham, A., Jatana, V., Alhyari, S.: COVID-19 Prediction and detection using deep learning. Int. J. Comput. Inf. Syst. Ind. Manag. Appl. 12, 168–181 (2020) 15. Maghdid, H.S., Asaad, A.T., Ghafoor, K.Z., Sadiq, A.S., Khan, M.K.: Diagnosing COVID19 pneumonia from X-Ray and CT Images using Deep Learning and Transfer Learning Algorithms, p. 8. 16. Tajbakhsh, N., Shin, J.Y., Hurst, R.T., Kendall, C.B., Liang, J.: Chapter 5 - automatic interpretation of carotid intima–media thickness videos using convolutional neural networks In: Zhou, S.K., Greenspan, H., Shen, D. (eds.) Deep Learning for Medical Image Analysis, pp. 105–131. Academic Press (2017) 17. Wanda, P., Jie, H.: RunPool: a dynamic pooling layer for convolution neural network. Int. J. Comput. Intell. Syst. 13, January 2020. https://doi.org/10.2991/ijcis.d.200120.002. 18. Srinivas, S., Sarvadevabhatla, R.K., Mopuri, K.R., Prabhu, N., Kruthiventi, S.S.S., Babu, R.V.: Chapter 2 - an introduction to deep convolutional neural nets for computer vision. In: Zhou, S.K., Greenspan, H., Shen, D. (eds.) Deep Learning for Medical Image Analysis, pp. 25–52. Academic Press (2017) 19. Alsaeedi, A., Al-Sarem, M.: Detecting rumors on social media based on a CNN deep learning technique. Arab. J. Sci. Eng. (2020). https://doi.org/10.1007/s13369-020-04839-2 20. Sewak, M., Karim, M.R., Pujari, P.: Practical Convolutional Neural Networks: Implement Advanced Deep Learning Models Using Python. Packt Publishing Ltd. (2018)

Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques Tariq Saeed Mian(B) Department of IS, College of Computer Science and Engineering, Taibah University, Madinah Almunwarah, Saudi Arabia [email protected]

Abstract. Correct and early diagnosing Parkinson’s Disease (PD) is vital as it enables the patient to receive the proper treatment as required for the current stage of the disease. Early diagnosis is crucial, as certain treatments, such as levodopa and carbidopa, have been proven to be more effective if given in the early stages of PD. At present the diagnosis of PD is solely based on the clinical assessment of a patient’s motor symptoms. By this stage however, PD has developed to such an extent that irreversible neurological damage has already occurred, meaning the patient has no chance of recovering. By implementing the use of machine learning into the process of assessing a potential PD patient the disease can be detected and diagnosed at a much earlier stage, allowing for swift intervention, which increases the chance of PD not developing to such damaging levels in the patient. Machine Learning is a subfield of artificial intelligence that provides different technique to scientists, clinicians and patients to address and detect diseases like PD at early stage. The main symptom of PD is the vocal impairment that distinguishes from the normal person. In this study, we used a PD vocal based dataset that has 755 features The Principal Component Analysis (PCA) and Linear Discriminate Analysis (LDA) techniques are used to reduce the dimensionality of the available Parkinson’s dataset to 8 optimal features. The study used four supervised machine learning algorithms, two algorithms are from the ensemble techniques, Random Forest, Adaboost Support Vector Machine and Logistic Regression. The Random Forest model with LDA and PCA shows the highest accuracy of 0.948% and 0.840% respectively. Keywords: Parkinson’s disease · Early detection · Machine learning · Linear Discriminate Analysis · Dimensionality reduction · Principal Component Analysis · Ensemble methods · Random forest · Adaboost Support Vector Machine · Logistic regression

1 Introduction Bioinformatics have been widely used in diagnosis and detection of fatal neural diseases in recent times. Machine Learning is the sub-field of artificial intelligence that is being utilized in Parkinson’s disease diagnosis. This disease is mostly found in people over © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 160–175, 2021. https://doi.org/10.1007/978-3-030-70713-2_17

Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques

161

65 years of age [1]. However, the symptom of PD disease arises during the age of 30– 50. PD is a chronic, neurodegenerative disease that majorly affects the motor system of the body. The PD progression rate is very slow; however, as the condition worsens it stimulates the non-motor symptoms, such as hallucinations, mood disorders, delusions and other cognitive and behavioral changes. The environment and genetic factors have a significant influence on the risk of developing Parkinson’s; however, the origin of disease remains unknown [2]. PD is caused due to loss of brain cells by the degeneration of dopamine producing neuron cells. Dopamine is chemical that is generated from substantial nigra that is responsible for transfer of signals within the brain. As the production of dopamine is reduced, movement disorders begin to develop in the patient. The Parkinson’s symptoms can be divided into motor and non-motor symptoms. Motor symptoms are movement related problems, and these are more perceptible to non-motor symptoms [3]. The patient has issue of rigidity, tremor and slower movement. Non-motor symptoms of PD are speech problem, sleep disorder and olfactory issues. The general symptoms that PD patient face is difficulty in walking, mental disorders and shaking in movement (tremors). PD patients also suffer from depression and anxiety. AS the condition of a PD patient worsens then dementia develops. In regard to gender, this disease is more common in males as compared to females [4]. There is no proper treatment of Parkinson’s disease in the market. Available treatments only help control the symptoms rather than treat them. There are different methods for diagnosis of PD that are adopted by different practitioners. Physicians prefer medical history and neurological examination to assess the condition of PD patients [5]. Medical imaging, such as Magnetic Resonance Imaging (MRI) and Computer Tomography (CT), are also useful for the diagnosis of PD, with MRI being the more effective of the two methods [6]. The diffusion MRI technique is effective in distinguishing between additional PD syndrome and PD [7]. As information technology is making progress and new computational systems are being introduced, More and more clinicians are interested in using an intelligent model to improve not only the accuracy of diagnosis, but also the quality of diagnosis. In recent years due to easy access to storage and communication tools enormous data has become available for domestic and industrial usage. The examples of applications that are providing further research are flight simulators, weather forecasting and earth simulators [8]. The cognitive ability can be achieved through a wide range of computer science branches like Artificial Intelligence, Machine Learning, Natural Language Processing and Computer Vision [9]. Now the signs of Parkinson can be identified through smart phone applications. The aim of this application is to detect PD and monitor the progress rate. Most of PD diagnosis model focus on clear symptoms that can be easily identified through medical equipment. PD can be detected by the vocal patterns of the user. The vocal disorder of PD individual can be identified in the early stages of the disease. The large set of available data about PD and clinical need for an intelligent system gave rise to develop a computational model that can be used for the early detection of PD [10]. The vocal features are passed to machine learning models to determined potential insight from the data. The automatic classification of PD is based on its severity. PD becomes life threatening due to late stage diagnosis. The early stage diagnosis of PD increases the chance of the patient’s condition not deteriorating further and becoming more severe. The researchers have used different

162

T. S. Mian

speech signal processing techniques to get the clinical like features of PD and then these mined features are passed on to machine learning-based models to classify the disease. Support Vector Machine (SVM) [11], Artificial Neural Network [12], Random forest [13] and K Nearest Neighbor (KNN) [14] are the more commonly used algorithms to classify PD. The above-mentioned algorithms use the feature selection technique and take optimal number of features as input. The vocal data have intrinsic properties and manual selection of features is a difficult task. There is need of a more novel approach to diagnose PD that should be simplified, less expensive and more reliable. In this study we used the vocal based feature dataset that is publicly available on the Kaggle website. This dataset is the property of the University of California, Irvine Machine Learning. The dataset consists of 754 different attributes. We used Principal Component Analysis and Linear Discriminate Analysis to reduce the dimensionality of the proposed dataset. We used only 8 features as input to the four supervised machine learning algorithms. We used two ensemble techniques; Random Forest and Adaboost. Like other health studies the datasets used on paper are also imbalanced, which means the number of one class instances is larger in distribution than the number of other class instances. The imbalance data set impact the classification performance of the ML algorithm due to its biasness towards the majority class. As the dataset is imbalanced, the accuracy may be misleading in measuring and predicting, and most outcomes may of the majority class. In order to check evaluation of proposed models, we used accuracy, confusion matrix, Precision, Recall, F1-Score and Receiver Operator Characteristics (ROC) as the performance evaluation metrics. The contribution of the proposed approach is: • In the proposed approach, we used an unsupervised approach to reduce the dimensionality of the data. PCA and LDA shows better results than simple used feature selection techniques. • We used two ensemble-based machine learning and two simple machine learning algorithms. We prove that ensemble model provides better results than simple supervised machine learning algorithms. • We used Accuracy, Precision, recall, F1-score to investigate the performance of LDA and PCA on ML algorithms. The paper is organized as follow; in Sect. 2, provides a literature review, Sect. 3 discuss and explain the methodological issues, Sect. 4 presents the proposed algorithms predictions through tables and graphs and Sect. 5 is devoted to conclusion and future direction.

2 Literature Review In this section we are going to discuss existing machine learning techniques used to diagnose Parkinson’s disease. Our main focus is to discuss intelligent methods powered by Machine learning and deep learning for the classification of PD. The author et al. [15] discuss about the fed-forward neural network used for Parkinson’s prediction. In this study, the prediction error of the model is discussed. Neural network output is a

Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques

163

project rule-based system. The unlearnt data is gathered separately during the process of learning the models and fed to the model in next batch of training. This approach also performs well on imbalanced datasets. Data Mining techniques are mostly used on structure data to make predictions. Three different techniques of data mining, tress based, statistical and support vector machine, are used to classify the effected individuals. The prediction is measured in term of accuracy in the data mining approaches [16]. In this study [17], Artificial Neural Network (ANN) and SVM are used to classify the effected individuals of PD. These approaches are helpful for medical practitioner to diagnose PD suffering individual at lower costs. There are different machine learning algorithms that are applied to vocal recordings of PD patients and make the decision boundary between target variable instance classes. The ensemble algorithm random forest model outperforms with candidate feature selection by using minimum Redundancy Maximum Relevance(mPMR) feature selection technique over the benchmark models [18]. The author et al. [19] presented a particular classification and prediction approach for Parkinson diagnosis. In this study, data preprocessing, cross validation and Machine Learning algorithms are used to find the hidden pattern from the data. The tremor data features and neuro data features were analyzed for symptoms prediction. Machine Learning provides very good results, yet have some shortcomings in PD detection and sensitivity rate. Mostafa et al. [20] proposed multi-agent data analysis technique. This technique evaluates vocal disorders. The effected individual vocal records were considered as important features for disease detection. Reinforcement Learning, Naïve Bayes, Random Forest and Decision Tree models were used to analyze the vocal variations. The dataset used in this study were collected from Tel Aviv Sourasky Medical Centre. However, the work was lagging with real-time issues. The author et al. [21] presented the novel approach for detection of PD symptoms using vocal based dataset. Naïve Bayes and SVM were used on vocal dataset to make the prediction. The proposed approach was attempted to predict the more accurate result yet this work has some limitation in dataset features. The author et al. [22] proposed information gain analysis technique for PD detection from benchmark datasets. In this study, different Machine Learning and Information gain techniques were combined for the detection of PD. This strategy has good result in PD diagnosis yet produced insignificant results compared to deep learning techniques used for PD diagnosis. Seppi et al. [23] presented the new technique for Parkinson treatment using non-motor symptoms. The proposed approach provides information and updates the next level treatment method for the future. The work was a collection from different treatment evidences and provided valuable suggestions. The PD classification results depend on the feature selection and artificial learning methods. In research, many researchers have used publicly available dataset [24] that consist of 31 instances and 195 sound recordings. Parisi et al. [25] proposed hybrid intelligence-based classifier for PD diagnosis. The dataset used in this paper was the property of University of California Irvine ML repository. They used MLP with custom cost function and were trained on the training instance. Then hybrid MLP_LSVM and predict the diagnosis of PD and get 100% accuracy rate. The author et al. [26] presented a new technique for PD detection with vocal features. They used different feature selection to filter 10 optimal features with high relevance score. They used feature selection technique such as Least

164

T. S. Mian

Absolute Shrinkage and Selection Operator, Minimum Redundancy Maximum Relevance (mRmR). These optimal features are then passed to the Random Forest model and Support Vector Machine. These models show a precision rate of 98% with following features: shimmer, HNR and vocal fold excitation. The author et al. [27] presented a novel approach for detection of PD using the vocal features. The input features were, jitter, shimmer, pitch and HNR. A different selection technique was used to get the high rank features such as Fisher’s Discriminant ratio, correlation rates, t-test and ROC. The optimal features were defined through the wrapper method that used support Vector machine model to project feature performance curve. In this study, KNN, SVM and discrimination-based classifiers were used with optimal features. In order to validate the performance of the algorithms, accuracy, error rate, sensitivity and specificity, metrics were used. The KNN model has the highest accuracy score with 93.82%. The author et al. [28] used advance method for clinical treatment of PD and his proposed study presented that supportive care, including rehabilitative and physical interventions, nursing care and speech therapy, are main process for improvement in the recovery of PD. The author et al. [29] proposed the SVM model with Gaussian Radical basis kernel for PD diagnosis. The dataset used in study was taken from UCI machine learning repository. The author et al. [30] used the non-linear model for classification of PD that is based on Dirichlet mixtures. The author et al. [31] used the feature selection technique mutual information gain with SVM. This technique obtained high classification accuracy yet tele-diagnosis of PD needs a better method with higher classification performance.

3 Methodology 3.1 Data Set In this proposed study, vocal based dataset is used that contains healthy and affected individual’s vocal recording instances. This dataset is accessed from California University, Irvine Machine Learning. This dataset has 188 patient records in which there are 81 female and 107 male participants. The participants’ individual age group ranges from 33 to 87. The healthy group has 64 samples with 23 male and 41 female individuals with age range of 41 and 82. The final version of the dataset contains 756 instances and 754 attributes [33, 34] (Table 1). Table 1. Dataset description Detail

Source information

Dataset property

University of California, Irvine Machine Learning

Dataset name

Parkinson’s Disease

Dataset attributes 754 Dataset records

756

Target variable

(0-control, 1-PD). Binary Class Problem

Task

Binary classification

Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques

165

We also analyze that class distribution of dataset is skewed, which means our dataset is imbalanced. The classification accuracy will be tending toward the majority class. To handle the issue of imbalanced dataset, we used an upsampling technique in which class distribution 0 and 1 are equal. The upsampling increase the distribution in the minority class and make an equal distribution of both classes. The Fig. 1 shows that class distribution of the target value without upsampling.

Fig. 1. Class distribution of the target value without upsampling

3.2 Proposed Model In the proposed approach, we first perform the preprocessing of the vocal dataset, explore the duplicate values, get statistical information from the dataset, do exploratory data analysis to get more insight information from the hidden pattern of the dataset. The dataset contain the highly correlated features. We set the value of thresh at 80% and remove the correlated feature that have strong correlation of more than 80%. We then split the dataset into training and test with the ratio of 70:30 respectively. A five fold cross validation is used to test the generalizability of the models and increase the accuracy of the proposed models. The dataset has large number of attributes. We performed the dimensioanlity reduction technique PCA and LDA to use the important dimension. We then implement supervised machine learning algorithm and check the performance evaluations of proposed algorithms in term of accuracy, precision, recall, f-score and AUC (Fig. 2). 3.3 Dimensionality Reduction Dimensionality reduction is the process of converting high dimensional data into low dimensional data. In this paper, we used two technique of dimensionality reduction, Linear Discriminant Analysis (LDA) and Principal Component Analysis (PCA). Linear Discriminant Analysis (LDA) LDA is prevalent dimensionality reduction technique for reducing the dimension of data in machine learning and data mining applications [39]. The goal of LDA is to project the large number of features into a reduced number of features with good class-separability and reduce computational costs. LDA maximize the variance of data and also perform maximization into separation of multiple classes. The main purpose of LDA is to project a dimension space into a reduced subspace i (where i < x−1) without losing the class information.

166

T. S. Mian

Fig. 2. Proposed algorithm

The steps of LDA are: • A d-dimensional mean vector is calculated for every class of dataset. • Computation of scatter matrices is performed. • The eigenvectors (E1, E2, E3…. Ed) and their related eigenvalues (ψ1, ψ2, ψ3,…. ψd) of the scatter matrices are calculated. • We sort the eigenvectors in descending order of eigenvalues and then opt for k eigenvectors which have maximum eigenvalues in order to form a d * I matrix WW. • We used d * I matrix of eigenvector for transformation of input samples into a new subspace • YY = XX * WW. Principal Component Analysis (PCA) PCA is a statistical technique that transforms the data orthogonal. PCA transformed a group of correlated features to uncorrelated group of features [38]. The basic function of PCA is to reduce the dimensionality of data and perform exploratory data analysis. PCA can be used to determine the relationship among variables. Let we have the dataset having features X = (x1, x2, x3…………xn) where n denotes input dimension. We can reduce the n-dimension data into k-dimension (k < n) by using PCA. Assume that we have raw data with unit variance and zero mean. j

j

xi =

xi − xj σj

We can calculate the co-variance matrix of the raw data.   1 m (xi )(xi )T , ∈ Rnxn = 1 m

(1)

(2)

We can compute the eigenvalue and eigenvector of the co-variance matrix. μT = λμ

(3)

Early Diagnosos of Parkinson’s Using Dimensionality Reduction Techniques

167

⎡

⎤ − − − μ = ⎣ μ1 μ2 μ3 ⎦, μi ∈ Rn − − −

(4)

We projected the row data into k-dimensional subspace and then choose the top k eigenvector from co-variance matrix. This matrix will be new from the original dataset. The PCA is method of converting the raw data of n-dimensionality into a reduce k dimensional representation of the data. ⎡

xinew

⎤ μT1 xi = ⎣ μT2 xi ⎦ ∈ Rk μTk xi

(5)

3.4 Machine Learning Algorithms Random Forest Random Forest model is an ensemble technique that is based on decision tree and use the set of splitting rules to build model that predict the value of target variable. Random Forest improves classification accuracy of singletree classifier by adding randomization and bootstrap aggregating method in the selection of data nodes during the decision tree construction [78]. A decision tree with M leaves divides the feature space into M regions Rm, 1 ≤ m ≤ M. The prediction function f(x) for each tree can be defined as: f (x) =

M m=1

cm



(x, Rm )

(6)

In Eq. (6) M denotes number of regions in the feature space, Rm is a region suitable to m, Cm is a constant to m.   1, if x ∈ Rm (7) (x, Rm ) = 0, Otherwise K-Nearest Neighbor K-Nearest Neighbor (KNN) model was introduced by Fix Hodges in 1951. KNN is simple distance based powerful, non-parametric lazy learning algorithms. KNN can also be for classification and regression tasks. It stores all available cased and divide into new cased based similarity score. KNN has been used in pattern recognition and statistical estimation before 1970. KNN takes n number of training instance and q as an unknown value. 1. Training samples are stored in any array of data points arr[] in which every element of the array shows a tuple(a, b) 2. For i = 1 to n, then compute Euclidean distance d(arr[i], q)

168

T. S. Mian

3. Obtained the smallest set of K where obtained distances correspond to dataset target variable. 4. Return majority label class from small set of S. Logistic Regression Logistic Regression was present by David Cox in 1958. Logistic Regression model is capable to solve only classification task that use the probability value of 0.5 to predict the target. Logistic Regression model can use observed numerical and categorical values. Logistic Regression optimal decisions are based on the posterior clasds probabilities p(y|x). We can represent outcome of logistic regression in case of binary classification as follows: y = 1 if log p(y = 1|X )p(y = 0|X ) > 0

(8)

or y = 0 if log

p(y = 1|x) 120 mg/dl (1 = true; 0 = false)

restecg

Resting electrocardiographic results; (Value 0: normal, Value 1: having ST-T wave abnormality (T wave inversions and/or ST elevation or depression of > 0.05 mV), Value 2: showing probable or definite left ventricular hypertrophy by Estes’ criteria

thalach

Maximum heart rate achieved

exang

Exercise induced angina; (1 = yes; 0 = no)

oldpeak

ST depression induced by exercise relative to rest

slope

the slope of the peak exercise ST segment; (Value 1: upsloping, Value 2: flat, Value 3: downsloping)

ca

number of major vessels; (0–3) colored by flourosopy

thal

3 = normal; 6 = fixed defect; 7 = reversable defect

num

diagnosis of heart disease (angiographic disease status); (Value 0: < 50% diameter narrowing, Value 1: > 50% diameter narrowing)

Hardware and Software The experiments conducted in this paper were implemented using Python programming language using the Jupyter notebook from Anaconda. It was run on an Intel® Core™ i7 CPU 2.60 GHz with 8.00 GB of RAM under 64-bit Windows 10 Enterprise operating system.

180

F. Kashif and U. K. Yusof

3.2 Classification Algorithms Random Forest The base learners in this ensemble model are the individual decision trees that comprise together to form a forest, hence the name random forest [12]. Random forest can be applied to both regression and classification problems. This method makes use of the bootstrap aggregating or bagging for the learning the data. For a training set X = x1 , x2 , . . . , xn with responses Y = y1 , y2 , . . . , yn bagging is done repeatedly for B times wherein a random sample is drawn with replacement [12]. For each of these samples a classification or regression tree is trained. From these trained models, the outcomes for unseen samples is made by taking average for regression and majority for  the classification tasks. The predictions for random forests for an unseen sample x in regression and classification tasks are shown in Eqs. 1 and 2 respectively.   1 B fb x b=1 B   f = MajorityVote{fb x }B1 

f =

(1)



(2)

Bagging methods work best when the base learners are not stable or there is a need to improve the accuracy of the model [11]. Generally random forests prove more stable in classification tasks as opposed to regression tasks. In this paper, random forest algorithm is employed on 500 base learners to achieve a higher classification accuracy as was achieved by decision tree alone. Support Vector Machine SVM is a classification algorithm that has a very high accuracy. It is primarily used for dichotomous response variables, such as binary or logical [13]. This algorithm provides linear and non-linear separators to find a better fit for the data. One of the main drawbacks is that this procedure is highly sensitive to noise and a little bias can dramatically affect the overall performance. SVM is a discriminant-based method, so in classification tasks, evaluating posterior probabilities is not the priority, rather the need is to estimate the decision boundaries [12]. For a binary classification problem with class labels −1 and + 1, the sample X = {xt , r t } where r t = +1 if xt ∈ C1 and r t = −1 if xt ∈ C2 , the classification in SVM follows the principles of weights associated with each class. The weights w and w0 are found such that: wT xt + w0 ≥ +1forr t = +1

(3)

wT xt + w0 ≤ −1forr t = −1

(4)

Theoretically there can be multiple hyperplanes that are able to separate the classes, but the optimal hyperplane is the one that maximizes the margin, which is the distance between the two closest data points of different classes [12]. SVM are powerful algorithms even for data that cannot be separated by a straight line. Transforming the data

Detection of Cardiovascular Disease Using Ensemble Machine Learning Techniques

181

into a higher dimension can separate the data points but in practice these transformations can get overly complicated. The kernel trick is a function that takes vectors from the original space as input and returns their dot product as the feature space. Mathematically, if x1 , x2 ∈ X and a map ∅ : X → RN , then x = (x1 , x2 ) → ∅(x) = {∅1 (x), ∅2 (x)}

(5)

There are several kernel functions one can choose to handle the data, however in this paper the radial basis function kernel is used in the proposed ensemble model. For two samples x1 , x2 in the original space, the RBF kernel is defined as   x1 −x2 2 (6) K(x1 , x2 ) = exp − 2σ 2 where σ is a free parameter: which cannot be accurately predicted and is not controlled by the model; this parameter can be estimated experimentally. Using this kernel function, the feature space is defined, and the ensemble model is created for this classification task. Ensemble Random Forest and Support Vector Machine Ensemble models use the weak or base learners as their foundation and can be used to generate more accurate results. In this paper an ensemble model is created by using the random forest model and support vector machine. This model is created using the Vote method wherein the model decides based on majority votes from the base learners as to which class to assign to each instance [12]. As described in the previous sections, random forest and SVM are powerful classification techniques and combining them in one model yielded better results than either model produced individually. The models are merged by using the Vote method where an unweighted approach is used i.e. both algorithms are given equal importance while selecting the final class for every new instance. The base learners for this ensemble model are selected based on their individual capabilities to handle classification tasks. These methods are also individually compared with the proposed ensemble model. Hyperparameters of the Proposed Ensemble Model The proposed model is an ensemble using the combined strengths of Random Forest and Support Vector Machine algorithms. The Random Forest model was created using 500 decision trees with a maximum depth of 5. The criterion used is the Gini Impurity which measures the likelihood of a new instance being wrongly classified by the trained model [12]. The hyperparameters of the SVM model were tuned wherein the kernel used was radial basis function. Gamma for SVM was set to “scale” for training this base learner. This parameter of SVM algorithms determines how far the influence of one instance prevails [12]. When this value is set to ‘scale’ the gamma is calculated by using the following formula: Gamma = 1/(NumberofFeatures)(VarianceofX )

(7)

182

F. Kashif and U. K. Yusof

3.3 Evaluation Metrics To build the models, 80% of the data was designated as the training set and 20% was used for testing. To evaluate these models, confusion matrix and the metrics associated with it are employed. Confusion matrix summarizes the actual and the predicted outcomes in a tabular format in terms of false positive, false negative, true positive and true negatives. Ideally the off diagonals i.e. the false positive and negative results should be zero [12]. The evaluation metrics used on this paper are described below: Precision = TP/(TP + FP)

(8)

Recall = TP/(TP + FN )

(9)

Error = (FN + FP)/(TP + FP + FN + TN )

(10)

F − 1 Score = 2(Precision)(Recall)/(Precision + Recall)

(11)

Where T stands for true, F for false, P for positive and N for negative. Precision of a model is the fraction of the predicted positive values that really are positive. Recall corresponds to the fraction of positives that were correctly predicted by the classifier. The threshold is the measure used to define a balance between the precision and the recall [13].

4 Results and Evaluation This research was conducted by using Heart Disease dataset from the UCI repository to predict the class label. The best performance was shown by the ensemble random forest and support vector machine model where the model was able to reach an overall accuracy of 0.89. This model was built by combining the random forest made with 500 decision trees and the SVM algorithm with the radial basis kernel function. The comparison of the base learners with the proposed ensemble model is shown in Table 2 and the detailed analysis of the proposed model is shown in Table 3. We can see from Table 2 that the ensemble model performs better than either of the base individual models. Table 2. Comparison of base learners with proposed ensemble model Algorithm

Accuracy Precision Recall F-1 score

SVM

0.81

0.88

0.85

0.87

Random forest

0.85

0.86

0.84

0.85

Proposed ensemble model 0.89

0.89

0.89

0.88

Detection of Cardiovascular Disease Using Ensemble Machine Learning Techniques

183

Table 3. Class-wise analysis for proposed ensemble model Class label

Precision Recall F-1 score Error rate

0

0.88

0.85

0.87

0.14

1

0.89

0.91

0.90

0.08

Average 0.89

0.89

0.88

0.11

As seen from Table 3 the accuracy, recall and precision of the model are quite close to each other which implies that the model is stable and can predict both classes well. The proposed model has scored a better accuracy as compared to the other algorithms. For further comparisons, some other algorithms were also trained and tested on the same dataset and the comparison was made with the proposed model. The comparison with other algorithms is shown in Table 4. We can see that the proposed model outperforms the rest of the algorithms and can be used to accurately detect heart disease in patients. Table 4. Accuracy comparison of various algorithms Algorithm

Accuracy Precision Recall F-1 score

Decision tree

0.77

0.77

0.77

0.77

Voting

0.84

0.84

0.83

0.83

Bagging

0.82

0.82

0.81

0.81

AdaBoost

0.72

0.72

0.72

0.72

Proposed ensemble model 0.89

0.89

0.89

0.88

The AdaBoost gives the lowest accuracy among the algorithms being compared. This ensemble model was made by using decision tree as the base classifier. The learning rate was set as 0.07 with 500 classifiers. There is an inherent trade-off between these two values and further experiments showed that tuning these parameters allow for some fluctuations in the classification report of this classifier. The problem in this instance is that this model in this form has overfit the data and thus shows subpar performance on the test set. This data has been used by other researchers to predict the heart disease diagnosis in patients. The results from [6, 14] and [15] are compared with the proposed model in Table 5. The complete evaluation metrics for [6] are not reported in the paper. The researchers in [6] have proposed the Vote ensemble as their proposed model and by comparison, their algorithm outperforms the Vote ensemble performed in this paper. This can be contributed to two reasons: different base learners and feature selection. In [6], the base learners for Vote are Logistic Regression model and Naïve Bayes whereas in our proposed model the base learners are Random Forest and SVM. Additionally,

184

F. Kashif and U. K. Yusof

the difference in accuracy can be due to the feature selection step. We have used all the features in the Cleveland dataset whereas in [6], 9 of these features were selected. Table 5. Comparison of proposed model with existing literature Best algorithm

Accuracy Precision Recall F-1 score

Vote [6]

0.874

HRFLM [14] Logistic regression [15]

–

–

–

0.884

0.901

0.928

0.90

0.8625

0.89

0.86

0.86

0.89

0.89

0.88

Proposed ensemble model 0.89

From the results, we can see that the overall accuracy is enhanced in detection of heart disease. The proposed model is accurately predicting both classes, in it that it can be considered as robust for either case. The ensemble model proposed is not favoring any class over the other which is a desirable property in classification tasks. The comparison in Table 5 shows that no other model is as consistent as the one proposed in this paper. The stability of this model makes it a good candidate for classifying patients in this domain.

5 Conclusion Cardiovascular diseases are quite prevalent in today’s day and age, early and accurate diagnostics can help people achieve a higher quality of life. In this paper ensemble learning using random forest and support vector machine were used to predict heart disease in patients based on medical records. In this study 303 records of the Cleveland Heart Disease dataset were observed and used to create classification models. The ensemble random forest and support vector machine model outperformed the other algorithms and showed an overall accuracy of 0.89 and F-1 score of 0.88. The results of the proposed model are compared with several other techniques. This research can be extended to address higher dimensional and real-world data. In this paper, all 13 attributes were used to create the model without any feature selection. In future, we intend to perform comprehensive feature selection to select appropriate attributes. Feature selection can impact the accuracy levels of the model. Additionally, the method proposed in this paper is an ensemble of random forest and SVM with radial basis function kernel, a different combination of algorithms can be tried to improve the accuracy. Furthermore, different preexisting or custom kernels can be applied to the data to improve prediction accuracy. Acknowledgements. The authors would like to thank Universiti Sains Malaysia (USM) for the support and encouragement to conduct this research through the Research University Grant (RUI) (1001/PKOMP/8014084).

Detection of Cardiovascular Disease Using Ensemble Machine Learning Techniques

185

References 1. Ali´c, B., Gurbeta, L., Badnjevi´c, A.: Machine learning techniques for classification of diabetes and cardiovascular diseases. In: 6th Mediterranean Conference on Embedded Computing (MECO) IEEE (2017) 2. Alpaydın, E.: Introduction to Machine Learning. MIT Press, (2015) 3. Amin, M.S., Chiam, Y.K., Varathan, K.D.: Identification of significant features and data mining techniques in predicting heart disease. J. Telematics Inf. 36, 82–93 (2019) 4. Arji, G., Safdari, R., Rezaeizadeh, H., Abbassian, A., Mokhtaran, M., Ayati, M.H.: A systematic literature review and classification of knowledge discovery in traditional medicine. J. Comput. Methods Programs Biomedicine 168, 39–57 (2019) 5. Durairaj, G., Oommen, A.T., Pillai, G.: Correlation between BMI, Hba1c and fasting lipid profile in patients presenting with acute coronary syndrome and their relationship with CVD Risk. J. Cardiovascular Disease Res. 2, 10 (2019) 6. Islam, S., Jahan, N., Khatun, M.E.: Cardiovascular Disease Forecast using Machine Learning paradigms. In 10th International Conference on Computing Methodologies and Communication (ICCMC), pp. 487–490. IEEE (2020) 7. Kumar, S., Sahoo, G.: Enhanced decision tree algorithm using genetic algorithm for heart disease prediction. J. Int. J. Bioinf. Res. Appl. 14(1–2), 49–69 (2018) 8. Mohan, S., Thirumalai, C., Srivastava, G.: Effective heart disease prediction using hybrid machine learning techniques. J. IEEE Access. 7, 81542–81554 (2019) 9. Nowbar, A.N., Gitto, M., Howard, J.P., Francis, D.P., Al-Lamee, R.: Mortality from ischemic heart disease: analysis of data from the World Health Organization and coronary artery disease risk factors From NCD Risk Factor Collaboration. J. Circulation: Cardiovascular Quality and Outcomes. 12 (6), (2019) 10. World Health Organization: Cardiovascular Disease. https://www.who.int/health-topics/car diovascular-diseases 11. UCI Machine Learning Repository: Heart Disease Dataset. https://archive.ics.uci.edu/ml/dat asets/heart+disease 12. Sisodia, D., Sisodia, D.S.: Prediction of diabetes using classification algorithms. J. Procedia computer science 132, 1578–1585 (2018) 13. Sutton, C.D.: Classification and regression trees, bagging, and boosting. Handbook Stat. 24, 303–329 (2005) 14. Xu, G., Liu, M., Jiang, Z., Söffker, D., Shen, W.: Bearing fault diagnosis method based on deep convolutional neural network and random forest ensemble learning. J. Sensors 19(5), 1088 (2019) 15. Zumel, N., Mount, J., Porzak, J.: Practical data science with R. Manning Publications, Shelter Island (2014)

Health Information Management

Hospital Information System for Motivating Patient Loyalty: A Systematic Literature Review Saleh Nasser Rashid Alismaili1,2(B) , Mohana Shanmugam1 , Hairol Adenan Kasim1 , and Pritheega Magalingam3 1 College of Informatics and Computing, Universiti Tenaga Nasional, Kajang, Malaysia

[email protected] 2 Directorate of Information Technology, Ministry of Health, Muscat, Sultanate of Oman 3 Razak Faculty of Technology and Informatics, Universiti Teknologi Malaysia, Skudai,

Malaysia

Abstract. Healthcare service institutions (HIS) seeking to motivate patient loyalty have identified Hospital Information Systems (HIS) as a potential solution to gather, measure, and analyze the healthcare data necessary for this goal. The purpose of this systematic review of the literature is to reveal how prevalent the use of HIS with respect to motivating patient loyalty, and to investigate the efficacy of HIS in doing so. To generate data, published empirical studies and conference papers from the past five years were compiled from the following online databases: Scopus, ACM Digital Library, IEEE Xplore, ScienceDirect, and Emerald Insight. The search results indicate that, while the use of HIS in motivating patient loyalty is rare relative to other topics within the general field of HIS, HIS use have a significant positive impact on patient satisfaction, which is understood in the literature to be directly related to patient loyalty. There remains a gap in empirical studies on the direct application of HIS with the purpose of increasing patient loyalty. Future research may be required on the development of an HIS focused on motivating patient loyalty, which can be empirically tested in a real-world HSI setting. Keywords: Hospital Information Systems · Patient loyalty · Patient satisfaction

1 Introduction Due to cheap travel costs and rising healthcare standards in developing countries, more patients across the globe are choosing to go overseas for healthcare. In response, local governments, hospital administrators, and other stakeholders are seeking ways to retain their patients and motivate loyalty to their healthcare service institutions (HSIs). There is a clear consensus in the relation between service quality, patient satisfaction, and patient loyalty. In brief, service quality variables influence patient satisfaction; patient satisfaction, in turn, influences patient loyalty [1–3]. Thus, it has been argued in the literature that, to motivate patient loyalty, it would be necessary to motivate patient satisfaction by meeting the service quality variables laid out by [4], namely, tangibility, reliability, responsiveness, assurance, and empathy [5, 6]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 189–198, 2021. https://doi.org/10.1007/978-3-030-70713-2_19

190

S. N. R. Alismaili et al.

A significant complication inherent in these variables is the observed differences, not just in the relative importance specific groups place on each variable, but also how the variables are conceived in the first place. For instance, [7] found that younger patients valued the variables of empathy and tangibles more than their older counterparts, while single patients valued reliability more than their married peers. Another example is the high value Ghanaian patients place on individual attention from HSI staff [8], which is markedly different from Japanese patients, who were most satisfied when HIS staff treated them the same as other patients [9]. Thus, if HSIs wish to motivate patient loyalty, the best result would likely require the data gathered from their own patient base, rather than from other HSIs that may cater to other segments of people. HSIs attempting to meet all service quality variables require the gathering, measurement, and analysis of massive amounts of data in a timely manner. Not only do HSIs have to handle large sets of healthcare data, they must do so while respecting patients’ confidentiality, and keeping their trust [10]. The emergence of Hospital Information Systems (HIS) may assist in accomplishing this task. HIS can be understood simply as the methods with which a hospital manages the information within their organization; it has been defined as a “comprehensive, integrated Information System (IS) designed to manage the administrative, financial, and clinical aspects of a hospital” [11]. HIS is intended to make sense of the large amounts of information that HSIs collect every day and present them in a way that can be utilized easily by HSI administrators and stakeholders. However, HSI investments in HIS have focused mainly on improving performance from the perspective of administrators rather than from the perspective of patients [12, 13]. The purpose of this systematic review of the literature is to determine how prevalent the use of HIS is for the goal of motivating patient loyalty, as opposed to focusing on administrative concerns. An additional purpose is to reveal how effective HIS has been in motivating patient loyalty in the studies that use it for that goal. To reveal the prevalence and effectiveness of HIS use with respect to motivating patient loyalty, the following research questions were formulated: • RQ1. How prevalent is the use of Hospital Information Systems to motivate patient loyalty? • RQ2. How effective is Hospital Information Systems in motivating patient loyalty? By reviewing the current state of the literature regarding both the prevalence and efficacy of HIS with respect to motivating patient loyalty, this review aims to reveal how commonly HSIs use HIS to increase patient loyalty and how effective such use has been. The review will be organized into four sections. The first section will discuss the methods utilized for the systematic review, the second will outline the results of the review, the third will expand upon the obtained results, and the fourth and final section will include the conclusion of the study, and also possible directions for future research.

Hospital Information System for Motivating Patient Loyalty

191

2 Methods 2.1 Literature Search A search of five online research databases was made to compile relevant articles for the review: ACM Digital Library, Emerald Insight, IEEE Xplore, ScienceDirect, and Scopus. The search was conducted from late February to early May 2020. The search results were reconfirmed on June 2020. To generate the articles, the following search string was used: “hospital information system AND (“patient loyalty” OR “patient satisfaction”)”. Initially, the search string used was simply “hospital information system” AND “patient loyalty”. However, there were too few results using this string. Due to the robust evidence linking patient satisfaction and patient loyalty, “patient satisfaction” was added to the search string in order to generate more results. 2.2 Inclusion and Exclusion Criteria As the review is focused on the actual use of HIS by HSIs with the express purpose of motivating or improving patient loyalty, only articles with the following characteristics were included: (1) the article must be published in English, and available in full text version, (2) the article must be published in a peer-reviewed academic journal within the last five years, or (3) the article is set to be published in a peer-reviewed academic journal this year, or (4) the article is a peer-reviewed article from an international computer science conference. Only empirical studies that focus on the use or effect of HIS on patient loyalty or satisfaction were included in the review, as the primary concern is the actual practice of HSIs with respect to HIS and patient loyalty. The inclusion criteria were intended to generate results pertinent to the current state of HIS use with respect to motivating patient loyalty in HSIs. Due to the rapid advancements in IT, extending the inclusion range beyond five years prior to the review may introduce obsolete data. Because the research questions of the review pertain to actual use of HIS by HSI staff, only empirical studies were included. Articles were excluded if they were published prior to 2015, not peer-reviewed, used HIS in a manner other than for motivating patient loyalty, or did not conduct an empirical study. Previous systematic literature reviews were also not included in the review as their findings may already be obsolete. 2.3 Data Extraction By applying the inclusion and exclusion criteria outlined above, a list of articles to be included in the review was identified and gathered for further refinement to build the final list. It was necessary to eliminate articles that emerged from more than one database to eliminate redundancies.

192

S. N. R. Alismaili et al.

To determine whether an article was to be included in the review, the titles of the results from the aforementioned search string were scanned; if they did not mention HIS, patient loyalty, or patient satisfaction, they were excluded. Next, the remaining results’ abstracts were examined to determine their relevance to the review. If the abstract did not contain reveal an empirical study on HIS, patient loyalty, or patient satisfaction, they were excluded. Finally, the remaining results’ full texts were read by to determine their relevance to the review. The first three authors of the review handled the initial search and abstract reviews. All four authors contributed to the full text readings and final selection of the articles. The full process is outlined below in Fig. 1.

Fig. 1. Data extraction process

3 Results This section presents a summary of the reviewed studies.

Hospital Information System for Motivating Patient Loyalty

193

3.1 Search Results

Table 1. Data extraction results Authors

Year

Study location

Purpose of the study

Relevant results

Limitations

Implications

Liang, Gu, Tao, Jain, Zhao and Ding [16]

2015

Large hospital in East China

To examine the influence of HIS on doctor-patient relationships and patient satisfaction through the lens of service fairness

Patient-accessible HIS increases the patients’ perception of service fairness, which in turn improves both doctor-patient relationships, as well as patient satisfaction

Data from a large hospital was utilized; data from smaller HSIs may lead to different results. Furthermore, because this is a Chinese hospital, Liang et al. (2015) noted that there may be cultural factors specific to the Chinese that may differ from other cultures with respect to perceptions of service fairness

There is a power imbalance between physicians and patients in health care, leading to potential tension when patients feel that their concerns are unappreciated or ignored. The use of HIS, which allows patients more access to pertinent medical and administrative information regarding themselves, may help remedy this imbalance and increase patient satisfaction and loyalty

Yoo, Jung, Kim, Kim, Lee, Ching and Hwang [18]

2016

A public tertiary general hospital in South Korea

To evaluate an HIS that addresses the difficulties of outpatients regarding the search for HSIs, keeping up with treatment regimens, and accessing tailored medical and administrative information

The authors conducted a survey on their satisfaction regarding the HIS, n = 43 (23 outpatients and 20 of their guardians). Participants exhibited a satisfaction level of roughly 4.0 on a 5-point Likert scale

An Android-based mobile app was used by the outpatients. The study did not discuss the HIS used by the hospital. The results may therefore apply only to HIS initiatives solely focused on patients, and not on HSI-wide efforts to utilize HIS

Outpatients value the easy access to pertinent medical and administrative information at a glance. Visiting an HSI can often be a stressful experience, particularly for older patients. If they can access information readily without having to ask a staff member, they may feel more empowered and thus more satisfied—the increased satisfaction may motivate them toward loyalty for the HSI

(continued)

194

S. N. R. Alismaili et al. Table 1. (continued)

Authors

Year

Study location

Purpose of the study

Relevant results

Limitations

Implications

Khalifa [15]

2017

Four hospitals in Saudi Arabia (2 private, 2 public)

To reveal the perceived benefits of HIS and electronic medical records (EMR) from the point of view of patients

After 153 valid survey responses, the patients perceived the following benefits for HIS and EMR: 1) Improved information access, 2) Increased healthcare professionals productivity, 3) Improved efficiency and accuracy of coding and billing, 4) Improved quality of healthcare, 5) Improved clinical management (diagnosis and treatment), 6) Reduced expenses associated with paper medical records, 7) Reduced medical errors, 8) Improved patient safety, 9) Improved patient outcomes and 10) Improved patient satisfaction

HIS was examined solely from the point of view of patients. Considerations from the HSI’s point of view were left out, which means there is limited data on whether the HIS would be financially viable on their end

Patients are more satisfied when they can feel empowered and exercise their informed autonomy in HSI interactions. HIS that increases information availability and convenience will likely lead to greater patient satisfaction, as well as greater patient loyalty

Meyerhoefer, Sherer, Deily, Chou, Chen, Sheinberg and Levick [17]

2018

An obstetrics and gynecology practice in eastern Pennsylvania

To examine the impact of installing an EHR system at OB/GYN practices

HSI staff were dissatisfied with the EHR system; physicians most especially. Patient satisfaction decreased after the installation of the EHR system

Only OB/GYN practices were considered; results may not apply to other branches of healthcare

The negative impact of EHR on patient satisfaction may be due to the dissatisfaction of HSI staff with the system, which may have impacted staff compliance. HIS tasked with motivating patient loyalty must have support from HSI staff to be viable

(continued)

Hospital Information System for Motivating Patient Loyalty

195

Table 1. (continued) Authors

Year

Study location

Purpose of the study

Relevant results

Limitations

Implications

Asagbra, Burke, & Liang [14]

2019

Acute care hospitals in the United States

To examine the relationship between HIS functionalities and the quality of care by the HSI

The more comprehensive the coverage of the different HIS functionalities, the higher the satisfaction of patients. The number of functionalities also correlated negatively with readmission rates for myocardial infarction, hearth failure, and pneumonia

Secondary data was utilized by Asagbra et al. (2019), all of which were surveys

HIS that meets patient needs leads to greater patient satisfaction, which may, in turn, lead to more robust patient loyalty

4 Discussion 4.1 Prevalence of HIS Tasked with Motivating Patient Loyalty In contrast to the thousands of results one obtains by searching for “hospital information system” in online research databases, limiting the search terms to mentions of “hospital information system” in combination with “patient loyalty” drastically decreased the search results. To remedy this, the search term “patient satisfaction” was added in order to expand the results. The final data extraction resulted in just five studies; this indicates that there is a clear lack of studies focused on HIS with the specific intention of patient loyalty. 4.2 Efficacy of HIS Tasked with Motivating Patient Loyalty The reviewed studies revealed that HIS has significant positive effects on patient satisfaction, which in turn, motivates patient loyalty, which supports the findings of previous literature. Four out of five studies revealed that the increased information availability of HIS led to greater satisfaction rates from patients. In [14], secondary data from acute care hospitals was analyzed to reveal whether the HSIs’ HIS functionalities predicted the satisfaction rates of patients. It was found that the more HIS functionalities are present in an HSI, the more likely it is for patients to report satisfaction. A potential contributor to higher patient satisfaction rates is the lower readmission rates among HSIs with more HIS functionalities. Because an HIS can present pertinent medical data tailored to the patients’ own needs, patients are better able to follow their treatment regimens, which lead to better patient outcomes. In fact the improved outcomes may be attributed by patients to better information availability. Similar results were also found in [15–18]. A common thread among the four studies that revealed the relationship between information availability and patient satisfaction [14–16, 18] is the subjects’ apparent priority of information availability. A potential reason for this was illuminated by [16],

196

S. N. R. Alismaili et al.

who noted that, in China, there is a wide gulf between patients and medical staff, especially physicians, in terms of power. Patients often feel powerless in the face of illness; if medical staff is unable to empower patients by providing them with pertinent information quickly, they feel less powerful. This may contribute to their lack of satisfaction, as patients can feel confused and uncertain if they feel uninformed about the specifics of their treatment or care. The outlier was the result obtained by [17], which examined the effect of HIS on an OB/GYN practice. They found that patients were less satisfied before the implementation of HIS on the practice than during, as well as after, the implementation. However, the result may have been influenced by the dissatisfaction exhibited by the medical staff with the HIS employed, particularly among physicians. Their dissatisfaction could have negatively impacted their performance, which in turn could have led to the loss of satisfaction reported by patients. Overall, it appears that the use of HIS leads to positive effects with respect to increasing patient loyalty. However, the effect revealed in the reviewed studies is indirect; that is, HIS positively impacts patient satisfaction, which can then be inferred to lead to improved patient loyalty. 4.3 Gaps in the Literature There appears to be a need to examine the use of HIS toward the specific use of motivating patient loyalty. Based on the short list of results for HIS and patient loyalty, it appears that much of the scholarly focus regarding the overall aim of HIS is centered on other issues. Some of the more common search results for HIS concerns technology adoption or conceptual frameworks. Traditionally, the healthcare industry has utilized HIS to tend to administrative concerns, such as the streamlining of billing procedures and the storage of medical records or patient data [12]. It is mostly assumed that the administrative benefits will result in greater patient satisfaction and loyalty, due to the efficiencies brought about by HIS. Despite the surge in patient-centered HIS in recent years, there is still a gap in empirical studies on the effect of such HIS on patient loyalty. Instead, much of the focus is on the conceptual development, implementation, and adoption of patient-centered HIS. One potential explanation for this focus is that most researchers believe that motivating patient loyalty is a byproduct of the improvements brought about by HIS, rather than the primary goal. The short list of empirical studies on HIS and patient loyalty indicate that, when the impact of HIS on patient loyalty is examined, the results are significantly positive. It is noteworthy, however, that the reviewed articles examined patient loyalty indirectly. That is, they did not investigate the impact of HIS on patient loyalty specifically; instead, this connection could only be inferred by the positive impact of HIS on patient satisfaction. There appears to be a literature gap in the empirical testing of an HIS that treats patient loyalty as its primary goal, rather than an incidental effect of improving HSI services.

Hospital Information System for Motivating Patient Loyalty

197

4.4 Limitations This review is limited by the short list of articles generated by the search terms utilized. The articles were also limited by publication date, specifically, within the last five years; it is possible that much of the empirical testing of the effects of HIS on patient loyalty was enacted prior to this period. The search results may also have been constrained by the selection of research databases, which was limited by the available resources of the school library.

5 Conclusions and Directions for Future Research Scholarly research on HIS yields a large number of results; however, when the research is limited to HIS in relation to generating patient loyalty, the number shrinks significantly. The review revealed that the prevalence of HIS for the express purpose of motivating patient loyalty is low. However, in the few studies where the impact of HIS was investigated, a majority found that HIS had a significant positive impact for patient satisfaction. Due to the robust literature on the relationship between patient satisfaction and patient loyalty, it would be reasonable to infer that HIS is likely effective in motivating patient loyalty. However, a direct examination of an HIS geared toward increasing patient loyalty was not found. Future research may be directed toward the development of HIS designed to motivate patient loyalty, as well as the empirical testing of this HIS in a real-world HSI setting.

References 1. Mohebifar, R., Hasani, H., Barikani, A., Rafiei, S.: Evaluating service quality for patients’ perceptions: application of importance-performance analysis method. Osong Public Health Res. Perspect. 7(4), 233–238 (2016) 2. Shabbir, A., Malik, S.A., Malik, S.A.: Measuring patients’ healthcare service quality perceptions, satisfaction, and loyalty in public and private sector hospitals in Pakistan. Int. J. Q. Reliab. Manag. 33(5), 538–557 (2016) 3. Tosyali, H., Sütcü, C.S., Tosyali, F.: Patient loyalty in the hospital patient relationship: the mediating role of social media. Erciyes˙Ileti¸simDergisi 6(1), 783–804 (2019) 4. Parasuraman, A., Zeithaml, V.A., Berry, L.L.: Servqual: a multiple-item scale for measuring consumer perceptions of service quality. J. Retail. 64(1), 12 (1988) 5. Asnawi, A., Awang, Z., Afthanorhan, A., Mohamad, M., Karim, F.: The influence of hospital image and service quality on patients’ satisfaction and loyalty. Manag. Sci. Lett. 9(6), 911–920 (2019) 6. Kulsum, U., Yanuar, T., Syah, R.: The effect of service quality on loyalty with mediation of patient satisfaction. Int. J. Bus. Manag. Invention 6(3), 41–50 (2017) 7. Ahmed, S., Tarique, K.M., Arif, I.: Service quality, patient satisfaction and loyalty in the Bangladesh healthcare sector. Int. J. Health Care Qual. Assur. 30(5), 477–488 (2017) 8. Nkrumah, S., Yeboah, F., Adiwokor, E.: Client satisfaction with service delivery in the health sector: the case of Agogo Presbyterian Hospital. Int. J. Bus. Adm. 6(4), 64–78 (2015) 9. Elleuch, A.: Patient satisfaction in Japan. Int. J. Health Care Qual. Assur. 12(7), 692–705 (2008)

198

S. N. R. Alismaili et al.

10. Bayer, R., Santelli, J., Klitzman, R.: New challenges for electronic health records: confidentiality and access to sensitive health information about parents and adolescents. J. Am. Med. Assoc. 313(1), 29–30 (2015) 11. Ahmadi, H., Nilashi, M., Shahxradi, L., Ibrahim, O.: Hospital information system adoption: expert perspectives on an adoption framework for Malaysian public hospitals. Comput. Hum. Behav. 67, 161–189 (2017) 12. Sheikh, A., Sood, H.S., Bates, D.W.: Leveraging health information technology to achieve the “triple aim” of healthcare reform. J. Am. Med. Inform. Assoc. 22(4), 849–856 (2015) 13. Wang, Y., Kung, L., Byrd, T.A.: Big data analytics: Understanding its capabilities and potential benefits for healthcare organizations. Technol. Forecast. Soc. Chang. 126, 3–13 (2018) 14. Asagbra, O.E., Burke, D., Liang, H.: The association between patient engagement HIT functionalities and quality of care: does more mean better? Int. J. Med. Inf. 130, 103893 (2019) 15. Khalifa, M.: Perceived benefits of implementing and using hospital information systems and electronic medical records. In: ICIMTH, pp. 165–168. IOS Press (2017) 16. Liang, C., Gu, D., Tao, F., Jain, H.K., Zhao, Y., Ding, B.: Influence of mechanism of patientaccessible hospital information system implementation on doctor–patient relationships: a service fairness perspective. Inf. Manag. 54(1), 57–72 (2017) 17. Meyerhoefer, C.D., Sherer, S.A., Deily, M.E., Chou, S.Y., Guo, X., Chen, J., Sheinberg, M., Levick, D.: Provider and patient satisfaction with the integration of ambulatory and hospital EHR systems. J. Am. Med. Inform. Assoc. 25(8), 1054–1063 (2018) 18. Yoo, S., Jung, S.Y., Kim, S., Kim, E., Lee, K.H., Chung, E., Hwang, H.: A personalized mobile patient guide system for a patient-centered smart hospital: lessons learned from a usability test and satisfaction survey in a tertiary university hospital. Int. J. Med. Informatics 91, 20–30 (2016)

Context Ontology for Smart Healthcare Systems Salisu Garba(B) , Radziah Mohamad, and Nor Azizah Saadon(B) School of Computing, Faculty of Engineering, Universiti Teknologi Malaysia, UTM, 81310 Skudai, Johor, Malaysia [email protected], {radziahm,azizahsaadon}@utm.my

Abstract. This paper proposes an improved Context Ontology for Smart Healthcare Systems. The main contribution of this work is the simplification, sufficiently expressiveness, and extendability of the smart healthcare context representation, in which only three contextual classes are required—compared to several classes in the related context ontologies. This is achieved by adapting the feature-oriented domain analysis (FODA) techniques of software product line (SPL) for domain analysis, and subsequently, the lightweight unified process for ontology building (UPON Lite) is used for ontology development. To validate the applicability of the proposed context ontology, sustAGE smart healthcare case study is used. It is found that the proposed context ontology can be used to sense, reason, and infer context information in various users, environments, and smart healthcare services. The ontology is useful for healthcare service designers and developers who require simple and consolidated ontology for complex context representation. This paper will benefit the smart healthcare service developers, service requesters as well as other researchers in the ontology-based context modeling domain. Keywords: Ontology · Smart healthcare · Context ontology · Healthcare service

1 Introduction The smart healthcare system is an intelligent system that makes use of modern technologies such as IoT, Big Data, advanced analytics with deep learning for better diagnosis of the disease, better treatment of the patients, and improved quality of lives with the aid of vital components (mHealth and eHealth) for efficient and effective communication between individuals and health service providers [1]. Wireless Sensor Network (WSN) serves as the enabling technology for the transformation of healthcare applications. WSNs consist of an array of sensors that can monitor all-natural phenomena such as body temperature, blood pressure, heart rate, glucose, breath rate in smart healthcare [2]. Healthcare systems require a high transmission rate and lower delay which mandates mobile network operators to shift from network monitoring to service monitoring device monitoring [3]. This leads to the generation of massive data that necessitate proper representation, intelligent analytics to enable appropriate decisions in smart healthcare systems [4].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 199–206, 2021. https://doi.org/10.1007/978-3-030-70713-2_20

200

S. Garba et al.

Among the resonant context space, query preference, and vector space, objectoriented and ontology are widely employed [5]. The most viable method to organize, represent, and generate inference from this massive context information is ontology [6]. An ontology is a machine-readable and precise representation of rigorous conceptual schema (relevant entities, properties, relations, and rules) that is derived from consensus to capture meaning (semantics) within the domain of discourse [7]. Despite the extensibility and expressive power of ontology, its application in context representation for smart healthcare systems exhibits several challenges such as constantly changing context, heterogeneous smart devices, building ontologies from scratch, and many more [8]. The existing context ontologies are mostly constricted and slightly fragmented as the smart healthcare context is user, service, and environment-dependent. This may be probably due to the impression that the smart healthcare context is similar to traditional healthcare systems. Given these shortcomings, an improved context ontology for smart healthcare systems in this paper. The idea is based on the previously defined context ontologies integrated using UPON, unlike other approaches thus far—which rely on object-oriented or creation of ontology from scratch for semantic reasoning in smart healthcare systems. The key benefits of the proposed context ontology for smart healthcare systems are expressiveness and expandability. The rest of this paper is organized as follows: The related works are discussed in Sect. 2; The smart healthcare context properties analysis is discussed in Sect. 3; The proposed context ontology is discussed in Sect. 4; The applicability and validity of the context ontology are illustrated through a typical case study in Sect. 5; Lastly, the conclusions are presented in Sect. 6.

2 Related Work Several authors have recognized the importance of ontologies in organizing and represent the massive context information in the smart healthcare domain to realize interoperability among smart healthcare systems. For instance, [9] proposed an ontology that represents valuable context suitable for the provision of healthcare monitoring services in pervasive healthcare systems. The ontology described four fundamental concepts (personal data, sensor data, services, and host) and also the relationships between these concepts, furthermore, physician, and developer rules are used for context reasoning. The authors in [10] developed an upper-level context ontology to model the daily activities of elderly people in a smart home. Context information such as user, activity, location, sensors, physical objects, and temporals form the upper-level ontology to provide a homogeneous view over the heterogeneous data and generate semantics for activity modeling and context representation. Other research contributions such as ontology-based teal-time data modeling and knowledge representation for a smart healthcare system [11], and medical ontology for the effective management of healthcare system during an emergency in the dynamic environment [12], focus more on the medical information rather than the user context necessary for the provision of personalized smart healthcare service.

Context Ontology for Smart Healthcare Systems

201

Although the related studies have contributed immensely to the context ontology for the ubiquitous healthcare domain. The usefulness of other ontologies should not be ignored. The evaluation of applicability and validation of the context ontology should be extended with real data, in a real case study, which will improve the quality of the smart healthcare system, thereby making possible to provide the required services to healthy individuals, elderly people, or patients with chronic diseases [13].

3 Smart Healthcare Context Properties Analysis To analyze the context properties in the smart healthcare domain, Feature-Oriented Domain Analysis (FODA) techniques of Software Product Line (SPL) is adapted to capture the commonality and variability of the smart healthcare domain. This involves domain planning, feature identification, Feature extraction based on six case studies from healthy living and assisted living smart healthcare, Commonality, and variability analysis. The activities in the Feature-Oriented Domain Analysis (FODA) technique are shown in Fig. 1. The six case studies identified for feature extraction are as follows: 1. Healthy-living case study: sustAGE—smart healthcare for sustainable well-being of employees in EU industries 2. Healthy-living case study: Sports performance monitoring and injury prevention framework 3. Healthy-living case study: Food and activity tracking for a disease prevention system 4. Assisted-living case study: SMART BEAR—smart living and healthcare system for elderly 5. Assisted-living case study: xVLEPSIS—Smart non-invasive healthcare monitoring system for infants 6. Assisted-living case study: Smart Heart disease monitoring system

Fig. 1. The Feature-oriented domain analysis (FODA) technique.

The existing commonality and variability in the smart healthcare domain are identified based on the fundamental concepts proposed in [14] and [15]. With the identified commonality and variability within the context properties of smart healthcare, the context ontology can be developed with flexibility and reusability.

202

S. Garba et al.

4 Context Ontology for Mobile Service Instead of reinventing the wheel in the development of the proposed context ontology for mobile web service in smart healthcare, the Unified Process for ONtology building (UPON Lite) methodology [16] which is a lightweight extension of Network Ontology (NeOn) [17] is adopted.

Fig. 2. The Six stages of steps in UPON-Lite methodology.

The first two steps of the UPON Lite methodology about the domain is archived in Sect. 3, while the other steps about the ontology are discussed in this section. Classes are used to concretely represent ontology concepts in the process of ontology construction. The main classes are; “User” class, “Services” class, and “Environment” class. Figure 2 presents the taxonomy (hierarchical representation of the concepts) for smart healthcare context ontology. In ontology construction, relationships such as “hasCharacteristics”, “isLocatedIn” are used to show the interaction between the ontology concepts based on the properties and by the attributes that describe the concepts. Figure 3 shows an overview of the context ontology classes, object properties, data properties, and individuals in Protégé (Fig. 4).

Context Ontology for Smart Healthcare Systems

Fig. 3. The taxonomy of the smart healthcare context ontology

203

204

S. Garba et al.

Fig. 4. An overview of the context ontology classes and properties in Protégé

5 The Evaluation of the Proposed Context Ontology To evaluate the proposed context ontology and demonstrate its applicability, the sustAGE case study is used. Figure 5 shows the system architecture of the sustAGE case study discussed in [18].

Fig. 5. The sustAGE system architecture

The context spectrums and the context situations of the sustAGE case study are organized based on the proposed context ontology to represent all the available context for inference generation for the smart healthcare system using Protégé together with

Context Ontology for Smart Healthcare Systems

205

Pellet reasoner. Given that, it is the most widely used tool for expressive, fast, and flexible ontology development [19]. Figure 6 shows the description and property assertion of an assembly line worker in the sustAGE case study.

Fig. 6. The description and property assertion of an assembly line worker

An inference can be generated for the sustAGE system, for example, CareMessage service can be used to alert users (assembly line worker) via SMS, alarm, etc., depending on the user preferences and environment devices, e.g., proximity to hazardous conditions. The rules of inference for the “CareMessage service” are the following, in the case that the proximity to hazardous conditions is 100 m. • Worker(?w)m ˆ isLocatedIn(?w, ?l) ˆ Location(?l) ˆ locationType(?t, “hazardous conditions”) ˆ proximity(?p, 100) - > AlertWorker(?w). The above Semantic Web Rule Language (SWRL) is one of the methods used to create rules that govern the inference generation and reasoning process for smart healthcare systems. Among the numerous axioms, classes, object properties, datatype properties, and individuals of the proposed context ontology, no inconsistency was detected.

6 Conclusion and Future Work This paper proposes an improved Context Ontology for Smart Healthcare Systems. The feature-oriented domain analysis (FODA) techniques of software product line (SPL) is adapted for domain analysis which also complements domain expert’s opinion from existing ontologies. The UPON-Lite methodology is used to develop the proposed ontology. To validate the applicability of the proposed context ontology, a sustAGE smart healthcare case study is used. Overall, the results demonstrate a strong effect of the proposed context ontology as it’s more consistent, expressive, extendable, and can be used for context reasoning in smart healthcare systems. Future investigations will consider smart systems case studies and the development of smart healthcare systems based on the proposed context ontology to further validate the conclusions that are drawn from this study. Acknowledgments. We would like to thank the Ministry of Education (MOE) Malaysia for sponsoring the research through the Fundamental Research Grant Scheme (FRGS) with vote number 5F080 and Universiti Teknologi Malaysia for providing the facilities and supporting the research. In addition, we would like to extend our gratitude to the lab members of Software Engineering Research Group (SERG), School of Computing, Faculty of Engineering, Universiti Teknologi Malaysia for their invaluable ideas and support throughout this study.

206

S. Garba et al.

References 1. Park, S.J., et al.: Development of the elderly healthcare monitoring system with IoT. In: Advances in Human Factors and Ergonomics in Healthcare, vol. 482, pp. 309–315. Springer (2017) 2. Khalaf, O.I., Sabbar, B.M.: An overview on wireless sensor networks and finding optimal location of nodes. Periodicals Eng. Natural Sci. 7(3), 1096–1101 (2019) 3. Salman, A.D., Khalaf, O.I., Abdulsahib, G.M.: An adaptive intelligent alarm system for wireless sensor network. Indonesian J. Electr. Eng. Comput. Sci. 15(1), 142–147 (2019) 4. Khalaf, O.I., Abdulsahib, G.M., Kasmaei, H.D., Ogudo, K.A.: A new algorithm on application of blockchain technology in live stream video transmissions and telecommunications. Int. J. e-Collaboration 16(1), 16–32 (2020) 5. Cabrera, O., Franch, X., Marco, J.: Ontology-based context modeling in service-oriented computing: a systematic mapping. Data Knowl. Eng. 110(May), 24–53 (2017) 6. Munir, K., Sheraz Anjum, M.: The use of ontologies for effective knowledge modelling and information retrieval. Appl. Comput. Inf. 14(2), 116–126 (2018) 7. Pradeep, P., Krishnamoorthy, S.: The MOM of context-aware systems: a survey. Comput. Commun. 137(January), 44–69 (2019) 8. Bagtharia, P., Bohra, M.H.: An optimal approach for web service selection. In: Proceedings of the 3rd International Symposium on Computer Vision and the Internet - VisionNet 2016, pp. 121–125 (2016) 9. HameurLaine, A., Abdelaziz, K., Roose, P., Kholladi, M.-K.: Ontology and rules-based model to reason on useful contextual information for providing appropriate services in U-healthcare systems. In: Intelligent Distributed Computing VIII, pp. 301–310. Springer (2015) 10. Ni, Q., García Hernando, A.B., De La Cruz, I.P.: A context-aware system infrastructure for monitoring activities of daily living in smart home. J. Sensors 2016, 1–9 (2016) 11. Abatal, A., Khallouki, H., Bahaj, M.: A smart interconnected healthcare system using cloud computing. In: ACM International Conference Proceeding Series (2018) 12. Zeshan, F., Mohamad, R.: Medical ontology in the dynamic healthcare environment. Procedia Comput. Sci. 10, 340–348 (2012) 13. Gubert, L.C., da Costa, C.A., da Rosa Righi, R.: Context awareness in healthcare: a systematic literature review. Universal Access in the Information Society, no. 0123456789 (2019) 14. Aguilar, J., Jerez, M., Rodríguez, T.: CAMeOnto: context awareness meta ontology modeling. Appl. Comput. Inf. 14(2), 202–213 (2018) 15. Lu, Z.J., Li, G.Y., Pan, Y.: A method of meta-context ontology modeling and uncertainty reasoning in SWoT. In: Proceedings - 2016 International Conference on Cyber-Enabled Distributed Computing and Knowledge Discovery, CyberC 2016, pp. 128–135 (2017) 16. De Nicola, A., Missikoff, M.: A lightweight methodology for rapid ontology engineering. Commun. ACM 59(3), 79–86 (2016) 17. Suárez-Figueroa, M.C., Gómez-Pérez, A., Fernández-López, M., Benjamins, V.R.: The NeOn methodology for ontology engineering. In: Ontology Engineering in a Networked World, pp. 9–34 Springer (2012) 18. Pateraki, M., et al.: Biosensors and Internet of Things in smart healthcare applications: challenges and opportunities. Wearable Implantable Med. Devices 5, 25–53 (2020) 19. Musen, M.A.: The protégé project. AI Matters 1(4), 4–12 (2015)

A Modified UTAUT Model for Hospital Information Systems Geared Towards Motivating Patient Loyalty Saleh Nasser Rashid Alismaili1,2(B) , Mohana Shanmugam1 , Hairol Adenan Kasim1 , and Pritheega Magalingam3 1 College of Informatics and Computing, Universiti Tenaga Nasional, Kajang, Malaysia

[email protected] 2 Directorate of Information Technology, Ministry of Health, Muscat, Sultanate of Oman 3 Razak Faculty of Technology and Informatics, Universiti Teknologi Malaysia, Skudai,

Malaysia

Abstract. Healthcare service institutions (HSIs) have sought ways to motivate patient loyalty in response to surging rates of medical tourism. Previous research indicates that Hospital Information System (HIS) is essential for HSIs to gather, measure, and analyze the massive amounts of data required to generate patient loyalty. There is currently no consensus on the factors that comprise HIS specifically geared towards motivating patient loyalty (HISPL). Furthermore, HIS requires full adoption by HSI staff to be effective. Thus, to reduce wastage of HSI resources, it is necessary to predict whether a given HIS specifically geared towards motivating patient loyalty is likely to be adopted. The purpose of this study is to reveal the factors that comprise HISPL and to modify the Unified Theory of Acceptance and Use of Technology (UTAUT) model to help predict the likelihood of an HISPL to be fully adopted by HSI staff. The results revealed that pertinent HISPL factors are capability, configurability, ease of use/help desk availability and competence (EU), and accessibility/shareability (AS). Using these factors, the UTAUT model was modified to fit the specific needs of HISPL. The modifications are theoretical and will have to be validated in future empirical studies. Keywords: Hospital Information System · Patient loyalty · UTAUT

1 Introduction Healthcare service institutions (HSI) seeking to motivate patient loyalty have turned to Hospital Information Systems (HIS) as a means of achieving their organizational goals. HIS is a “comprehensive, integrated Information System (IS) designed to manage the administrative, financial, and clinical aspects of a hospital [1]. When implemented correctly, HIS can assist greatly in improving healthcare quality and efficiency, as well as improve patient outcomes and reduce medical errors by freeing medical staff to focus solely on their jobs [2]. HIS that is designed with the express purpose to motivate patient loyalty will be referred to in this study as HISPL. The primary goal of an effective © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 207–216, 2021. https://doi.org/10.1007/978-3-030-70713-2_21

208

S. N. R. Alismaili et al.

HIS is to meet the needs of patients [3]. However, due to the lack of direct interaction between patients and HIS, patients can only perceive the effects of HIS indirectly, that is, through the improved performance of HSI staff empowered by the HIS [4]. HIS can motivate patient loyalty by empowering HSI staff, which improves patients’ perceived service quality; the higher patients’ perceptions of service quality are, the more likely they are to report being satisfied with their care, which would then motivate patient loyalty [5–10]. Prior research indicates that the effectiveness of HIS depends largely on its full-scale adoption by its end-users, the HSI staff [11, 12]. An HIS may be well-designed and fully functional, but if HSI staff refuse to use it across the board, it will be no better than a malfunctioning HIS [11, 12]. To be effective, an HIS must be adopted across the board by HSI staff [12]. It is therefore necessary to ensure that an HIS meets the needs of HSI staff to ensure its efficacy in a real world context [13]. Before an HIS can be utilized by an HSI, it is essential to predict whether the HIS will likely be adopted by HSI staff to avoid wasting resources. To predict the likelihood of an HIS’ adoption, the Unified Theory of Acceptance and Use of Technology (UTAUT) model was utilized. The UTAUT model was selected as it is the culmination of previous technology acceptance models in the field [14, 15]. The UTAUT is the most widely cited model of individual technology acceptance and use [15]. However, the original UTAUT model is not sufficient to predict the adoption of HIS, as it is a general model that must be revised depending on the context in which it is used [14]. The popularity of the UTAUT model is due in no small part to its adaptability into different contexts. The original UTAUT model was developed outside the healthcare industry, and focuses on individual adoptive behaviors [14]. As HSIs are large organizations, the UTAUT model must therefore be modified if it is to be used in this capacity [16]. There is currently no research on the UTAUT model modified to predict the adoption of HISPL. To modify the UTAUT model for evaluating the adoption of HISPL, it will be necessary to expand the original model’s focus on individuals to entire HSIs. Furthermore, the modified model will have to account for the needs of HSI staff. Due to the paucity of research on HISPL and the lack of consensus on the factors of HISPL, conducting a literature review is necessary to modify the UTAUT model for predicting the adoption of HISPL. The objective of this study is to reveal the relevant factors of HISPL according to the literature, and to modify the original UTAUT model to accommodate the relevant factors of HISPL. To that end, the following research questions were generated for this study: 1. What are the factors of HISPL revealed in the literature? 2. How can the original UTAUT model be modified to address the HISPL factors revealed in the literature? The rest of the study will be organized into the following sections: a review of the literature on factors that comprise HISPL, and the modification of the UTAUT model to account for the HISPL factors revealed in the literature review.

A Modified UTAUT Model for Hospital Information Systems

209

2 Literature Review This section contains the pertinent findings of the review, which reveal factors of HISPL. The findings will be divided into the four factors found to have a significant impact on HIPL, namely: capability, configurability, ease of use/help desk availability and competence, and accessibility/shareability. 2.1 Capability Capability refers to the technical aspects of the HIS, or the ability of the HIS to accomplish healthcare goals. This includes the HIS’ ability to collect pertinent medical data, facilitate communication between different individuals and departments within an HSI, appropriate system architecture, and fast response times for requests while ensuring that the data it collects is confidential and secure from outsiders [17]. Collecting, keeping, and analysis of data is a necessary component of any effective HSI functioning, especially as healthcare data accrues more and more worldwide [18]. While one effect of HIS is easing the procedure of billing and payments, some scholars have argued that past investments in IT among HSIs have focused on this issue, and have so far failed to capitalize on the other benefits IT offers to the clinical needs of HSI, especially from the perspective of patients [2, 19]. The lack of investment centered on easing the medical process for the customer manifests itself in the difficulty of procuring medical treatments in different parts of the world relative to how similarly information-rich industries function [19]. Thus, it is fundamental for HIS to be designed around the fact that it is patient-centered, in that it assists HSIs to meet the needs of patients, and not just meet the needs of HSIs. 2.2 Configurability Configurability refers to the ability of an HIS to be modified or changed depending on the needs of the medical personnel using the technology. Systems and work practices must be able to adapt to each other in order for an HIS to be effective [11]. Work practices refer to the “practices, procedures, and norms” at a given HIS [20]. Because each HSI can have different areas of focus and requirements, an effective HIS must be able to be configured easily in order to adapt to any foreseeable situation [21, 22]. In short, due to the multitudes of different contexts and situations of individual HSIs, an effective HIS must be an open, generic system that prioritizes flexibility so that the system can be configured for the HSI’s specific needs [20]. An added complication is that requests about configurability may be routinely ignored by IT vendors without explanation, leading to difficulties with formal templates present in the IT solutions that they may feel fails to address the core tasks that necessitated the solutions in the first place. HIS developers must endeavor to make changes to the system intuitive for most HSI IT departments. IT vendors must also pitch in if IT departments need assistance configuring the HIS to their own specifications. Similarly, if end-users have issues with the HIS, their IT departments must be able to resolve their issues, with the assistance of IT vendors if needed [23, 24].

210

S. N. R. Alismaili et al.

2.3 Ease of Use/Help Desk Availability and Competence (EU) While younger medical staff have been seen to be more likely to engage proactively with IT solutions in healthcare, it is important to note that older medical staff may have the desire to do so as well but lack the competence that could be easily acquired if IT training is presented properly to staff [24]. Thus, it is essential that an IT solution for HSIs be accommodating of staff members who are not fully competent in the new technology by making them user-friendly to medical staff, to encourage full compliance and encourage their learning process. In case medical staff retains difficulties adapting, a help desk tasked with dealing with issues that arise from the IT solution must be ready to assist competently and promptly in a manner that will not dissuade staff from asking for help [23, 24]. The aforementioned difficulty of focusing on older medical staff can be compounded further in HSIs from developing countries, where computer literacy, technological competence, and willingness to learn among non-IT staff may be low. While newer IT technologies tend to be more streamlined and thus much simpler to learn and use, older staff may prefer older systems with which they are already familiar [21]. EU is important to assist in the adoption of new systems, whether from on-site technical support or remotely, e.g. a call center [21]. The reason is because an HIS’ ease of use can reduce the learning curve, while help desk availability ensures that any difficulties learners face will be addressed promptly—thus avoiding the fears of some medical staff that learning a new system would be slower for them, potentially jeopardizing their patients’ outcomes. 2.4 Accessibility/Shareability For an HIS to be effective, it is necessary that the information it is tasked to handle can be accessed easily and conveniently by the many different departments and individuals within an HIS, to avoid delays, inconveniences, and potential harms [22]. The importance of efficiency in terms of granting efficient access to medical information to the relevant medical staff not only improves service time, but also improves service quality by allowing different medical teams to coordinate with each other without losing time [25, 26]. An illustration of this is the fact that vascular surgery outpatient appointments are becoming more difficult to manage in recent years due to a lack of information [27]. As this information is essential for the success of the operation, additional work tracking them down is required of the medical staff, time that could be spent more productively. Some of the sources of this information include patient notes, referral letters from general practitioners or the patients’ original doctor, and recent test results and scans [27]. The information must also be shareable to staff because inpatient and outpatient departments in HSIs tend to be structured independent of each other, and records from each may have to be repeated once the patient enters the other department, wasting time and resources [28]. This is especially important when a patient, in the course of an illness, enters many different phases of treatments, located in different departments within the HSI. For example, a person who enters the emergency room for a broken leg is confined to the inpatient department for further observation, then is transferred to home care for rehabilitative treatments. Each of these departments may not have access

A Modified UTAUT Model for Hospital Information Systems

211

to information obtained in other departments, necessitating the collection of redundant information, causing delays and lack of transparency, both of which can impact the quality of healthcare, as well as the patients’ perceptions of it [28]. Furthermore, reliable access of staff to medical data depends on a robust computer network with minimal delays. The accessibility/shareability of medical information is essential to the pursuit of quality healthcare [25, 26]. The more comprehensive one’s medical record is, the more likely it is for medical staff to make well-informed decisions about the patient’s care. The use of computerized medical records possess the ability to improve the quality and efficiency of HSIs drastically because they are much easier to share across different healthcare providers compared to paper records [19]. The increased shareability of patient records and other important medical information could help facilitate shorter waiting periods and more efficient medical interactions between different departments and HSIs [25].

3 Modified UTAUT Model for HISPL The Unified Theory of Acceptance and Use of Technology (UTAUT) model was developed by [14, 15], and aimed to consolidate the disparate views on technology acceptance to a single coherent model. UTAUT theorized that all such models utilized four core constructs that motivate behavioral intention—performance expectancy, effort expectancy, social influence, and facilitating conditions—and that these four core constructs were, in turn, moderated by individual characteristics, namely, age, gender, voluntariness, and experience [14]. The original UTAUT model is provided in Fig. 1. Performance expectancy refers to the degree an individual perceives the helpfulness of adopting technology in their job performance. Effort expectancy refers to the degree an individual perceives the ease of adopting a technology. Social influence refers to the degree in which an individual perceives others’ beliefs that they should adopt a technology. Facilitating conditions refer to the degree an individual perceives the technical and organizational support they receive in adopting a technology [14]. HISPL requires full adoption to be effective [12]. Without full compliance by HSI staff, even an excellent HISPL will fail to bring about its goals [11]. Thus, to ensure that HSI resources are not wasted needlessly in developing HISPL, the use of a technology adoption model is needed to ensure that the HISPL will be likely adopted by HSI staff. The choice to use UTAUT as the study’s base technology adoption model is justified by the current status of UTAUT as the most cited model of individual technological acceptance and use [16]. Within the healthcare context, UTAUT has been used most often in predicting the adoption of electronic medical records [29–31], but also in predicting acceptance of Information Systems among healthcare professionals [14, 32]. Within healthcare, UTAUT has been most associated with predicting the individual use of electronic medical records [30, 31], but has also been used in predicting the acceptance of Information Systems (IS) among healthcare professionals [14, 32, 33]. To modify the UTAUT model, results from the literature review were used to reveal the factors that influenced adoption and use of HISPL. The revealed factors were capability, configurability, EU, and AS.

212

S. N. R. Alismaili et al.

Fig. 1. Original UTAUT model [14–16]

UTAUT was modified by eliminating the original UTAUT variables that had no relation to HISPL, as revealed by the results of the literature review. This meant eliminating social influence and facilitating conditions, and splitting performance and effort expectancy into two distinct constructs, to mirror HSI professionals’ distinction between capability and configurability, as well as the distinction between EU and AS. Figure 2 shows the modified UTAUT model, based on the factors of HISPL revealed in the previous section. The choice to simplify the original model by reducing the moderating factors to just one—namely, age—is supported by [16] in a study that tailored the UTAUT model for electronic health records. The modification is validated by the natural comparison between the modified constructs and the original UTAUT constructs, which have already been validated in previous studies. Assimilating the factors of HISPL as revealed by the literature review into the original UTAUT model to generate a modified model for HISPL is justified in [14] as a way of creating an HISPL-specific UTAUT model, a method previously utilized by other studies of UTAUT in healthcare contexts [34, 35]. An additional factor in keeping age as the sole moderating factor is the strong evidence found for age’s significant effects on technology adoption in developing countries, such as Cameroon [36], Ghana [35, 37], and Brazil [38]. The factor of age may play a larger role than other moderating factors due to age affecting HSI staffs’ ability to learn a new system [21, 24], while factors such as gender, voluntariness, and experience are superseded by the organizational requirement to adopt new technologies such as HIS. The modified UTAUT model presented in this study is theoretical and will have to be validated in a future empirical study within a real life healthcare context. Based on the preceding literature review and discussion, it can be hypothesized that: H1: Capability has a significant positive effect on HISPL adoption among HSI staff. H2: Configurability has a significant positive effect on HISPL adoption among HSI staff. H3: Ease of use/help desk availability and competence has a significant positive effect on HISPL adoption among HSI staff. H4: Accessibility/Shareability has a significant positive effect on HISPL adoption among HSI staff.

A Modified UTAUT Model for Hospital Information Systems

213

Fig. 2. Modified UTAUT model for HISPL

H5: The influence of capability on HISPL adoption among HSI staff is moderated by age. H6: The influence of configurability on HISPL adoption among HSI staff is moderated by age. H7: The influence of ease of use/help desk availability and competence on HISPL adoption among HSI staff is moderated by age. H8: The influence of accessibility/shareability on HISPL adoption among HSI staff is moderated by age. The modified UTAUT model is part of an ongoing study on patient loyalty. The potential effect of HIS on patient loyalty is illustrated in Fig. 3:

Fig. 3. Relationship between HIS and Patient Loyalty [5–10]

4 Conclusions and Directions for Future Work The study conducted a literature review on the factors that comprise a Hospital Information System specifically geared towards motivating patient loyalty (HISPL) in health service institutions (HSIs). HISPL is distinguished from HIS in general as no HISPL currently exists. While HIS can and does motivate patient loyalty indirectly by improving patient satisfaction, such effects are often secondary to the HIS’ primary purpose of meeting administrative, financial, and clinical needs. For HISPL, attaining patient

214

S. N. R. Alismaili et al.

loyalty is the primary purpose. To that end, the literature review revealed that capability, configurability, ease of use/help desk availability and competence (EU), and accessibility/shareability (AS) are possible factors to HISPL. Given that the efficacy of HIS in general depends largely on its full-scale adoption by HSI staff, the technology adoption model UTAUT was modified to fit the factors of HISPL revealed in the literature review. In the modified UTAUT model, the factors of facilitating conditions and social influence were dropped due to the lack of support found in the literature review. Performance expectancy was divided into two distinct components to mirror the distinction found in the literature between capability and configurability. Effort expectancy was similarly divided into EU and AS. Only age was retained as a moderating factor due to the strong support it received in UTAUT studies conducted within the healthcare field—especially in developing countries, where HIS can make the most improvements—while others received mixed evidence. Future research will test the variables used in the modified UTAUT model for validity, with the ultimate goal of developing an HISPL for public HSIs in Oman.

References 1. Ahmadi, H., Nilashi, M., Ibrahim, O.: Organizational decision to adopt hospital information system: an empirical investigation in the case of Malaysian public hospitals. Int. J. Med. Inform. 84(3), 166–188 (2015) 2. Zakaria, N., Yusof, S.A.M.: Understanding technology and people issues in hospital information system (HIS) adoption: case study of a tertiary hospital in Malaysia. J. Infect. Public Health 9(6), 774–780 (2016) 3. Cantiello, J., Kitsantas, P., Moncada, S., Abdul, S.: The evolution of quality improvement in healthcare: patient-centered care and health information technology applications. J. Hosp. Adm. 5(2), 62–68 (2016) 4. Wijaya, E., Sulistyowati, N.: The effect of application of hospital information systems on operational performance through user satisfaction. Eur. J. Bus. Manag. 11(36), 71–78 (2019) 5. Juhana, D., Manik, E., Febrinella, C., Sidharta, I.: Empirical study on patient satisfaction and patient loyalty on public hospital in Bandung, Indonesia. Int. J. Appl. Bus. Econ. Res. 13(6), 4305–4326 (2015) 6. Lubis, A.N., Lumbanraja, P., Lubis, R.R., Hasibuan, B.K.: A study of service quality, corporate social responsibility, hospital image, and hospital value creation in Medan. Eur. Res. Stud. 20(4B), 125–133 (2017) 7. Meesala, A., Paul, J.: Service quality, consumer satisfaction and loyalty in hospitals: thinking for the future. J. Retail. Consum. Serv. 40, 261–269 (2018) 8. Asnawi, A., Awang, Z., Afthanorhan, A., Mohamad, M., Karim, F.: The influence of hospital image and service quality on patients’ satisfaction and loyalty. Manag. Sci. Lett. 9(6), 911–920 (2019) 9. Rahmadita, A., Yanuar, F., Devianto, D.: The construction of patient loyalty model using Bayesian structural equation modeling approach. CAUCHY 5(2), 73–79 (2018) 10. Tosyali, H., Sütcü, C.S., Tosyali, F.: Patient loyalty in the hospital patient relationship: the mediating role of social media. Erciyes˙Ileti¸simDergisi 6(1), 783–804 (2019) 11. Handayani, P.W., Hidayanto, A.N., Pinem, A.A., Hapsari, I.C., Sandhyaduhita, P.I., Budi, I.: Acceptance model of a hospital information system. Int. J. Med. Inform. 99, 11–28 (2017)

A Modified UTAUT Model for Hospital Information Systems

215

12. Narattharaksa, K.C., Speece, M.: Vendor relations and implementation of health IT projects. https://www.researchgate.net/profile/Mark_Speece2/publication/292988984_Vendor_relati ons_and_implementation_of_health_IT_projects/links/56b49ad008ae922e6c020216.pdf. Accessed 16 Aug 2020 13. Shahzad, K., Jianqiu, Z., Zia, M.A., Shaheen, A., Sardar, T.: Essential factors for adoption hospital information system: a case study from Pakistan. Int. J. Comput. Appl. 1–12 (2018) 14. Venkatesh, V., Morris, M.G., Davis, G.B., Davis, F.D.: User acceptance of information technology: toward a unified view. MIS Q. 27, 425–478 (2003) 15. Venkatesh, V., Thong, J., Xu, X.: Consumer acceptance and use of information technology: extending the unified theory of acceptance and use of technology. MIS Q. 36(1), 157–178 (2012) 16. Venkatesh, V., Sykes, T.A., Zhang, X.: ‘Just what the doctor ordered’: a revised UTAUT for EMR system adoption and use by doctors. In: 2011 44th Hawaii International Conference on System Sciences, p. 10. IEEE, January 2011 17. Farzandipour, M., Meidani, Z., Gilasi, H., Dehghan, R.: Evaluation of key capabilities for hospital information system: a milestone for meaningful use of information technology. Ann. Trop. Med. Public Health 10(6), 1579 (2017) 18. Roesems-Kerremans, G.: Big data in healthcare. J. Healthcare Commun. 1(4), 33 (2016) 19. Wang, Y., Kung, L., Byrd, T.A.: Big data analytics: understanding its capabilities and potential benefits for healthcare organizations. Technol. Forecast. Soc. Chang. 126, 3–13 (2018) 20. Hertzum, M., Simonsen, J.: Configuring information systems and work practices for each other: what competences are needed locally? Int. J. Hum. Comput. Stud. 122, 242–255 (2019) 21. Bawack, R.E., Kamdjoug, J.R.K.: Adequacy of UTAUT in clinician adoption of health information systems in developing countries: the case of Cameroon. Int. J. Med. Inform. 109, 15–22 (2018) 22. Malik, S.A., Nordin, A., Al-Ehaidib, R.N.: Requirements engineering (RE) process for the adaptation of the hospital information system (HIS). Int. J. Adv. Sci. Eng. Inf. Technol. 9(1), 8–17 (2019) 23. Bezboruah, K.C., Hamann, D.: Health IT adoption in nursing homes: the role of IT vendors. Int. J. Innov. Technol. Manag. 15(01), 1850001 (2018) 24. Abramson, E.L., Edwards, A., Silver, M., Kaushai, R.: Trending health information technology adoption among New York nursing homes. Am. J. Managed Care 20(11 Spec No. 17), eSP53-9 (2014) 25. Holmgren, A.J., Patel, V., Charles, D., Adler-Milstein, J.: US hospital engagement in core domains of interoperability. Am. J. Manag. Care 22(12), e395–e402 (2016) 26. Liebe, J.D., Esdar, M., Hübner, U.: Measuring the availability of electronic patient data across the hospital and throughout selected clinical workflows. Stud. Health Technol. Inform. 253, 99–103 (2018) 27. Hurst, K., Kreckler, S., Handa, A.: Improving information availability in vascular surgical clinics. A service evaluation and improvement project. BMJ Open Qual. 5(1), u210012– w4177 (2016) 28. Kranz, A.M., Dalton, S., Damberg, C., Timbie, J.W.: Using health IT to coordinate care and improve quality in safety-net clinics. Joint Comm. J. Qual. Patient Saf. 44(12), 731–740 (2018) 29. Jewer, J.: Patients intention to use online postings of ED wait times: a modified UTAUT model. Int. J. Med. Inform. 112, 34–39 (2018) 30. Alam, M.Z., Hu, W., Barua, Z.: Using the UTAUT model to determine factors affecting acceptance and use of mobile health (mHealth) services in Bangladesh. J. Stud. Soc. Sci. 17(2), 137–172 (2018)

216

S. N. R. Alismaili et al.

31. Cimperman, M., Brenˇciˇc, M.M., Trkman, P.: Analyzing older users’ home telehealth services acceptance behavior—applying an extended UTAUT model. Int. J. Med. Inform. 90, 22–31 (2016) 32. Sharifian, R., Askarian, F., Nematolahi, M., Farhadi, P.: Factors influencing nurses’ acceptance of hospital information systems in Iran: application of the unified theory of acceptance and use of technology. Health Inf. Manag. J. 43(3), 23–28 (2014) 33. Williams, M.D., Rana, N., Dwivedi, Y.K.: The unified theory of acceptance and use of technology (UTAUT): a literature review. J. Enterp. Inf. Manag. 28(3), 443–488 (2015) 34. Ahlan, A.R., Ahmad, B.I.E.: User acceptance of health information technology (HIT) in developing countries: a conceptual model. Procedia Technol. 16, 1287–1296 (2014) 35. Zhou, L.L., Owusu-Marfo, J., Antwi, H.A., Antwi, M.O., Kachie, A.D.T., Ampon-Wireko, S.: Assessment of the social influence and facilitating conditions that support nurses’ adoption of hospital electronic information management systems (HEIMS) in Ghana using the unified theory of acceptance and use of technology (UTAUT) model. BMC Med. Inform. Decis. Mak. 19(1), 230 (2019) 36. Ahlan, A.R., Ahmad, B.I.E.: An overview of patient acceptance of health information technology in developing countries: a review and conceptual model. Int. J. Inf. Syst. Project Manag. 3(1), 29–48 (2015) 37. Antwi, H.A., Yiranbon, E., Lulin, Z., Maxwell, B.A., Agebase, A.J., Yaw, N.E., Vakalalabure, T.T.: Innovation diffusion among healthcare workforce: analysis of adoption and use of medical ICT in Ghanaian tertiary hospitals. Int. J. Acad. Res. Bus. Soc. Sci. 4(7), 63 (2014) 38. Duarte, J.G., Azevedo, R.S.: Electronic health record in the internal medicine clinic of a Brazilian university hospital: expectations and satisfaction of physicians and patients. Int. J. Med. Inform. 102, 80–86 (2017)

Teamwork Communication in Healthcare: An Instrument (Questionnaire) Validation Process Wasef Matar1(B) and Monther Aldwair2 1 University of Petra, Amman, Jordan

[email protected] 2 College of Technological Innovation, Zayed University, Abu Dhabi, UAE

[email protected]

Abstract. Healthcare face many problems, one of these problems is embodied in teamwork communication systems, the current HISs lack of teamwork communication tools. To introduce a teamwork communication instrument (questionnaire) in healthcare which plays a key role in health information system area. The proposed a research model for this study applied a quantitative approach using a survey method. To formulate the problem a preliminary data was collected by survey method to test and introduce a validated instrument (questionnaire). This study proposed and validated an instrument (questionnaire) to be used in healthcare teamwork communication studies. The findings of this study will be contributed to teamwork communication in healthcare and will be a reference for any healthcare communication related study. This study is the first of its kind in Jordan and has added a new dimension in the teamwork communication in healthcare. Keywords: Teamwork communication · Clinical Pathways · Instrument (questionnaire) · Communication tools

1 Introduction Teamwork communication in healthcare is an important process to prevent medical errors. According to the report of the Institute of Medicine (IOM), 70% of medical errors are related to a teamwork communication, and 30% related to other factors. The current Health Information Systems (HISs) is lack of supporting teamwork communication among medical staff. The term “To err is human” has been applied to medical errors due to diagnostics, treatment, prevention and others. In the United States, 30%–45% of patients do not receive appropriate health care [1]. Effective use of HIS has improved quality of treatment, improved patient safety, better team climate, and better clinical outcomes. Clinical Pathways may improve teamwork communications [2, 3], Clinical Pathways is a system developed to apply patientcentered approach. Clinical Pathways is defined as “A complex intervention for the mutual decision making and organization of predictable care for a well-defined group of patients during a well-defined period.” (E-P-A, www.E-P-A.org). Clinical Pathways © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 217–229, 2021. https://doi.org/10.1007/978-3-030-70713-2_22

218

W. Matar and M. Aldwair

support teamwork communication, improve healthcare quality and reduce the cost [4–6], thus, this system plays a key role in quality of healthcare. By implementing CP, healthcare teamwork communication will be enhanced and improved [4] in Jordan. Teamwork communication in healthcare has become a critical issue in healthcare sector. Hence, there is a pressing need to investigate this issue. Based on the literature and the best of researcher knowledge, there is a lack of validated instrument and/or questionnaire to measure the teamwork communication in healthcare. Therefore, the aim of this paper is two folds; first, to develop an instrument (questionnaire) which based on a developed model to implement electronic Clinical Pathways (CP). Second aim is to validate the developed instrument (questionnaire) to be a reference in teamwork communication in healthcare research. The proposed and validated instrument (questionnaire) is based on the model in Fig. 3. The proposed model and the instrument (questionnaire) were tested in healthcare sector in Jordan, which is based on implementing Clinical pathways system. The study was conducted on two University’s hospitals use the computerized systems for healthcare. According to the researchers’ inspections for these two hospitals, it was found that they do not apply the Clinical Pathways. Moreover, it was found that they still use mobile phones and emails for communication rather than using HIS. Comparatively, these two types of communications have disadvantages and lack in providing information and support communications than the computerized information system. Thus, it is necessary to implement an effective communication system such as Clinical Pathways to improve teamwork communication based on the proposed model.

2 Literature Review The related work on Information Systems (IS) theories and teamwork communication models were reviewed and investigated to develop an instrument (questionnaire) to be a reference in teamwork communications in healthcare. This research study attempts to validate the proposed model of CP and an instrument (questionnaire) for healthcare teamwork communications. Many of HISs have failed around the world, and very few models integrate electronic clinical pathways in HISs [7, 8]. In addition, there is a trend to improve communication and decision making in HISs, which has not yet been achieved [9]. Therefore, there is a need for a model for successful communication and decision making, and a valid instrument (questionnaire) to measure the model in healthcare communications to be a reference. The current HISs was designed and developed for administrative purposes, thus there is a need to differentiate between the administrative and clinical processes. Most HISs in the hospitals and medical centers supports the administrative processes, the involve, but to a limited extent, they embrace some clinical processes that can support teamwork communication. Consequently, treatment processes, communication and coordination of teamwork does not supported in the applied HISs [10–13]. There are two approaches in HISs, patient-centered and disease-centered. In disease-centered approach, the patients are treated individually, without considering other circumstances and treatments. To develop and improve the current HISs, There is a necessity to switch from diseasecentered to patient-centered approach to treat the patient’s case as a complete instead of

Teamwork Communication in Healthcare: An Instrument (Questionnaire)

219

isolated diseases [10, 11, 14, 15]. Disease-centered approach supports the administrative process, but this approach is poor in supporting clinical processes; there is a need for a system supporting the clinical processes. Teamwork activities are an essential process which lack in disease-centered approach, especially communication, and lacks information on the flow of treatment process. In addition, in patient-centered approach, the patients are treated by considering all his diseases as a whole, not in isolation from other illnesses. The aim of patient-centered approach is to improve the healthcare quality and to decrease the medical errors. Patient-centered approach supports teamwork activities, communication and coordination are the main activities for teamwork. Patient-centered approach has two key requirements, teamwork communication and care coordination [10, 11, 14, 16, 17]. This study has reviewed information system studies and teamwork communications based on implementing Clinical Pathways as a clinical process. There is a lack of studies on Clinical Pathways from information systems perspective given the problematic history of HISs issues [18], and a lack of specific details and failure in HISs implementations, and we surveyed the issues related to HISs and Clinical Pathways [7, 19]. In previous studies have not identified the importance of Clinical Pathways as a communication tool [2, 3]. In addition, there is no validated healthcare communication instrument (questionnaire) to test the model in Fig. 3 [10].

3 Model As mentioned before that the aim of this study is to develop and validate an instrument (questionnaire) for teamwork communication in healthcare. Therefore, Socio-technical theory and Donabedian model were integrated to support the propose model to enhance teamwork communication. There is a need for a valid instrument (questionnaire) to test the proposed model. And to be a reference in healthcare communications discipline by implementing electronic Clinical Pathways. Social and technical aspects are the two subsystems for socio-technical theory which depict on Fig. 1, the two subsystems interact with each other with a relationship, every aspect support the other aspect by the interrelated relationship. Donabedian model depict in Fig. 2, this model has three dimensions (structure, process, and outcomes). The main objective for this model is to enhance healthcare system’s quality in general, and to be as reference in improving models and frameworks in healthcare in particular [20]. For this study, we develop an instrument (questionnaire) to be used in healthcare to enhance teamwork communication. Socio-technical theory is a recommended theory by many researchers in healthcare to light the requirements of the 21st century [21–23]. Healthcare has a set of characteristics and these characteristics are seen in three dimensions. The nature of work is one these characteristics that is viewed based on socio-technical theory. Based on the literature Clinical Pathways used as a tool to support teamwork communications can be seen from the perspective of socio-technical theory [24]. Integrate the human factors and healthcare quality models can improve and add to performance, healthcare quality and patient safety [25]. Socio-technical has two aspects (systems) without factors or dimensions. Based on

220

W. Matar and M. Aldwair

this, there is no related questions or a validated instrument (questionnaire) were found to test the model. There is another synonym for socio-technical which called socio-technical systems (STS) approach which is integrated of technical and social factors should be considered in organizational systems design [26, 27]. The STS approach eases a better understanding of how social factors affect of technical systems usage. Figure 1 presents socio-technical theory.

Social tem

Sys-

Structure

People

Technical tem

Sys-

Technology

MIS (Direct)

Process

Fig. 1. Socio-technical theory

Donabedian Model has a set of factors that are useful in implementing and designing HIS [20]. Donabedian model addresses the structure and the process for interdisciplinary teamwork model and provides a set of factors that support teamwork structure and teamwork process that lead to support implementation and designing HIS. Moreover, this model provides a solution for teamwork communication issues such as lack of information systems that support treatment flow, shared care requires, etc. Figure 2 presents the Donabedian model. This model was tested based on qualitative method, in this paper, the model would have been tested based on quantitative approach.

Structure

Process Fig. 2. Donabedian model

Outcome

Teamwork Communication in Healthcare: An Instrument (Questionnaire)

221

Validation and testing of the instrument (questionnaire) are main objective of this study in the context of teamwork communication in healthcare utilizing socio-technical theory in Jordan. Previous studies in healthcare communication consider the communication between physicians and nurses by using SBAR (Situation, Background, Assessment, and Recommendation). Previous research did not consider the communication among physicians and nurses and between them based on Clinical Pathways system [28]. Therefore, the current instrument (questionnaire) are based on such aspects. In this study provides an instrument (questionnaire) that is based on electronic Clinical Pathways (CP). The following is the model to test and validated the instrument that has been developed for teamwork communication in healthcare sector.

Communication Protocols

Social Factor

Internal Communication External Communication Teamwork Structure

Enhanced Teamwork Communication in Healthcare

Technical Factor Care Planning Disease Planning Discharge Planning Information Sharing Information Exchange

Fig. 3. Research model

222

W. Matar and M. Aldwair

4 Discussion The main objective of this study is to propose and validate an instrument (questionnaire) to be a reference in healthcare teamwork communications, this instrument (questionnaire) is based socio-technical theory and validated through the model mentioned in Fig. 3. The main contribution of this study is the establishment of a validated developed instrument (questionnaire) based on a review of literature in teamwork communication in healthcare. The instrument includes 48 items to measure 10 factors. The only items excluded from developed instrument; (CP1, CP2) from communication protocols, (EC1) from external communication, (IE3, IE6) from information exchange. This research contributes to the area by developing and validating an instrument (questionnaire) using as a sample of physicians and nurses in Jordanian hospitals. HIS in Jordanian hospitals is needed to be successful and there is a need to develop its functions to support teamwork communication among medical staff. Without validated instrument and responses from the users of HIS system this process yields misleading results. This study, therefore, overcome the shortcomings of the instrument in HIS with the model from the responses of nurse and physicians from two hospitals in Jordan. However, and despite the limitations, many directions are highlighted in this paper as future work such as the use of the instrument in doing and enhance the research on healthcare teamwork communications, which are likely to enhance the research and studies in not only on communication by implementing electronic pathways but in teamwork communication in general.

5 Conclusion The developed instrument (questionnaire) in this work opens the door for researchers to explore teamwork communication in healthcare. In addition, this instrument (questionnaire) is the first step to build a block that can contribute to teamwork communication in healthcare and other domains. To generalize the findings of this research, more research is encouraged in other hospitals in Jordan and in developing countries. HIS in Jordan is a new technology that emerged from the needs of healthcare organizations to better serve their patients and to improve healthcare quality and prevent medical errors by enhancing teamwork communications between physicians and nurses. Limitation of this study is that this study only considers physicians and nurses as main users of Clinical Pathways. Future research can consider top management and communication between medical staff and patients.

6 Appendix A

Teamwork Communication in Healthcare: An Instrument (Questionnaire) Internal communication

Coming from

Using information Process planning technology will improve internal communication and coordination

223

Reference

Used in

[29]

Process-oriented model 1–5

Scale

[30]

Self-categorization theory

Using information technology will strengthen care processes planning Using information technology will enable your hospital to adopt new organizational structures Using information technology will improve decision-making Using information technology will streamline care process External communication External communication

1–5

Using information technology facilitates members to communicate to other members during off time hours Using information technology facilitates to get all information from other team members during off time hours Using information technology makes it easy for members to communicate during off time hours Using information technology facilitates communication with other agencies (continued)

224

W. Matar and M. Aldwair

(continued) Internal communication

Coming from

Reference

Used in

Scale

Communication

[31]

TeamSTEPPS framework

1–5

Communication protocols

[32]

This study

Formalization

[33, 34]

TeamSTEPPS framework

Using information technology facilitates streamlining of clinical processes during off time hours Communication protocols Using specific terminologies in communication will improve patient care Staff follows a standardized method of sharing information when handing off patients Team member document and verify information that they receive from one another Using briefing (surgical checklist) will improve patient care Using debriefing (document treatment method) will improve patients care Using method of SBAR (Situation, Background, Assessment, and Recommendation) will improve patient care Teamwork structure

1–5

Hospital rules and regulations effect teamwork structure composition (continued)

Teamwork Communication in Healthcare: An Instrument (Questionnaire)

225

(continued) Internal communication

Coming from

Reference

Used in

Scale

Information sharing

[35]

Socio-technical theory emergency system

1–5

Care planning

[36, 20]

Donabedian model

1–5

Using information technology will improve the structure and composition of teamwork Using information technology will improve leadership in teamwork structure composition Using information technology will improve care coordination in teamwork structure composition Using information technology will improve communication in teamwork structure composition Information sharing Using information sharing enables the hospital to work with other agencies cooperatively Using information sharing enables better allocation of the resources Using information sharing among physicians should be timely Using information sharing during off time hours should be timely Care planning Hospital’s policies have positive effect on process of care planning (continued)

226

W. Matar and M. Aldwair

(continued) Internal communication

Coming from

Reference

Used in

Scale

Information exchange

[37, 38]

Donabedian model

1–5

Discharge planning

[20, 39]

Donabedian model

1–5

Standardized care processes will improve care planning Clinical guidelines will improve the structure of care planning Using information technology will improve documentation care planning Information technology will improve administrative procedures Information exchange Physicians benefit from exchanging and combining ideas with one another Physicians believe that by exchanging ideas they can improve healthcare quality Nurses believe that by exchanging ideas they can improve healthcare quality Hospital policies facilitate information exchange My hospital has a standardized system for information exchange Discharge planning Using information technology facilitates pre-discharge instruction Using information technology facilitates discharge planning process (continued)

Teamwork Communication in Healthcare: An Instrument (Questionnaire)

227

(continued) Internal communication

Coming from

Reference

Used in

Scale

Disease planning

[40]

Donabedian model

1–5

Effective emergency system

[35]

This study

1–5

Using information technology will improve coordination with some governmental agencies Using information technology facilitates patients to reintegrate into the community Disease planning Using information technology will improve disease planning process Using information technology will improve evaluation of disease planning process Using information technology will improve physiological sessions that are needed to reduce potential disease Other therapies are needed to manage diseases and it should be planned Teamwork communication enhancement Using information technology will improve teamwork communication quality Using information technology will provide communication in a timely manner Using information technology will streamline the care plan Using information technology will improve quality of care

228

W. Matar and M. Aldwair

Acknowledgements. This work was supported in part by Zayed University Research Office, Research Cluster Award #R18054.

References 1. Gaddis, G.M., Greenwald, P., Huckson, S.: Toward improved implementation of evidencebased clinical algorithms: clinical practice guidelines, clinical decision rules, and clinical pathways. Acad. Emerg. Med. J. 14(11), 1015–1022 (2007) 2. Abu-Rish Blakeney, E., et al.: Purposeful interprofessional team intervention improves relational coordination among advanced heart failure care teams. J. Interprof. Care 33(5), 481–489 (2019) 3. Lawal, A.K., et al.: Development of a program theory for clinical pathways in hospitals: protocol for a realist review. Syst. Rev. 8(1), 1–7 (2019) 4. Deneckere, S., et al.: Care pathways lead to better teamwork: results of a systematic review. Soc. Sci. Med. 75(2), 264–268 (2012) 5. Adeyemi, S., Demir, E., Chaussalet, T.: Towards an evidence-based decision making healthcare system management: modelling patient pathways to improve clinical outcomes. Decis. Support Syst. 55(1), 117–125 (2013) 6. Shoji, F., et al.: Assessing a clinical pathway to improve the quality of care in pulmonary resections. Surg. Today 41(6), 787–790 (2011) 7. Logan, J.: Electronic health information system implementation models-a review. Stud. Health Technol. Inform. 178, 117–123 (2012) 8. Sherifi, D.: FITT Model, in Portable Health Records in a Mobile Society, pp. 199–208. Springer (2019) 9. Sacchi, L., et al.: Personalization and patient involvement in decision support systems: current trends. Yearb. Med. Inform. 10(1), 106 (2015) 10. ALsalamah, H.: Supporting integrated care pathways with workflow technology. Cardiff University (2012) 11. Ward, M.M., et al.: Clinical information system availability and use in urban and rural hospitals. J. Med. Syst. 30(6), 429–438 (2006) 12. Lenz, R., Reichert, M.: IT support for healthcare processes – premises, challenges, perspectives. Data Knowl. Eng. 61(1), 39–58 (2007) 13. Swenson, K.D., Palmer, N., Silver, B.: Taming the Unpredictable: Real World Adaptive Case Management: Case Studies and Practical Guidance. Future Strategies Inc. (2011) 14. Executive, N.: The NHS Plan: a plan for investment, a plan for reform, in London: Department of Health. Colegate (2000) 15. Marjanovic, S., et al.: Innovating for improved healthcare: sociotechnical and innovation systems perspectives and lessons from the NHS. Sci. Public Policy 47(2), 283–297 (2020) 16. Majidi, M., Mahdavi, H., Siamian, H.: Patients’ information needs in affiliated hospitals of Tehran University of Medical Sciences, vol. 723 (2012) 17. Luxford, K., Safran, D.G., Delbanco, T.: Promoting patient-centered care: a qualitative study of facilitators and barriers in healthcare organizations with a reputation for improving the patient experience. Int. J. Qual. Health Care 23(5), 510–515 (2011) 18. Ibrahim, W.M.R.: Situation analysis for clinical pathways and teamwork communication in healthcare. Indian J. Sci. Technol. 9, 28 (2016) 19. Robert, G., et al.: Organisational factors influencing technology adoption and assimilation in the NHS: a systematic literature review. Report for the National Institute for Health Research Service Delivery and Organisation programme (2009)

Teamwork Communication in Healthcare: An Instrument (Questionnaire)

229

20. Kuziemsky, C.E., et al.: An interdisciplinary team communication framework and its application to healthcare ‘e-teams’ systems design. BMC Med. Inform. Decis. Mak. 9(1), 43 (2009) 21. Sittig, D.F.: A socio-technical model of health information technology-related e-iatrogenesis. In: AMIA Annual Symposium Proceedings (2008) 22. Nøhr, C., Aarts, J.E.: Information Technology in Health Care: Socio-Technical Approaches 2010: From Safe Systems to Patient Safety, vol. 157. IOS Press, Amsterdam (2010) 23. Aarts, J., et al.: Information technology in health care: socio-technical approaches. Int. J. Med. Inform. 79(6), 389–390 (2010) 24. Berg, M.: Patient care information systems and health care work: a sociotechnical approach. Int. J. Med. Inform. 55(2), 87–101 (1999) 25. Carayon, P., et al.: Work system design for patient safety: the SEIPS model. Qual. Saf. Health Care 15(suppl 1), i50–i58 (2006) 26. Baxter, G., Sommerville, I.: Socio-technical systems: from design methods to systems engineering. Interact. Comput. 23(1), 4–17 (2011) 27. Bostrom, R.P., Heinen, J.S.: MIS problems and failures: a socio-technical perspective, Part II: the application of socio-technical theory. MIS Q. 1, 11–28 (1977) 28. Thomas, C.M., Bertram, E., Johnson, D.: The SBAR communication technique: teaching nursing students professional communication skills. Nurse Educ. 34(4), 176–180 (2009) 29. Van Der Vegt, G.S., Bunderson, J.S.: Learning and performance in multidisciplinary teams: the importance of collective team identification. Acad. Manag. J. 48(3), 532–547 (2005) 30. Tallon, P.P., Kraemer, K.L., Gurbaxani, V.: Executives’ perceptions of the business value of information technology: a process-oriented approach. J. Manag. Inf. Syst. 16, 145–173 (2000) 31. Battles, J., King, H.B.: TeamSTEPPS Teamwork Perceptions Questionnaire (T-TPQ) Manual. American Institute for Research, Washington, DC (2010) 32. Lisha Lo, M.: Teamwork and Communication in Healthcare, p. 68. Canadian Patient Safety Institute Edmonton (2011) 33. Hirst, G., et al.: How does bureaucracy impact individual creativity? A cross-level investigation of team contextual influences on goal orientation–creativity relationships. Acad. Manag. J. 54(3), 624–641 (2011) 34. Hamilton, H., Chou, W.-y.S.: The Routledge Handbook of Language and Health Communication, 1st edn. Routledge, New York (2014) 35. Kim, M., et al.: Assessing roles of people, technology and structure in emergency management systems: a public sector perspective. Behav. Inf. Technol. 31(12), 1147–1160 (2012) 36. Segars, A.H., Grover, V.: Profiles of strategic information systems planning. Inf. Syst. Res. 10(3), 199–232 (1999) 37. Collins, C.J., Smith, K.G.: Knowledge exchange and combination: the role of human resource practices in the performance of high-technology firms. Acad. Manag. J. 49(3), 544–560 (2006) 38. Gold, A.H., Arvind Malhotra, A.H.S.: Knowledge management: an organizational capabilities perspective. J. Manag. Inf. Syst. 18(1), 185–214 (2001) 39. Dai, Y.T., et al.: Effectiveness of a pilot project of discharge planning in Taiwan. Res. Nurs. Health 26(1), 53–63 (2003) 40. Glasgow, R.E., et al.: The chronic illness resources survey: cross-validation and sensitivity to intervention. Health Educ. Res. 20(4), 402–409 (2005)

Potential Benefits of Social Media to Healthcare: A Systematic Literature Review Ghada Ahmad Abdelguiom1(B) and Noorminshah A. Iahad2 1 School of Computing, Faculty of Engineering, Universiti Teknologi Malaysia, Johor Bahru,

Malaysia 2 Azman Hashim International Business School (Information Systems),

Universiti Teknologi Malaysia, Johor Bahru, Malaysia [email protected]

Abstract. Social media offers a rich online experience, dynamic content, usability, and knowledge that attracts more users. The use of social media in the health sector is indeed attracting more and more attention. Over the last ten years, researchers have attempted various topics related to the health sector via social media that have contributed beneficially to the healthcare domain. There is a demand for a study to identify the potential benefits of social media to healthcare. Thus, this paper surveys research papers related to the social media platform in the healthcare domain that were published between the years (2014–2020). The primary objective of this study is to review the range, nature, and extent of current research activity on the role of social media in healthcare. Therefore, this paper outlines the recent approaches to the utilization of social media to provide solutions for health-related issues. Also, it discusses the role of social media in promoting health care services. The study addresses the key issues addressed in the latest research, provides an overview of their shortcomings, restrictions, and finally, outlines the opportunities for future research. Keywords: Health care · Social networks · Digital communication · Social media · e-Health · Web 2.0

1 Introduction Social Media (SM) platforms can be described as a community of Internet-based applications that build on the Web 2.0’s that enable the creation and sharing of content that has been created by users (Carr and Hayes 2015). The new approach in which it provides, however, also promotes discussion and identification of medical problems (Naslund et al. 2016). The social media’s idea of offering immediate, direct input to the members allows the site to play an enticing and exciting role in delivering relevant content to promote a healthier lifestyle (Chung et al. 2017). SM has been very cautious to share information during training, practice and surveys and emergencies from healthcare providers and public health practitioners (Vaterlaus et al. 2015). The emergence of Internet of things © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 230–241, 2021. https://doi.org/10.1007/978-3-030-70713-2_23

Potential Benefits of Social Media to Healthcare

231

lead to development in the field of healthcare. The available resources or studies lacks generation of awareness to people to utilize SM to interact with medical professionals and avail better treatment. One of its main features is that it enables members who cannot communicate with their doctors about certain medical problems and want to utilize web applications. They are not free to contact them. On the other hand, SM platforms can potentially be used to establish medication coherence and improve compliance with medications, training, and patient care programs, including topics related to sexual health, alcohol, and drug addiction. This can also be used to encourage institutional patient aid, support groups for patients, facilitate institutional loyalty, increase overall interaction between the physician and the patient, and enhance general physician communication (Lee et al. 2016; Sarker et al. 2016). The use of SM in a targeted way may contribute to a stronger quality education for patients, which in turn allows them to better understand their disease predictions and their respective treatment options. In conjunction with highlighting the latest treatments, facilities, installations that can be used when needed (Glover et al. 2015) Sensitive medical providers do need to see the value of participating, cooperation, and encouragement to chronic diseases such as diabetes by the inclusion of different SM (Knight et al. 2015; Santoro et al. 2015). They are seen as interesting tools that offer the medical community enormous benefits to help health professionals deliver adequate service and practice. The literary analysis framework for narrative synthesis was used for evaluating the current evidence (McCaughey et al. 2014; Merolli et al. 2015; Panahi et al. 2016; Xu et al. 2016). We searched for indexed scientific literature using healthcare and SM keywords. The articles included those that analyzed SM application in the area of healthcare. We analyzed the studies based on the criteria such as user retention, acceptance, level of participation, and the findings of SM efficacy in the health sector. In the last nine years, public engagement in SM has risen dramatically (Allington et al. 2020).In the USA, there has been a rise from 8% to 72% for adults using SM since 2005 (Depoux et al. 2020; Tursunbayeva et al. 2017). In 2012 Facebook users have more than one billion people worldwide, representing a seventh of the world’s population (Wyche and Baumer 2017). SM is widely distributed among all ages and occupations and is widespread across the globe. Moreover, 100 million active Twitter users send over 65 million Twitter messages every day, and two billion YouTube videos (Alhabash and Ma 2017). SM have been related to major political events such as the Arab Spring revolution, as well as to common developments in culture, including the decline of public attention and drop in print media (AlSayyad and Guvenc 2015). The intention of this review is to address the following Research Question (RQ). RQ1: What are the existing research articles that presents the application of SM in healthcare. RQ2: How to classify the Healthcare Professionals (HCP) and non-HCP. RQ3: What are the opportunities available to connect both HCP and non-HCP.

232

G. A. Abdelguiom and N. A. Iahad

The objective of this study to extract literature related to the role of SM in the health care domain. In addition, it presents the available methods for connecting both HCP and non-HCP. This paper is organized as follows. The information regarding the benefits of SM and its application in health care domain is discussed in the first section. The second section provides information about methods for reviewing the literature with relevant decisions on selecting studies for consideration. The results of this review are discussed in the third and fourth sections. Finally, the review is concluded with future direction.

2 Methodology The narrative synthesis approach was used to review and analyze the existing evidence as presented in the related literature to usage of SM platforms in health-related issues. The databases includes Medline, Web of Science, Emerald, and Applied Social Sciences Index and Abstract were searched using the keywords: Social Media, Digital Media, Web 2.0, Facebook, Internet, Blog, Twitter, Forum, content marketing, Wiki, Email, Health, medication, and Medical. Table 1 presents the inclusion and exclusion criteria employed in the study. We easily sorted the literature based on a title to rule out irrelevant papers and remove studies from medical professionals using social networking as an intervention or using social networking within closed support groups. Non-English literature has been excluded. Meanwhile, Google Scholar Search engine features were used to focus the search publication year (2014 to 2020). We used the following combination of keywords to extract studies. (“Social media’ OR “Social software” OR “Social network” OR “Facebook” OR “Twitter” OR “Youtube “’ OR “Digital media” OR “Social tagging” OR Wiki) AND (Healthcare domain). The papers included were those involving a considerable investigation on SM and different health themes as well as reporting the results of the intervention or detailing evidence-based plans. As such, the general papers about social networking and health were excluded. Table 1. Inclusion and exclusion criteria Inclusion

Exclusion

Complete research

In – complete or partial Research content

Published between the year 2014–2020

Articles published before the year 2014

Articles related to the objective

Irrelevant articles

Research articles published in English Language Research articles in non – English language

Studies focused on SM and/or health conditions. Online forums discuss various health issues. These studies monitor network activity over a given period, compile user interaction scenarios, and analyze trend content. Statistics suggest that participants

Potential Benefits of Social Media to Healthcare

233

appear to seek out social support networks and find solace from people with similar health problems. Having experience in the SM with others can lead to interactions that are less judgmental than in other social arenas. Individuals are willing to discuss socially very sensitive or embarrassing conditions openly, for example on an online men’s eating disorder forum (Moessner et al. 2018). An analysis of the social networking site, where a health problem is of a rather personal nature. We have also focused at how the online networking operation of users could lead to specific social changes and how they formed dimensions that could capture campaigning activities and crowd funding activities. RQ1 is addressed using Table 2 that illustrates the process of selection of research studies used in the study. Table 2. Selection process to identify research articles Steps

Number of papers

Extraction of studies from digital platforms

120

Removal of duplicates using Endnote

70

Manual selection of studies using Title and Abstract

50

Manual selection of studies based on the content of the research

49

3 Results It’s worthy to mention, as a general observation, that the number of articles published on the usage of SM in the health domain has expanded outstandingly in the span of the present review. Most literatures addressed the topic related to either perspective of health care providers or patients. 3.1 User Classification To address the RQ2, we classified the users in SM plays an important role in this modern environment. All over the world, people are communicating with the support of SM. The purpose of SM is to provide a platform to users to discuss about a common or specific issue. In this study, we have classified SM users into two major classification includes HealthCare Professional (HCP) and Non – HCP. The reason for the classification of users is to categorize the benefits of SM according to the users (Fig. 1). HCP. An HCP may provide health care treatment and advice based on formal training and experience. The field includes those who work as a physician, surgeon, physician assistants, nurse, physiotherapist, dentist, midwife, psychologist, psychiatrist, or pharmacist or who perform services in allied health professions. A health professional may also be a public health or community health practitioner. The following part of this section will facilitate the benefits of SM for each user under HCP category

234

G. A. Abdelguiom and N. A. Iahad

Fig. 1. Classification of users

i.

Doctors: Doctors or medical professionals, especially in the western world, are linked to their patients and peers through SM, including Twitter and Facebook. This allows them, in addition to answering questions on the health perspective, to engage with their patients on an individual basis and share ideas with their peers. ii. Nurses: Nurses use SM to interact at the personal and social level, and to monitor events related to public health. They rely on SM in all fields of practice to communicate and exchange knowledge with colleagues. iii. Counselors and Volunteers: Over recent years, SM platforms have imbued several facets of social and cultural life. Such tools, and Web 2.0 in general, allow interactivity, interoperability and collaboration and encourage user-generated health-related content to be created and shared in support of patients and lay users. The openness and participation are perceived to be the underlying logic of those technologies. There is an increasing amount of knowledge that shows that SM provides significant opportunities for the charitable and community sectors to improve participation. iv. Psychologist: Some of SM’s most enticing facets is that it’s an unobtrusive way for psychologists to analyze individual behavior. A new research indicates that people on SM prefer to act somewhat like oneself, rather than as an idealized version of themselves. It makes the information obtained from all these networking sites more reliable than anyone would think it is. Nevertheless, it should be noted that certain forms of studies, such as experimental research, are not quite conducive to SM sites.

Potential Benefits of Social Media to Healthcare

235

v. Trainers and Trainees: Trainers have initiated to use the SM channels to train trainees as part of their training process. Trainees are reinforced to use unique hashtags on Twitter, or to join other groups to participate in training sessions. This makes the training process more immersive and enjoyable. Those preparation strategies have a central place for trainees to ask questions and get prompt responses. SM empowers learners to provide direct feedback on training sessions to their trainers. Non – HCP. A Non-HCP is a health care recipient carried out by medical practitioners. The non-HCP is most frequently ill or injured and requires treatment from a physician, nurse, psychologist, or dentist. The utilization of SM by each Non – HCP is as follows: i.

Patient: Patients are becoming more active with the introduction of SM and have access to online health records. Patients’ use of SM has strengthened their interaction with their health professionals. ii. Health Seekers: The health seeker is mindful of any individual’s encouragement for better or worse health. Health seeking is the natural pursuit of one’s proper wellbeing balance, the continuous step toward one’s own core and the appreciation of “normal” health. iii. Lay users: Lay users are the ultimate beneficiaries of the developments in SM in everyday settings. It is easy for them to communicate the health service officials and improve their health.

4 Discussion The proliferation of social networks like Facebook makes it much harder to keep distance between the doctor and patient (Paul et al. 2016). SM improves the bond between the patient and the doctor with professionalism. RQ3 raises a question, will healthcare professionals consider a patient as a social network friend, and if not would refuse the patient to be misunderstood and create issues in the relationship? Many members of the medical community agree that it is unwise for healthcare professionals to communicate with patients on social networking platforms or various reasons, however, primarily to prevent obscuring the distinction between personal and professional life. SM members are encouraged to connect, find, and understand each other (Jordan et al. 2019). ‘Sharing’ means updating the online profile with all health-related medical history material, including disease progression and symptoms, and supporting therapy. Participants use their online health accounts to monitor their progress and may opt to share the details with others or even their physicians. “Search” refers to the search engine for the website that allows a member to locate people who have the same medical conditions. Using the medical profiles, searches can be reduced to diagnosis, geographical location, age, and gender, not just illness or symptoms. SNSs and patient decision-making help correct assessment methods should be used to determine the quality of healthcare interactions or opinions. Responsive healthcare providers have also started to see the value of engaging and collaborating with patients through various SM platforms to foster self-management

236

G. A. Abdelguiom and N. A. Iahad

support, achieve positive health outcomes, and empower patients with chronic disease such as diabetes (De Martino et al. 2017). There are empirical underpinnings which support the claim that the utilization of SM in the context of health-related support groups have more noteworthy results, particularly, to those in need of emotional support. It can connect the health care providers to patients’ perspectives and needs to enable and empower them to develop strategies of coping with their health concerns (Grajales III et al. 2014) Interestingly, SM proved highly effective in hard-to-reach clinical populations, such as patients with rare diseases it can offer information on those issues (Davies 2016; Jacobs et al. 2016). Interestingly, SM proved highly effective in hard-to-reach clinical populations, such as patients with rare diseases (Davies 2016; Jacobs et al. 2016). The most striking outcome is that SM plays a vital role in supporting public health practices by facilitating reaching and identifying the targeted population for an intervention in case of disease outbreaks. Additionally, it can be used to incorporate investigation, analysis, management, and surveillance of disease outbreak (Hossain et al. 2016; McGough et al. 2017). There is a consensus that the utilization of SM can be a means of data sources for public health surveillance. It can also provide complementary information on health situations. Not to mention the considerable impact on the detection of epidemic diseases through the sharing of users’ information (Aiello et al. 2020). A large number of publications have emphasized that SM can offer new instruments to health authorities to support decision-making on the global, domestic, regional, and corporate levels (Lewin et al. 2015).This technology channel may have a profound impact on the advancement of healthcare-related issues, such as safe eating habits, physical activity, and overweight prevention (Gamache-OLeary and Grant 2017). The study found that self-management has become simpler for people who want to share their medical information, because health information is clearly organized according to actual data. Persons were often drawn with the opportunity to locate other patients who matched the advice or reassurance on treatment or symptoms both demographically and medically. While patients may feel empowered by websites such as SM, because of their ability to quantify and understand the everyday decision they make about their health, there is also a commercial operation behind these tools which sell patients’ data to private companies that might not prioritize the same goals as patients (Myrick et al. 2016). Experts advocate that regardless if participants state they do not care who has access to their medical data, or how public it may be, the users are vulnerable to unintentional and intentional security threats and misuse (Omary 2018), raising discriminations in health care, employment, etc.

5 Conclusion The aim of the proposed systematic literature review is to examine, analyze, and survey the current researches to explore the influence of SM on the health domain then summarizes the current evidence of understanding of using SM in the health domain, and how this platforms can contribute to promoting inflow of delivering health services and describes the main topics being discussed among the articles included in this review.

Potential Benefits of Social Media to Healthcare

237

Likewise reviews the shortcomings of previous researches and recommends to the direction of the future research. Prior studies found that generally SM has opened new opportunities for the strengthening the empowerment model and for participatory healthcare, and Allows shared decision making between patients and healthcare providers which in turn can lead to autonomy, improved communication, greater self-efficacy, and patient satisfaction. One of the more significant findings to emerge from this study is that usage of SM has significant potential to enhance the well-being of patients with long and short-term ailments. The Researches must trend to the adoption of SM platforms in the health domain Because of it acceptable and promising. Utilize this conceivably great innovation as a useful health knowledge transfer mechanism. The significance of SM in the context of healthcare is clearly supported by the current findings that people who suffer from having poor health more likely to share health information from online sources. The researchers have highlighted pieces of evidence that health care services are productively utilize the SM as a primary tool to deliver messages regarding health information to people who are residing in economically overburdened (low-income) countries. In those countries, individuals are often helpless to immediate health-related access and patients also may not have the health- insurance. A greater focus on SM in countries that lack state-of-the-art healthcare systems could produce interesting findings that account more for may be used to facilitate patient-centric healthcare by involving the patient in fulfilling personal healthcare needs. The findings of this research provide insights into this platforms can promote a health awareness and empowering patients on sensitive health issues such as (homosexuals). Furthermore, it can be contributed viably to control and prevent of the non- communicable diseases that resulting of lack of awareness and which have a negative impact on the health economics of those countries that already suffer from financials risk., Future researches should focus on how health institutions in developing and low-income countries, can take advantage of the proliferation and growth rates of SM platforms for harness them for helping to establish healthy behaviors and the healthy lifestyle among citizen. The findings of this study have a number of various imperative ramifications implications for future practice given that the SM new open doors opportunities for the empowerment model and for participatory HealthCare. That Usage of SM in health could produce interesting findings that account more it can improve health literacy and selfmanagement of health at the individual level and increase the efficiency in the provision of health services at the institutional level.

References Alhabash, S., Ma, M.: A tale of four platforms: motivations and uses of Facebook, Twitter, Instagram, and Snapchat among college students? Soc. Media+ Soc. 3(1) (2017). https://doi.org/10. 1177/2056305117691544 Allington, D., Duffy, B., Wessely, S., Dhavan, N., Rubin, J.: Health-protective behaviour, social media usage and conspiracy belief during the COVID-19 public health emergency. Psychol. Med. 1–7 (2020) AlSayyad, N., Guvenc, M.: Virtual uprisings: on the interaction of new social media, traditional media coverage and urban space during the ‘Arab Spring.’ Urban Stud. 52(11), 2018–2034 (2015)

238

G. A. Abdelguiom and N. A. Iahad

Aschbrenner, K.A., Naslund, J.A., Shevenell, M., Kinney, E., Bartels, S.J.: A pilot study of a peergroup lifestyle intervention enhanced with mHealth technology and social media for adults with serious mental illness. J. Nerv. Ment. Dis. 204(6), 483 (2016) Carr, C.T., Hayes, R.A.: Social media: defining, developing, and divining. Atlantic J. Commun. 23(1), 46–65 (2015) Chung, A.E., Skinner, A.C., Hasty, S.E., Perrin, E.M.: Tweeting to health: a novel mHealth intervention using Fitbits and Twitter to foster healthy lifestyles. Clin. Pediatr. 56(1), 26–32 (2017) Dadgar, M., Joshi, K.D.: The role of information and communication technology in selfmanagement of chronic diseases: an empirical investigation through value sensitive design. J. Assoc. Inf. Syst. 19(2), 2 (2018) Dai, X., Bikdash, M., Meyer, B.: From social media to public health surveillance: word embedding based clustering method for Twitter classification. Paper presented at the SoutheastCon 2017 (2017) de Almeida Marques-Toledo, C., Degener, C.M., Vinhal, L., Coelho, G., Meira, W., Codeço, C.T., Teixeira, M.M.: Dengue prediction by the web: tweets are a useful tool for estimating and forecasting dengue at country and city level. PLoS Negl. Trop. Dis. 11(7), e0005729 (2017) De Martino, I., D’Apolito, R., McLawhorn, A.S., Fehring, K.A., Sculco, P.K., Gasparini, G.: Social media for patients: benefits and drawbacks. Curr. Rev. Musculoskelet. Med. 10(1), 141–145 (2017) Dekker, R., Engbersen, G., Klaver, J., Vonk, H.: Smart refugees: how Syrian asylum migrants use social media information in migration decision-making. Soc. Media+ Soc. 4(1) (2018). https:// doi.org/10.1177/2056305118764439 Deng, Z., Liu, S.: Understanding consumer health information-seeking behavior from the perspective of the risk perception attitude framework and social support in mobile social media websites. Int. J. Med. Inform. 105, 98–109 (2017) Depoux, A., Martin, S., Karafillakis, E., Preet, R., Wilder-Smith, A., Larson, H.: The pandemic of social media panic travels faster than the COVID-19 outbreak. Oxford University Press (2020) Fernández-Luque, L., Bau, T.: Health and social media: perfect storm of information. Healthc. Inform. Res. 21(2), 67–73 (2015) Fung, I.C.-H., Hao, Y., Cai, J., Ying, Y., Schaible, B.J., Yu, C.M., Tse, Z.T.H., Fu, K.-W.: Chinese social media reaction to information about 42 notifiable infectious diseases. Plos ONE 10(5), e0126092 (2015) Fung, I.C.-H., Tse, Z.T.H., Fu, K.-W.: The use of social media in public health surveillance. West. Pac. Surveill. Response J. WPSAR 6(2), 3 (2015) Gamache-OLeary, V., Grant, G.: Social media in healthcare. Paper presented at the Proceedings of the 50th Hawaii International Conference on System Sciences (2017) Glover, M., Khalilzadeh, O., Choy, G., Prabhakar, A.M., Pandharipande, P.V., Gazelle, G.S.: Hospital evaluations by social media: a comparative analysis of Facebook ratings among performance outliers. J. Gen. Intern. Med. 30(10), 1440–1446 (2015) Gore, R.J., Diallo, S., Padilla, J.: You are what you tweet: connecting the geographic variation in America’s obesity rate to Twitter content. PLoS ONE 10(9), e0133505 (2015) Grajales III, F.J., Sheps, S., Ho, K., Novak-Lauscher, H., Eysenbach, G.: Social media: a review and tutorial of applications in medicine and health care. J. Med. Internet Res. 16(2), e13 (2014) Gruver, R.S., Bishop-Gilyard, C.T., Lieberman, A., Gerdes, M., Virudachalam, S., Suh, A.W., Kalra, G.K., Magge, S.N., Shults, J., Schreiner, M.S.: A social media peer group intervention for mothers to prevent obesity and promote healthy growth from infancy: development and pilot trial. JMIR Res. Protoc. 5(3), e159 (2016) Hossain, L., Kam, D., Kong, F., Wigand, R., Bossomaier, T.: Social media in Ebola outbreak. Epidemiol. Infect. 144(10), 2136–2143 (2016)

Potential Benefits of Social Media to Healthcare

239

Jordan, S.E., Hovet, S.E., Fung, I.C.-H., Liang, H., Fu, K.-W., Tse, Z.T.H.: Using Twitter for public health surveillance from monitoring and prediction to public response. Data 4(1), 6 (2019) Kaufman, M.R., Cornish, F., Zimmerman, R.S., Johnson, B.T.: Health behavior change models for HIV prevention and AIDS care: practical recommendations for a multi-level approach. J. Acquir. Immune Defic. Syndr. 66(Suppl 3), S250 (2014) Keir, A., Bamat, N., Patel, R.M., Elkhateeb, O., Roland, D.: Utilising social media to educate and inform healthcare professionals, policy-makers and the broader community in evidence-based healthcare. BMJ Evid.-Based Med. 24(3), 87–89 (2019) Keles, B., McCrae, N., Grealish, A.: A systematic review: the influence of social media on depression, anxiety and psychological distress in adolescents. Int. J. Adolesc. Youth 25(1), 79–93 (2020) Kim, S.H., Utz, S.: Effectiveness of a social media-based, health literacy-sensitive diabetes selfmanagement intervention: a randomized controlled trial. J. Nurs. Scholarsh. 51(6), 661–669 (2019) Knight, E., Werstine, R.J., Rasmussen-Pennington, D.M., Fitzsimmons, D., Petrella, R.J.: Physical therapy 2.0: leveraging social media to engage patients in rehabilitation and health promotion. Phys. Ther. 95(3), 389–396 (2015) Lee, J.L., Choudhry, N.K., Wu, A.W., Matlin, O.S., Brennan, T.A., Shrank, W.H.: Patient use of email, Facebook, and physician websites to communicate with physicians: a national online survey of retail pharmacy users. J. Gen. Intern. Med. 31(1), 45–51 (2016) Lehmiller, J.J., Ioerger, M.: Social networking smartphone applications and sexual health outcomes among men who have sex with men. PLoS ONE 9(1), e86603 (2014) Lewin, S., Glenton, C., Munthe-Kaas, H., Carlsen, B., Colvin, C.J., Gülmezoglu, M., Noyes, J., Booth, A., Garside, R., Rashidian, A.: Using qualitative evidence in decision making for health and social interventions: an approach to assess confidence in findings from qualitative evidence syntheses (GRADE-CERQual). PLoS Med. 12(10), e1001895 (2015) Li, C., Chen, L.J., Chen, X., Zhang, M., Pang, C.P., Chen, H.: Retrospective analysis of the possibility of predicting the COVID-19 outbreak from Internet searches and social media data, China, 2020. Eurosurveillance 25(10), 2000199 (2020) Li, Y., Wang, X., Lin, X., Hajli, M.: Seeking and sharing health information on social media: a net valence model and cross-cultural comparison. Technol. Forecast. Soc. Change 126, 28–40 (2018) Lim, S., Tucker, C.S., Kumara, S.: An unsupervised machine learning model for discovering latent infectious diseases using social media data. J. Biomed. Inform. 66, 82–94 (2017) Liu, S., Young, S.D.: A survey of social media data analysis for physical activity surveillance. J. Forensic Leg. Med. 57, 33–36 (2018) McCaughey, D., Baumgardner, C., Gaudes, A., LaRochelle, D., Wu, K.J., Raichura, T.: Best practices in social media: utilizing a value matrix to assess social media’s impact on health care. Soc. Sci. Comput. Rev. 32(5), 575–589 (2014) McGough, S.F., Brownstein, J.S., Hawkins, J.B., Santillana, M.: Forecasting Zika incidence in the 2016 Latin America outbreak combining traditional disease surveillance with search, social media, and news report data. PLoS Negl. Trop. Dis. 11(1), e0005295 (2017) McGregor, F., Somner, J.E., Bourne, R.R., Munn-Giddings, C., Shah, P., Cross, V.: Social media use by patients with glaucoma: what can we learn? Ophthalmic Physiol. Opt. 34(1), 46–52 (2014) Merolli, M., Gray, K., Martin-Sanchez, F., Lopez-Campos, G.: Patient-reported outcomes and therapeutic affordances of social media: findings from a global online survey of people with chronic pain. J. Med. Internet Res. 17(1), e20 (2015) Moessner, M., Feldhege, J., Wolf, M., Bauer, S.: Analyzing big data in social media: text and network analyses of an eating disorder forum. Int. J. Eat. Disord. 51(7), 656–667 (2018)

240

G. A. Abdelguiom and N. A. Iahad

Mohanty, S., Leader, A.E., Gibeau, E., Johnson, C.: Using Facebook to reach adolescents for human papillomavirus (HPV) vaccination. Vaccine 36(40), 5955–5961 (2018) Munger, K., Bonneau, R., Nagler, J., Tucker, J.A.: Elites tweet to get feet off the streets: measuring regime social media strategies during protest. Polit. Sci. Res. Methods 7(4), 815–834 (2019) Nagar, R., Yuan, Q., Freifeld, C.C., Santillana, M., Nojima, A., Chunara, R., Brownstein, J.S.: A case study of the New York City 2012–2013 influenza season with daily geocoded Twitter data from temporal and spatiotemporal perspectives. J. Med. Internet Res. 16(10), e236 (2014) Naslund, J., Aschbrenner, K., Marsch, L., Bartels, S.: The future of mental health care: peer-to-peer support and social media. Epidemiol. Psychiatr. Sci. 25(2), 113–122 (2016) Oh, S.-H., Lee, S.Y., Han, C.: The effects of social media use on preventive behaviors during infectious disease outbreaks: the mediating role of self-relevant emotions and public risk perception. Health Commun. 1–10 (2020) Omary, R.A.: Social media and education in radiology: let’s start with why. Acad. Radiol. 25(6), 744–746 (2018) Panahi, S., Watson, J., Partridge, H.: Social media and physicians: exploring the benefits and challenges. Health Inform. J. 22(2), 99–112 (2016) Paul, M.J., Sarker, A., Brownstein, J.S., Nikfarjam, A., Scotch, M., Smith, K.L., Gonzalez, G.: Social media mining for public health monitoring and surveillance. Paper presented at the Biocomputing 2016: Proceedings of the Pacific Symposium (2016) Pershad, Y., Hangge, P.T., Albadawi, H., Oklu, R.: Social medicine: Twitter in healthcare. J. Clin. Med. 7(6), 121 (2018) Rhodes, S.D., McCoy, T.P., Tanner, A.E., Stowers, J., Bachmann, L.H., Nguyen, A.L., Ross, M.W.: Using social media to increase HIV testing among gay and bisexual men, other men who have sex with men, and transgender persons: outcomes from a randomized community trial. Clin. Infect. Dis. 62(11), 1450–1453 (2016) Sarker, A., O’Connor, K., Ginn, R., Scotch, M., Smith, K., Malone, D., Gonzalez, G.: Social media mining for toxicovigilance: automatic monitoring of prescription medication abuse from Twitter. Drug Saf. 39(3), 231–240 (2016) Tursunbayeva, A., Franco, M., Pagliari, C.: Use of social media for e-Government in the public health sector: a systematic review of published studies. Gov. Inf. Q. 34(2), 270–282 (2017) Vannucci, A., Simpson, E.G., Gagnon, S., Ohannessian, C.M.: Social media use and risky behaviors in adolescents: a meta-analysis. J. Adolesc. 79, 258–274 (2020) Vaterlaus, J.M., Patten, E.V., Roche, C., Young, J.A.: # Gettinghealthy: the perceived influence of social media on young adult health behaviors. Comput. Hum. Behav. 45, 151–157 (2015). https://doi.org/10.1016/j.chb.2014.12.013 Ventola, C.L.: Social media and health care professionals: benefits, risks, and best practices. Pharm. Ther. 39(7), 491 (2014) Weiss, D., Rydland, H.T., Øversveen, E., Jensen, M.R., Solhaug, S., Krokstad, S.: Innovative technologies and social inequalities in health: a scoping review of the literature. PLoS ONE 13(4), e0195447 (2018) Wyche, S., Baumer, E.P.: Imagined Facebook: an exploratory study of non-users’ perceptions of social media in rural Zambia. New Media Soc. 19(7), 1092–1108 (2017) Xu, S., Markson, C., Costello, K.L., Xing, C.Y., Demissie, K., Llanos, A.A.: Leveraging social media to promote public health knowledge: example of cancer awareness via Twitter. JMIR Public Health Surveill. 2(1), e17 (2016) Zhang, J., Brackbill, D., Yang, S., Centola, D.: Efficacy and causal mechanism of an online social media intervention to increase physical activity: results of a randomized controlled trial. Prev. Med. Rep. 2, 651–657 (2015)

Potential Benefits of Social Media to Healthcare

241

Zhang, X., Wen, D., Liang, J., Lei, J.: How the public uses social media wechat to obtain health information in China: a survey study. BMC Med. Inform. Decis. Mak. 17(2), 71–79 (2017) Zhou, L., Zhang, D., Yang, C.C., Wang, Y.: Harnessing social media for health information management. Electron. Commer. Res. Appl. 27, 139–151 (2018)

Exploring the Influence of Human-Centered Design on User Experience in Health Informatics Sector: A Systematic Review Lina Fatini Azmi1(B) and Norasnita Ahmad2 1 UTM Research Computing, Department of Deputy Vice-Chancellor (Research and

Innovation), Universiti Teknologi Malaysia, 81310 Skudai, Johor, Malaysia [email protected] 2 Azman Hashim International Business School, Universiti Teknologi Malaysia, 81310 Skudai, Johor, Malaysia [email protected]

Abstract. Integrating human-centered design (HCD) approach in healthcare informatics solution are changing the landscape of the e-services and e-satisfaction among the users. Major evolution of informatics system in healthcare organization helps to revolve the role of design by changing it into key element that centralize on user’s capability on improving their e-service. Ample studies on implementing element of humanities into user experience-based design are now being adapted in order to enhance satisfaction and utmost benefits to users. This paper is built on a systematic literature review of academic papers that seeks to explore the influence of human-centered design approach towards user experience in health informatics sector. The total number of selected literatures using PRISMA process for this study is n = 64. The obtained results of this study highlighted the relation between human-centered design approach and user experience. This study also illustrates the process of human-centered design flow adapted from selected studies focus on healthcare sector in a unique approach to developing user-friendly informatics system to bridge the user experience gap. Keywords: Systematic review · Human-centered design · Health informatics system · Health informatics sector

1 Introduction Lately, user experience is playing a major role in usage of health informatics system that create a leaning trend that incline toward optimization of user experience through humancentered design. Human-centered design play a central role by helping system developer to develop a human-centric information system based on listening to and understanding user experiences, needs, and expectations [1]. Human-centered design is an innovative approach that was first used by corporations to create new products and services but in the same time it is intensely preoccupy the healthcare sector [2]. Human-centered design begins on how individuals understand their needs and designing from their viewpoint © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 242–251, 2021. https://doi.org/10.1007/978-3-030-70713-2_24

Exploring the Influence of Human-Centered Design on User Experience

243

[3]. Unfortunately, user satisfaction is difficult to measure. Therefore, user experience is crucial in designing human-centric informatics which can be illustrate on vast volume of exploratory research that being conducted to gather evidence [4] and various guideline have been developed to meet varying levels of user satisfaction. Moreover, study by [5] proved that the concept of user experience is a slightly unattended area in healthcare technology sector. Currently, studies on human-centered design in healthcare sector are widely discussed. Based on literature review of previous works, human-centered design approach usually starts with a brainstorming among team members on how to design an informatics system that works closer to human being. Then, prototypes are developed as a tool to accumulate user feedback. Hence, this review aims to explore the influence of system design towards feedback of user experience on human-centric health informatics system. This review also seeks to prove that system design contributes high influence to gain positive experience.

2 Previous Works Several preceding works that revolve around healthcare industry have been reviewed. Review by [6] also use PRISMA model to locate, select and include their selected studies. The review was conducted to assess the adequacy HCI and user-centered design in the development of e-mental health interventions. Next, study by [7] systematically reviewed using the Cochrane Handbook and PRISMA methods and guidelines. This goal of this study was to establish a measure of the user-centeredness of development processes and to define optimal practices. Another study by [8] also use PRISMA process for cross study analysis and synthesis. This study comprehensively details the problems with wearable Ventricular Assist Devices (VAD) systems and recommends a way to close the gap through human-centered design.

3 Method This study aims to systematically review the latest qualitative literature of user experience on human-centered design within health informatics system, analyze current evidence, and identify further research issues. The protocol for this systematic review is based on Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [9] which is a protocol that was developed and reviewed by the co-author until consensus was reached about the research questions and methods. According to [10], PRISMA statement was use to guarantee high-quality reporting of systematic reviews or metaanalyses published in studies in the healthcare field. In this study, user experience was defined as a user’s feedback about the implementation of human-centered design on the health informatics system. Human-centered design was defined as an innovation to humanize the design of informatics system to become more useful for people. Finally, the health informatics system was defined as e-services platform provided by healthcare providers. Protocol for this study is arrange as follows: i) research question developed ii) outlined search strategy iii) described inclusion and exclusion criteria iv) developed quality assessment and data extraction v) and, data analysis.

244

L. F. Azmi and N. Ahmad

3.1 Research Questions The review was guided by two specific questions: 1. What is the relation between human-centered design and user experience? 2. Does human-centered design approach influence user experience in health informatics system, and if so, how? 3.2 Search Strategy Five databases were searched using a broad search strategy; ScienceDirect, Scopus, Web of Science, SpringerLink, and Emerald Insight. The search terms were composed of variations of the following key terms: 1) user experience or satisfaction or perspective 2) user feedback 3) human-centered design 4) health informatics sector 5) health informatics system or platform. A brief grey literature search in Google Scholar was also conducted to expand the searching of relevant articles. As a result, Mendeley Desktop was used to store the findings of each database and to prevent duplicate studies and store all the citations. 3.3 Inclusion and Exclusion Criteria The flow of selected studies was conducted as in (see Fig. 1.). Full text screening of all articles was assessed using predefined inclusion criteria. The following studies were eligible for inclusion; (1) studies in health informatics sector area, (2) studies with full text articles in English, (3) studies assessing user experience within the health informatics system, (4) studies involving human-centered design of health informatics system, (5) studies published between 2015–2021, (6) studies published in journals and conference proceeding only. Studies were excluded if: (1) studies in other sector, (2) studies with no full text available and written in other language, (3) studies did not mention about user experience, (4) studies did not implement human-centered design on health informatics system, (5) studies published did not within determined period. 3.4 Quality Assessment and Data Extraction Selected studies were assessed for quality to ensure that the final list of studies answered the research questions before data extraction and analysis step. The following data was extracted from each selected study; study details (i.e. authors, title, publication year, source, methodology, issues/topics, theories, result/findings, future work). The data were tracked, gathered and recorded using Mendeley Desktop and Microsoft Excel. 3.5 Data Analysis The abstract sections of each study were read line-by-line to identify their aims, methodology and findings. In the beginning of identification phase, the database searching process combined with backward and forward searching method, 495 selected studies were

Exploring the Influence of Human-Centered Design on User Experience

245

Identification

subjected to a screening phase to remove irrelevant articles using inclusion and exclusion criteria. Following the screening phase, 105 studies were retained for full reading and 42 were removed throughout the eligibility phase after the assessment for quality stage conducted. As for the result in included phase, only 64 studies were selected in this review study as the flow of data analysis shown in Fig. 1.

Records identified through database searching (n=511)

Additional records identified through other sources (n=22)

Records after duplicates removed (n=495)

Screening

Records excluded (n=390)

Eligibility

Full-text articles assessed for eligibility (n=105)

Included

Records screened (n=105)

Studies included in SLR (n=64)

Full-text articles excluded, with reasons (n=41)

Fig. 1. PRISMA flow chart

4 Result 4.1 Year Published Over 83% of the papers (n = 53) that reported on user experience using human-centered design in health informatics sector were published in the latest of our chosen time period

246

L. F. Azmi and N. Ahmad

(2018 = 6; 2019 = 24; 2020 = 23). The remaining papers (n = 11) were published in the earlier years of chosen time period (2015 = 4; 2016 = 2; 2017 = 5). Therefore, it clearly shows that studies on this topic are widely discussed in latest previous year. 4.2 Methodology Only 10% of the selected studies used mixed mode method while none of them use quantitative method in the area of study. Most of the selected studies in this domain area used qualitative method in various approaches includes case study, focus group, interview, and observation. This is strongly shows that qualitative method is the best method to be used in human-centered design studies. The reason of these because humancentered design is an approach to develop informatics system, therefore, interview is the main technique in data collection for this area of study. 4.3 Publication Sources Almost all of the selected studies (between 2015–2020) were published in journals with majority distribution of 73%. Then, it followed by conference papers with 23%, while research studies only 4%. 4.4 Context of Study The context of the reported studies were: general health issues (n = 25), people with mental illness (n = 6), chronic health condition (n = 4), older adults (n = 3), rehabilitation treatment (n = 3), children (n = 3), hypospadias surgery (n = 2), HIV (n = 2), people with cancer (n = 2), people with asthma (n = 2), woman (n = 2), glucose patient (n = 1), people undergoes dialysis treatment (n = 1), psychotherapies (n = 1), pelvic exam (n = 1), spine surgery (n = 1), people with dementia (n = 1), people with hearing impaired (n = 1), people with arthritis (n = 1), radiology (n = 1), pulmonary embolism (n = 1). 4.5 Discussion of SLR RQ1: What is the relation between human-centered design (HCD) and user experience (UX)? In order to prove the influence of human-centered design towards user experience, evidence of their relation is needed. There are several evidences of relation between human-centered design and user experience reported by various studies in health informatics sector. Table 1 shows some evidences of relation between HCD and UX highlighted in the selected studies. Based on the Table 1, it proved that HCD have its own influences towards positive user experience and HCD can helps in improving healthcare system. RQ2: Does human-centered design approach influence user experience in health informatics system, and if so, how? Yes. Various studies using human-centered design approach as the methodology in their studies helps to improve their services and user experience instantly. Based on

Exploring the Influence of Human-Centered Design on User Experience

247

Table 1. Relation between HCD and UX. Author

Relation

[11]

“Stakeholders in healthcare system started realizing the importance of user experience …. required involving human-centered design and engineering to structured healthcare design paradigm”

[12]

“Applies HCD thinking to solve critical healthcare problems that is more focused on the patient and the patient experience”

[8]

“The user experience resulting from the design of the wearable system … by positioning human-centered design opportunities at the intersection of human factors and user experience”

[13]

“Human-centered design research methodologies put user experience and needs at the forefront of the development process … result in the next generation of patient-friendly healthcare”

[14]

“HCD make systems usable and useful by focusing more on the design analysis to improve user experience”

[15]

“The team strived to design usable interface using HCD approach to support learnability …, while providing a pleasant user experience …”

[16]

“Suggesting the need for new design approaches for population of low user experience smoking cessation app”

[17]

“.. human-centered design approach combines with design thinking methodology focuses on the end-user experience to generate innovation …”

[18]

“Incorporating HCD methodology into this problem allows for the consideration of user experiences when making design decisions”

[19, 20]

“HCD able to addresses the whole user experience, including the context in which the user finds his/herself”

[21]

“…. requires thoughtful design of user-friendly interfaces that consider user experience and present data in personalized ways …”

[22]

“Design processes such as HCD, ….. can be crucial in ensuring that the product meets the needs …, in terms of safety and user experience.”

[7]

“HCD is a highly iterative method for optimizing the user experience and the effectiveness of the system, service or product”

[6]

“… understanding HCI and HCD is an important factor in developing successful computer user experience”

[23]

“… apps/wearable … should be designed to leverage and further improve the user experience …”

[24]

“.. to improve designer’s abilities and tools to search for user experience, active participation of users to design process and sharing their experiences with designers … is critical to develop more inclusive environments..”

[25]

“… an appropriate selection of game-design principles, … may improve the usability and user experience of a system” (continued)

248

L. F. Azmi and N. Ahmad Table 1. (continued)

Author

Relation

[26]

“In contrast, HCD focuses primarily on individual user experiences.”

[1]

“… using HCD to optimize user experience at a tertiary academic medical center.”

[27]

“The human-centered design thinking methodology … identifying and defining the problem… a deep understanding of user experience”

[28]

“A review of the design … reducing the burden of cognitive strain experience”

selected studies, many of them have been explained the way on how they implemented human-centered design approach in their development phase of health informatics system. To conclude the human-centered design method used by selected studies, see Fig. 2 shows the flow of the process.

Positive User Experience

Inspiration • • • •

Complaining Sharing Information Listening & Feedback Consulting & Advising

Human Centered Design

Delivery

Ideation • • •

Identify Needs Research & Analysis Design Concept

Implementation • • •

Develop Prototype Pilot Testing User Experience Evaluation

Fig. 2. Flow of process human-centered design (self-adaption)

Reported on various studies, human-centered design (HCD) consists of 3 main phases; inspiration, ideation, and implementation. At the end of these phases, the system will be delivered. First phase is inspiration, which is the beginning of ideas by listening to complains, sharing information by user, listening user’s feedback and consultation from the expert in order to improve their products/services. Next, ideation phase that converge all the ideas that helps to cater the needs of users which lead to research and analysis stage of the results and reshape it into vital piece on designing the concept of

Exploring the Influence of Human-Centered Design on User Experience

249

HCD. Third phase is implementation, which is to implement the ideas by developing prototype, pilot testing, and user experience evaluation. These 3 phases are iterative until the final evaluations by the users are reaching their goals. Lastly, they can deliver the product to the user and gain positive user experience.

5 Discussion Result shows that implementing human-centered design principle in development process of informatics system phase able to expand the services productivity. All the selected studies focus on the same aim which is developing products based on human-centered design as their methodology in order to improve the user experience and satisfaction. The upshot of evolution in technologies nowadays forcing many sectors to change their nature of working from paper-based to system-based. Thus, designing a system is challenging as they need to consider the needs, expectation and experience of the users. Moreover, this domain area of studies still debatable actively on previous research and studies. Therefore, it is proven that further study about user experience on human-centered design implementation to the technologies is paramount to improve the e-services in healthcare sector. In addition, this study helps in proving the influences of human centered design (HCD) towards positive user experience (UX).

6 Conclusion This systematic literature review is conducted for studies published in health informatics sector between 2015 until 2020. Even though this domain area has varieties of discussion on previous year studies but through these days, only few studies particularly focusing on applying human-centered design approach on health informatics sector. Mostly, studies focused on improving their e-services in health sector using human-centered design approach. Out of all publication within the selected time period, only 64 studies met the inclusion criteria for the review. This study also points out relation between humancentered design approach, user experience and also the process of influence of humancentered design implementation to the informatics system.

References 1. Vagal, A., Wahab, S., et al.: Optimizing patient experience using human-centered design. J. Am. Coll. Radiol. 17(5), 668–672 (2020). https://doi.org/10.1016/j.jacr.2019.11.020 2. Kim, S.H., Myers, C.G., Allen, L.: Health care providers can use design thinking to improve patient experiences. Harv. Bus. Rev. 95, 222–229 (2017) 3. Searl, M.M., Borgi, L., Chemali, Z.: It is time to talk about people: a human-centered healthcare system. Health Res. Policy Syst. 8, 35 (2010) 4. Matheson, G.O., et al.: Leveraging human-centered design in chronic disease prevention. Am. J. Prev. Med. 48(4), 472–479 (2015). https://doi.org/10.1016/j.amepre.2014.10.014 5. Trauzettel, F., Minge, M.: Usability in the lifecycle of medical software development. Curr. Dir. Biomed. Eng. 2(1), 583–586 (2016). https://doi.org/10.1515/cdbme-2016-0129

250

L. F. Azmi and N. Ahmad

6. Søgaard Neilsen, A., Wilson, R.L.: Combining e-mental health intervention development with human computer interaction (HCI) design to enhance technology-facilitated recovery for people with depression and/or anxiety conditions: an integrative literature review. Int. J. Ment. Health Nurs. 28(1), 22–39 (2019). https://doi.org/10.1111/inm.12527 7. Witteman, H.O., et al.: User-centered design and the development of patient decision aids: protocol for a systematic review. Syst. Rev. 4(1) (2015). https://doi.org/10.1186/2046-40534-11 8. Dunn, J.L., et al.: Human factors and user experience issues with ventricular assist device wearable components: a systematic review. Ann. Biomed. Eng. 2431–2488 (2019). https:// doi.org/10.1007/s10439-019-02303-3 9. Shamseer, L., et al.: Preferred reporting items for systematic review and meta-analysis protocols (PRISMA-P) 2015: elaboration and explanation. BMJ (Online) g7647 (2015). https:// doi.org/10.1136/bmj.g7647 10. Panic, N., et al.: Evaluation of the endorsement of the preferred reporting items for systematic reviews and meta-analysis (PRISMA) statement on the quality of published systematic review and meta-analyses. PLoS ONE (2013). https://doi.org/10.1371/journal.pone.0083138 11. Dey, N., Rautray, P., Soni, M.: Patient-centered design in a connected healthcare world: a case study. Smart Innovation, Systems and Technologies, pp. 967–976. Springer Science and Business Media Deutschland GmbH (2019). https://doi.org/10.1007/978-981-13-5974-3_83 12. Gomes, N., Patwardhan, V.: Applying human-centered design and human-machine integration techniques to solve key healthcare problems. Advances in Intelligent Systems and Computing, pp. 3–9. Springer (2019). https://doi.org/10.1007/978-3-030-02053-8_1 13. Park, T., et al.: Living profiles: an example of user-centered design in developing a teenoriented personal health record. Pers. Ubiquit. Comput. 19(1), 69–77 (2015). https://doi.org/ 10.1007/s00779-014-0812-1 14. Khakurel, J., et al.: Human-centered design components in spiral model to improve mobility of older adults, pp. 83–104. Springer, Cham (2019). https://doi.org/10.1007/978-3-319-934 91-4_5 15. Sebillo, M., et al.: Human-centered design of a personal medication assistant - putting polypharmacy management into patient’s hand! Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), pp. 685–699. Springer (2017). https://doi.org/10.1007/978-3-319-57186-7_49 16. Vilardaga, R., et al.: User-centered design of learn to quit, a smoking cessation smartphone app for people with serious mental illness. J. Med. Internet Res. 20(1) (2018). https://doi.org/ 10.2196/games.8881 17. Carter, J., Bababekov, Y.J., Majmudar, M.D.: Training for our digital future: a human-centered design approach to graduate medical education for aspiring clinician-innovators. npj Digital Med. 1(1), 26 (2018). https://doi.org/10.1038/s41746-018-0034-4 18. Taylor, G.A., McDonagh, D., Hansen, M.J.: Improving the pelvic exam experience: a humancentered design study. Des. J. 20(sup1), S2348–S2362 (2017). https://doi.org/10.1080/146 06925.2017.1352750 19. Holeman, I., Kane, D.: Human-centered design for global health equity. Inf. Technol. Dev. 26(3), 477–505 (2020). https://doi.org/10.1080/02681102.2019.1667289 20. Persson, J.: A review of the design and development processes of simulation for training in healthcare – a technology-centered versus a human-centered perspective. Appl. Ergon. 314–326 (2017). https://doi.org/10.1016/j.apergo.2016.07.007 21. Hartzler, A.L., et al.: Integrating patient-reported outcomes into spine surgical care through visual dashboards: lessons learned from human-centered design. eGEMs (Generating Evid. Methods Improve Patient Outcomes) 3(2), 2 (2015). https://doi.org/10.13063/2327-9214. 1133

Exploring the Influence of Human-Centered Design on User Experience

251

22. Harte, R., et al.: Human-centered design study: enhancing the usability of a mobile phone app in an integrated falls risk detection system for use by older adult users. JMIR mHealth uHealth 5(5), e71 (2017). https://doi.org/10.2196/mhealth.7046 23. Wulfovich, S., et al.: “I must try harder”: design implications for mobile apps and wearables contributing to self-efficacy of patients with chronic conditions. Front. Psychol. 10 (2019). https://doi.org/10.3389/fpsyg.2019.02388 24. Özten Anay, M.: Design thinking to familiarize hearing-impaired architectural drafting students with human-centered design concept. ICONARP Int. J. Archit. Plan. 8(1), 62–87 (2020). https://doi.org/10.15320/iconarp.2020.105 25. Schulz, R., Martinez, S., Hara, T.: Towards a game-design framework for evidence-based clinical procedure libraries. In: 2019 IEEE 7th International Conference on Serious Games and Applications for Health, SeGAH 2019. Institute of Electrical and Electronics Engineers Inc. (2019). https://doi.org/10.1109/SeGAH.2019.8882474 26. Chen, E., et al.: Enhancing community-based participatory research through human-centered design strategies. Health Promot. Pract. 21(1), 37–48 (2020). https://doi.org/10.1177/152483 9919850557 27. Vagal, A., Wahab, S.A., et al.: Human-centered design thinking in radiology. J. Am. Coll. Radiol. 17(5), 662–667 (2020). https://doi.org/10.1016/j.jacr.2019.11.019 28. Faiola, A., Srinivas, P., Duke, J.: Supporting clinical cognition: a human-centered approach to a novel ICU information visualization dashboard. In: AMIA ... Annual Symposium Proceedings, AMIA Symposium, pp. 560–569 (2015). https://pubmed.ncbi.nlm.nih.gov/26958190/. Accessed 9 Sep 2020

An Emotional-Persuasive Habit-Change Support Mobile Application for Heart Disease Patients (BeHabit) Bhavani Devi Ravichandran(B) and Pantea Keikhosrokiani School of Computer Sciences, Universiti Sains Malaysia, 11800 Minden, Penang, Malaysia [email protected], [email protected]

Abstract. Heart disease is stated as the world’s biggest killers. The risk factors of this deadly disease are due to some bad habits such as being overweight, bad eating diet, smoking, assumption of alcohol, etc. Nevertheless, patients can live a healthy lifestyle if they have the proper guidance of persuasive-emotional featured technologies. In line with this, this study focuses on developing an emotionalpersuasive habit-change support mobile application called BeHabit to improve heart disease patients’ lifestyles. Persuasive-emotional features are two different features that are integrated with BeHabit to distinguish this application from the existing ones. The proposed system is designed, implemented, tested, and evaluated by 10 users. In conclusion, the users are satisfied to used BeHabit to change their bad habits. Emotional and persuasive features which are integrated into BeHabit are the key to help patients to change their bad habits. BeHabit and the integrated feature can be used as a guideline for healthcare developers and providers for the improvement of mHealth services. Keywords: Heart disease · mHealth · Habit-change · Persuasive · Emotional features · Mood · Medical information system

1 Introduction According to the press release statistics on causes of death, Malaysia [1], Malaysians affected by coronary illness at a very young age of 41 as compared to other nations whereas in Thailand it is at the age of 65, in China 63, in western countries 66 and Canada at the age of 68. Therefore, more attention is required for heart disease patients in Malaysia. For instance, Mobile health technologies can impact the health of a chronic disease patient [2]. Having bad habits of diet, sleep, smoking, exercise, etc. might be the main cause of many diseases such as heart disease. Many existing mobile applications provide a platform for heart disease patients mostly to record patient’s activities; however, there is no current mobile application that can influence to change heart disease patients’ bad habits by adding persuasive and emotional features. The lifestyle of heart disease patients is very different from a normal individual. A heart disease patient cannot perform vigorous physical activities, nor they should not have their heart rate beat faster than a certain rate. Therefore, this study aims to develop a mobile application called BeHabit © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 252–262, 2021. https://doi.org/10.1007/978-3-030-70713-2_25

An Emotional-Persuasive Habit-Change Support Mobile Application

253

to assist heart disease patients by implementing persuasive and emotional features as a motive to form a new healthy lifestyle. BeHabit guides patients to follow daily healthy routines as well it keeps track of the patient’s moods and symptoms after each activity. It is programmed to suggest carrying out light physical activities as it is suggested by medical expertise [3]. BeHabit also provides a summary of the day to motivate the patients emotionally so that they will be enthusiastic to continue to live a healthy lifestyle. Besides, this application also helps the doctor to monitor their patients in real-time as well as provide prescriptions to their patients. This paper firstly introduced some existing mobile health (mHealth) applications followed by the proposed emotional-persuasive habit-change support system with a mobile application which is called BeHabit. BeHabit is developed in June 2020 to assist heart disease patients to change their bad habits of exercise and mood to reduce further development of heart disease. System design, methodology, implementation, test, and user acceptance evaluation are summarized in this paper. Finally, concluding remarks and future works are added as well. 1.1 Background The use of mHealth in the medical world has been a massive game-changer to the healthcare industry as patients have access to the latest and best reliable medical resources, treatments, good communication with doctors, and many more. Emerging mobile technology provides a platform for patients to be more aware of self-care, reduce hospitalization and mortality rates by 21% and 20% respectively [4]. The rate of heart disease is relentlessly expanding as well as becoming one of the biggest killers. Nonetheless, there are many ways to prevent heart disease by following good habits such as exercising 30 min a day on most days of the week, eating a healthy diet, maintaining an ideal weight, reduce stress, and many countless efforts that can be taken. Persuasive-emotional features are two different technologies with each being researched individually. Persuasive technology is the study of computers known as Captology. It is the study of interactive technology that helps to change the user’s habit. A study by [5], has presented five perspectives on computers and persuasion where it is the primary research to emphasize further to understand more on persuasive computing. Nevertheless, emotional technology is the new platform for further improvement in artificial intelligence area where it measures biometric information to define emotion as a computation for different computer applications. Recently, there has an increase in demand for its application to various fields. In this scenario, emotionally featured technology is key to help patients to guide their emotions in a way so that they could change their bad habits. The health habits of heart disease patients play a major role in mortality. Patients are advised to exercise regularly. According to medical experts, when it comes to exercising, it is different for heart disease patients as they should only maintain light exercise [3]. Hence, heart disease patients should be guided differently based on their lifestyle. Instant Heart Rate mobile application is designed to measure pulse accurately and heartbeat zone with heart rate and health monitor after sleeping or during workouts & training. Instant Heart Rate doesn’t require heart rate straps. It monitors blood circulation with accurate heart health monitors (like ECG or EKG). Functions similarly to pulse

254

B. D. Ravichandran and P. Keikhosrokiani

oximeters, detecting the change in your finger to provide accurate heartbeat measurements. It can measure the instant heart rate in less than 10 s. Most of the functions are in-app purchase and has limited functionalities [6, 7]. Cardiio helps to measure the user’s pulse using a phone camera. This application will help users gain insights on how heart rate relates to fitness and endurance. It will also improve the user’s fitness by building high-intensity circuit training exercises that will take around 7 min to complete. It keeps track of personal dashboards with history for daily, weekly, and monthly. This application is only available in iOS and it has in-app purchases where most functionalities are limited to normal users [8]. iCardio application helps users to keep track of runs, rides, and many activities related to cardio at the gym, daily step count, and activity all in one application. Users can add heart rate for more accuracy to count calories. The main motivation for this application is for users to lose weight. This application can also track workouts indoors and outdoors [9, 10].

2 Proposed Solution BeHabit is a specialized solution for heart disease patients to improve their lifestyle for being healthy. This system connects to Samsung Health with Samsung Smartwatch which retrieves data from Samsung Health Cloud. The smartwatch can provide more accurate data of the user such as heart rate, calories, distance traveled, and step counts. Samsung Health platform is a useful tool as it provides a centralized databased for developers to work on their projects. All information is retrievable from one Samsung account. BeHabit proposes a solution for heart diseases patient as this system analyses user heart rate in real-time and able to identify any abnormal changes. After the detection, the system automatically sends the user’s abnormal maximum heart rate to the doctor as well as alerts the user in the application. The user can also communicate with the doctor by sending messages and retrieve prescriptions directly from their doctor. Moreover, this system implements persuasive and emotional features to persuade patients for changing their bad habits. As for implementing persuasive features, the system applied a point collecting system known as BeHabit point. The BeHabit point is a technique to measure a user’s activeness which is presented in Table 1. This feature will motivate users persuasively to improve their health by either increasing step count or carrying more activity. The system also sends user persuasive messages to complete their targeted achievements. Another persuasive feature in BeHabit is that the system will let the user choose an activity from some listed choices and the targeted minutes to complete is at most 30 min for each. At the end of the activity, the system will prompt praise to the user. More persuasive features are defined by [11]. BeHabit uses dialogue support, credibility support, and primary task to support the persuasion context. In short, an application should be appealing, pleasurable, memorable, and effective. Nevertheless, with the same ideology, BeHabit has implemented similar methods to emotionally affect the user. Moreover, the system keeps track of the user’s mood before and after an activity. The system also keeps user symptoms if any after an activity. According to [12], mobile apps are developed for mood tracking in which the application features can be mapped

An Emotional-Persuasive Habit-Change Support Mobile Application

255

Table 1. Habit-points calculation Habit-points 0 to 20 range

21 to 40

41 to 60

61 to 80

81 to 100

Remarks

Very bad

Bad

Average

Good

Excellent

Types of activity achieved

– 20% of the targeted step count – Preference of user mood

– 40% of the targeted step count – Preference of user mood

– 60% of the targeted step count – Preference of user mood

– 80% of the targeted step count – Preference of user mood

– 100% or more of the targeted step count – Preference of user mood

into stages of mood tracking. Table 2 shows the stages to be implemented in the BeHabit system. For example, when the system prompts user mood tracking form, it is a stage of preparation where it provides fundamental information on how to conduct mood tracking. Next, the system will show a range of emoticons, pictures, and texts to define how the user feels. This is a collection stage. Table 2. Stages of mood tracking

The system will also notify users emotionally using motivational messages to improve user’s emotional mentality. This feature is also a collection stage. Moreover, the system will display a summary of user activity. The main purpose of this feature is to emotionally motivate the user to keep up the work or to remind them to work out more which according to Table 2 which is at the reflection stage. Finally, the system will export user data to the doctor to be referred and receive a prescription to improve the user’s health. 2.1 System Design BeHabit system architecture design is illustrated in Fig. 1. End users will be using android mobile devices and connect to the internet via an access point. The mobile applications have access to the online database which is in this case Firebase and Samsung Health cloud through the Internet connection. However, the user’s health data retrieved from Samsung Health are not stored in any database to ensure the user’s confidentiality and data access control.

256

B. D. Ravichandran and P. Keikhosrokiani

Fig. 1. System architecture diagram

Application architecture design is divided into 3 layers, namely (1) the view layer, (2) domain layer, and (3) the data layer. Each of the layers places significant roles in the overall architecture of the project. The view layer is responsible for the interaction between the users and the system. The mobile application needs to be installed in a smartphone that is equipped with a stable cellular network with a GPS sensor. The application will be connected to the Internet via the gateway. The domain layer is responsible for processing the data obtained from the view layer and pass it to API for data operation. The API implemented here is Google Firebase services, Samsung Health services, and YouTube API. Finally, the data layer is responsible for storing and managing the data to be used in the application. The main database used in this system is Firebase Real-Time database which is a flexible scalable new SQL cloud database to store and think data for the client. The firebase database functions as online storage for the application to retrieve the basic information on the user. The firebase cloud messaging (FCM) Service is also used to send push notification. The component in every layer needs to work together seemingly to ensure the application can perform at a desirable level. Nonetheless, Samsung Health Data Store and Tracker Service have been responsible to provide a platform for sharing health data from users to Android phones. The health data and services are retrieved in real-time to the BeHabit system. Health data sharing is shared with the user’s knowledge.

3 System Development Methodology The development methodology for this application is the Software Development Life Cycle (SDLC) methodology designed by [13–17]. SDLC is a framework that characterizes the various advances or procedures. The various steps involved in SDLC are modeling, assessment, design, and prototype as shown in Fig. 2. The SDLC can be applied

An Emotional-Persuasive Habit-Change Support Mobile Application

257

Fig. 2. SDLC methodology of the project [13]

to both hardware and software which will deliver high-quality products or services. This will guarantee the smooth running of the organizations. 3.1 System Implementation System Requirement There are some functional and non-functional requirements for developing BeHabit. Based on the functional requirements, the system should be able to retrieve real-time Samsung Health data and display them in an understandable user interface. It must provide an option for the user to view a history of health data. Furthermore, the user should be able to receive push notification which contains a daily quote to encourage the user. The user can send messages to and receive from the doctor in real-time. The system provides a platform for users to track and view their mood. The system will check the user’s heart rate if the user has the device to check it. The system displays helpful and encouraging tips to motivate users as a persuasive feature. It should recommend the user the types of activity to carry out based on heart rate data User is also able to share their achievement of the week as well as month. Finally, it must notify the doctor and user in case of any abnormalities in the user’s heart rate in real-time. As for non-functional requirements, the mobile application is designed with a minimum SDK version API 24: Android 7.0 Nougat which is 100% compatible with all portable devices. Every user is required to sign into the system to access the functionalities of the mobile application in the line of protecting the personal data. The user’s health data is retrieved directly from the Samsung Health API and is not stored anywhere in the device to ensure the user’s confidentiality. The mobile application should be designed with a consistent graphical user interface that is user friendly. The system always performs at an optimal level unless there is no Internet connection or have a failure in the online database. This requirement also always ensures system availability.

258

B. D. Ravichandran and P. Keikhosrokiani

Algorithms, Pseudocodes, APIs In this section, only the operations that have more sophisticated procedures with the use of a specific library or APIs are discussed. The straightforward operation such as user sign in, display encouraging tips, manage user profile, and more are considered as self-explanatory, thus it will not discuss in this section. Recommend Activity In recommendation activity which is a implementation of persuasive feature, the application retrieves binning heart rate data from Samsung Health data of the user. BeHabit has a built-in method to calculate the user’s estimated heart rate at a vigorous level and compare the user’s current heart rates with the value. The methods are called is onCalculateEstimatedHR() and onCompareHR(). Both methods are called when user requests for activity recommendation. The calculation to identify the user’s estimated heart rate value at vigorous value logic starts with converting the binning heart rate value into arrays and obtaining the user’s age from the firebase database. The method compares the age and the arrays of heart rate with the estimated value. If the array of heart rate is less than the estimated value, then BeHabit will recommend regular activities to carry out. However, if there’s any heart rate value in the array that exceeds the estimated value, then BeHabit will immediately recommend calming activities. The user can carry out the activity when the tap on the activity icon and the application will link directly to the Samsung Health activity tracker. Alert User An alert user activity, the primary objective is to alert the user when the heart rate during exercise exceed the estimated heart rate value. BeHabit uses the following table as a guide for detecting the abnormal heart rate. Firstly, the user’s age is retrieved from the Firebase database. Secondly, BeHabit received binning heart rate data which is converted to arrays to check if there is any heart rate that exceeds the heart rate zone of vigorous-intensity and maximum. The method implemented here is called checkExerciseHeartRate() which returns either true or false. In a scenario where the method returns true, the notificationAlert() method will be triggered. This method will trigger sending out a notification to the user to inform the user immediately. Furthermore, the exceeded heart rate value will be pushed to the Firebase database. So, the user can view the heart rate values by data later in a list view. Send/Receive Message to/from Doctor BeHabit allows communication between user/patients and their respective doctors. The user can send a message to the doctor. The user has to select the Send message button which triggers the sendMessage() method. This method opens another layout to allow the user to write the message. When the user is done, he/she can tap on the send button. However, if the message is empty, the system will prompt the message when not the message will be pushed to the Firebase database. This message will later be received at the doctor’s site and be displayed in the application. The system will also display

An Emotional-Persuasive Habit-Change Support Mobile Application

259

messages received from the doctor site which are pushed in the Firebase database in the application. Retrieve Current and Historical Health Data BeHabit application is primarily dependent on Samsung Health SDK which provides Data and Services to Android API. The Samsung Health Data Store syncs data with the user’s Samsung Account. The health data is retrieved from the Samsung Health Server that implements Rest API. For first time users, BeHabit will request Rest API Oauth2 from the Samsung Health Server SDK. The users must approve the authentication to fully utilized BeHabit functionalities. Initially, health data service is required to be initialized and ensure health data store connection is connected. Next, it is vital to set the listener to retrieve the required health data to use. In this system, daily step count, heart rate, and exercise health data are retrieved. This activity implements a persuasive feature as it motivates the user to be more active. Receive Daily Quotes In this part, the system sends user inspirational daily quotes as persuasion features. Every day at 8 am, the user receives motivational quotes using Firebase Cloud Messaging. The system randomly selects a quote from a long list of quotes and triggers the quoteFCM() method to start. This method starts the service and sends out a daily quote to the user. Moreover, the user can view and share the quote in the application to other social media platforms. Mood and Symptoms Tracker The system tracks the user’s mood and symptoms daily as emotional features. User is required to select the date to enter entry for either mood or symptom. The system will trigger moodTrack() and symptomsTrack(). The user must select one mood or symptom and tap on the submit button. Next, the submitted information will be pushed to Firebase real-time database. The user can also view the submitted data in the application. Calculate BeHabit Points The main purpose of calculating BeHabit points is to identify the activeness and mental evaluation of the user. This point system is dependent on both the total steps taken by the user and the mood of the day. The system retrieves the health data and reads the user’s input for symptoms. The implementation of BeHabit points applies persuasive features. The formula for BeHabit as shown as below: BeHabit points = ((Total steps/target steps) ∗ 0.5) + ((Mood value for the day/Number of moods) ∗ 0.5)

(1)

4 Testing and Evaluation 4.1 Unit Testing Unit testing in defies the smallest units of codes for its functionality, the purpose is to identify that each unit of the software performs as designed. Unit testing is very

260

B. D. Ravichandran and P. Keikhosrokiani

important because it can ensure the system runs seamlessly who starred in the Android application which uses Android studio as the environment tools, unit testing is carried out on a method. The functionality of a method is tested one by each other to minimize the error when methods are integrating to become a subsystem. 4.2 Integration Testing Integration testing is a process of combining the units of code and carry out their testing process that will produce the result of combination functions correctly. The purpose of this level of testing is to expose faults in the interaction between integrated units. Integration testing provides a systematic technique for assembling a software system while conducting tests to uncover errors associated with interfacing. It can ensure the parameter, function, run-time Exceptions an incompatibility between the interaction of objects. In this application, the integration between the subsystems is important because we divided the work to different members. Integration testing needs to be carried out from time to time starting from the development of the project. It is advisable to carry out integration testing when every functionality of the subsystem is complete to ensure the efficiency for implementation of the application. 4.3 System Testing and User Acceptance Evaluation System testing is a level of software testing where a complete and integrated software is tested. We performed system testing at the end of each iteration to ensure system compliance with the specific requirement. If that any error, immediate action needs to be done to fine-tune the system. In this project, an example of system testing carried out was the application system to ensure it can retrieve real-time data about the user’s health data and process the data to calculate if the user is active or not. Generally, user acceptance testing (UAT) on the system is carried when the system is integrated and completed. It is carried up by randomly selecting 10 users. This testing

Fig. 3. User interfaces designs of BeHabit application

An Emotional-Persuasive Habit-Change Support Mobile Application

261

is carried out by first to introduce all the available features to the respondent to make sure they understand the usage of each function implemented. After that, each of them we’ve given about 20 min. Next, the respondents are asked to fill up a questionnaire that consists of 7 questions to find user satisfaction and acceptance of BeHabit. After gathering all the responses, the results of UAT are analyzed Lastly, the result of UAT will be taken into consideration in future enhancement or development of the project. Figure 3 shows the user interfaces of the application.

5 Conclusion and Future Work The system with the BeHabit application has been developed successfully by meting requirements to provide a platform for improving heart disease patients’ lifestyles. BeHabit has equipped with push notification that can alert the user when the heart rate increases as well as receive a prescription from the doctor. At the same time, the application benefits patients by providing emotional and persuasive feature implantation. The user can keep track of their daily mood from time to time and symptoms if there’s any. All this information will be sent to the doctor’s site to observe and provide prescriptions according to the individual user. The user can also send messages to the doctor if they wish to ask any questions or to update. The application provides a visual presentation of the user’s health data. Moreover, the system can provide recommended activities to users according to their health and more. BeHabit was tested and evaluated by 10 users who were satisfied to use the app. Emotional and persuasive features are very important for changing bad habits. In the future, communication with a doctor should be improved. The system can increase the usability to keep track of the user’s other unhealthy habits such as smoking, diet, etc. Furthermore, a food intake tracking feature can be added to track of user’s calorie intake. Finally, BeHabit point calculation can be improved. Acknowledgment. The authors are thankful to School of Computer Sciences, and Division of Research & Innovation, USM for providing financial support from Short Term Grant (304/PKOMP/6315435) granted to Dr Pantea Keikhosrokiani.

References 1. Department of Statistics Malaysia. Press Release Statistics on Causes of Death, Malaysia (2019). https://dosm.gov.my/v1/index.php?r=column/pdfPrev&id=RUxlSDNkcnRVazJnak NCNVN2VGgrdz09. Accessed 30 Nov 2019 2. Nilsen, W., et al.: Advancing the science of mHealth. J. Health Commun. 17(sup1), 5–10 (2012) 3. Fuezeki, E., Engeroff, T., Banzer, W.: Health benefits of light-intensity physical activity: a systematic review of accelerometer data of the National Health and Nutrition Examination Survey (NHANES). Sports Med. 47(9), 1769–1793 (2017) 4. Clark, R.A., et al.: Telemonitoring or structured telephone support programmes for patients with chronic heart failure: systematic review and meta-analysis. BMJ 334(7600), 942 (2007) 5. Fogg, B.J.: Persuasive computers: perspectives and research directions. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (1998)

262

B. D. Ravichandran and P. Keikhosrokiani

6. Instant Heart Rate: HR Monitor & Pulse Checker. Apps on Google Play, Google (2019). https://play.google.com/store/apps/details?id=si.modula.android.instantheartrate&hl=en. Accessed 28 Oct 2019 7. Azumio Inc, Instant Heart Rate: HR Monitor. App Store (2019). https://apps.apple.com/us/ app/instant-heart-rate-hr-monitor/id409625068. Accessed 28 Oct 2019 8. Cardiio, Inc, Cardiio: Heart Rate Monitor. App Store (2019). https://apps.apple.com/us/app/ cardiio-heart-rate-monitor/id542891434. Accessed 28 Oct 2019 9. iCardio Workout Tracker & Heart Rate Trainer. Apps on Google Play, Google (2019). https://play.google.com/store/apps/details?id=com.fitdigits.icardio.app&hl=en. Accessed 28 Oct 2019 10. Fitdigits Inc, iCardio Workout Tracker. App Store (2019). https://apps.apple.com/us/app/ica rdio-workout-tracker/id314841648. Accessed 28 Oct 2019 11. Lehto, T., Oinas-Kukkonen, H.: Persuasive features in web-based alcohol and smoking interventions: a systematic review of the literature. J. Med. Internet Res. 13(3), e46 (2011) 12. Caldeira, C., et al.: Mobile apps for mood tracking: an analysis of features and user reviews. In: AMIA Annual Symposium Proceedings. American Medical Informatics Association (2017) 13. Keikhosrokiani, P.: Perspectives in the Development of Mobile Medical Information Systems: Life Cycle, Management, Methodological Approach, and Application. Academic Press, Cambridge (2019) 14. Keikhosrokiani, P.: Chapter 6 - Emotional-persuasive and habit-change assessment of mobile medical information Systems (mMIS). In: Keikhosrokiani, P. (ed.) Perspectives in the Development of Mobile Medical Information Systems, pp. 101–109. Academic Press (2020) 15. Keikhosrokiani, P., et al.: User behavioral intention toward using mobile healthcare system. In Consumer-Driven Technologies in Healthcare: Breakthroughs in Research and Practice, pp. 429–444. IGI Global (2019) 16. Keikhosrokiani, P.: Chapter 4 - Behavioral intention to use of mobile medical information system (mMIS). In: Keikhosrokiani, P. (ed.) Perspectives in the Development of Mobile Medical Information Systems, pp. 57–73. Academic Press (2020) 17. Keikhosrokiani, P., Mustaffa, N., Zakaria, N.: Success factors in developing iHeart as a patientcentric healthcare system: a multi-group analysis. Telematics Inform. 35(4), 753–775 (2018) 18. Keikhosrokiani, P., et al.: Assessment of a medical information system: the mediating role of use and user satisfaction on the success of human interaction with the mobile healthcare system (iHeart). Cogn. Technol. Work 22(2), 281–305 (2020)

A Systematic Review of the Integration of Motivational and Behavioural Theories in Game-Based Health Interventions Abdulsalam S. Mustafa(B) , Nor’ashikin Ali, and Jaspaljeet Singh Dhillon Universiti Tenaga Nasional, Selangor, Malaysia {nora.ali08,jaspaljeet}@uniten.edu.my

Abstract. M-Health interventions designed for healthcare can potentially increase participation and behaviour outcomes. However, interventions need to incorporate a theoretical perspective of behavioural change to enhance their perceived efficacy. Although behavioural outcome theories have gained interest in the health and fitness literature, the implementation of theoretical integration remains largely under-studied. Therefore, we reviewed the efficacy of behavioural gamified interventions based on integrated theories in various contexts, such as healthcare and fitness. Studies were included if an integrated theoretical intervention was implemented to change behaviour in specific contexts. The review aims to uncover the effectiveness of integrated theory in predicting behaviour outcome in interventions. Our findings reveal that in 39 studies, Self Determination Theory (n = 19) and Theory of Planned Behaviour (n = 16) outnumbered other theories in integrated models. Overall, 77% of studies showed evidence that integrated theoretical-based behaviour change interventions can be successful for a short time, with only a few studies that tested these interventions’ long term effects. We discuss the implication of our findings, and also propose potential future directions. Keywords: Integrated theories · Hybrid · Gamification · Intervention · Behaviour change · Health and fitness

1 Introduction Essentially all lifestyle-related health risks, like non-communicable diseases, are significantly affected by an individual’s health behaviours such as physical activity and food intake. It was reported that globally, approximately 80% of adolescents are physically inactive [1]. Evidence indicates that our wellbeing can be regulated by individual behaviours [2]. Thus, considering the potential risks of sedentary behaviour and physical inactivity [3], behavioural improvement becomes critical to maintaining a healthy lifestyle. One of the critical drivers of behavioural change in an individual is motivation. A key challenge in the health and fitness context is to keep users interested, inspired and engaged in maintaining physical activity. Gamification which refers to the use of game elements in a non-gaming context has been proposed to positively impact on both health © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 263–278, 2021. https://doi.org/10.1007/978-3-030-70713-2_26

264

A. S. Mustafa et al.

behaviour change and adherence [4, 5]. The concept of gamification has been applied in several context, such as education and online learning [6, 7], workplace [8] and travel [9], through game elements such as leaderboards, badges, and challenges. Gamification seeks to make activities more engaging, exciting, and promote a long-term behaviour outcome. Examples of apps and services that use gamification for user engagement include Fitbit, Duolingo, Foursquare, Khan Academy, Coursera, and Grab. However, for gamification to be effective in these conditions, the platform needs to encourage users to adhere to the interventions and remain engaged. This refers to continued usage behaviour or post-adoption behaviour, which is the repeated usage activity of users after the system is adopted, often measured by the frequency of use. Recent research, however, affirmed that the strengthening of the continuous use behaviour is related to maintaining long-term behaviour outcome, thus promoting the long-term stability and overall efficacy of the intervention [10, 11]. In total, interventions aim to increase motivation for sustained use of apps towards performing certain behaviours [12]. Nevertheless, to achieve this, it has become critical to provide a theoretical understanding of predicting behaviour. Therefore, theory application is essential in identifying the causal determinants of change and the effects of the intervention [13]. Indeed, theoretical approaches that enable theories to be tested to determine the most effective and best fit are now widely used in the health and fitness-related literature [14–16]. In this regard, Schoeppe and colleagues [17] noted varying degrees of effectiveness in behavioural change theories to explain behavioural health improvements and these theories’ efficacy. According to [18], addressing multiple theories helps develop more extensive interventions by leveraging each theory’s strength. Thus, identifying the integrated theories in this review will strengthen the use of relevant theory and the objectivity with which they are implemented. This will improve understanding of how multiple theories have been revised and combined over time and their usefulness in predicting behaviour. Moreover, the findings have implications for creating a better understanding of predicting behaviour through theory integration. This paper is organised as follows: Sect. 2 outlines and discusses theory integration in behaviour change interventions. We then include an overview of our review procedure in Sect. 3. Section 4 summarises the key conclusions of our study. Section 5 discusses the results of our review. We end with a brief conclusion, identifying limitations and potential future directions in Sect. 6.

2 Integration of Theories Theory integration refers to combining the variables or constructs of two or more theories to form an integrated theory, a hybrid or modified model. Essentially, integrated theories illustrate key psychological factors and processes that help predict and explain a behavioural change in hybrid-based health interventions by eliminating redundancy while leveraging each theory’s strength [19, 20]. Notably, some authors have extended existing theories by adding useful constructs from one model and combining them with another model [21, 22]. In other words, this can be referred to as theory advancement.

A Systematic Review of the Integration of Motivational and Behavioural Theories

265

While the literature shows that multiple theories’ constructs can work together to predict behaviour, they have the common limitation of testing only one or more constructs of each theory instead of all variables [19]. In addressing this, Noar and Zimmerman [19] suggest integrating theories to expand our understanding of the constructs’ influence while evaluating individual theories to help guide their integration. This helps us to equate the individual theoretical models with integrated ones and examine their benefits. Thus the main goal is to provide modern-day views on applied theoretical research in the health domain to understand the processes better. In turn, it may lead to improvements in health behaviour and related outcomes. Indeed, this intends to provide a simplified and detailed view of the factors that influence health behaviour, thus eliminating theoretical gaps, reducing complexities and increasing parsimony [23]. Therefore, while theoretical integration is feasible, further studies are necessary before drawing clear conclusions. In this context, the first step is analysis of the existing literature.

3 Review Procedure 3.1 Search Strategy A systematic search of the scholarly literature was performed using DOAJ, Emerald Insight, Google Scholar, IEEE Xplore, GSPI, PubMed, Science Direct, Scopus, Web of Science and Wiley from April to July 2020. To locate relevant papers on integrated theories in behavioural interventions, we used a mixture of search terms to identify all related literature in the following categories: “integrated,” “theories,” “integ,” “comb,” “models,” “behaviour” “gamification”. We reviewed 88 potentially relevant articles, but only 22 of them met the study criteria. Specifically, we identified 76 records via database search, with 12 additional records from other sources. We then excluded 44 articles with a different focus. We also conducted a forward and backward search and identified an additional 17 related articles. Finally, we selected 39 articles from June 1999 to June 2020 for this review (see Fig. 1 and Table 4). 3.2 Data Analysis Studies were categorised as quantitative, qualitative or mixed-method, using either subjective or objective indicators based on the concepts of [24]. We were especially interested in how the studies integrated theories into the behavioural change process through theoretical lenses, including the discovery of theoretical approaches as well as assessing their degree of effectiveness. Thus, an intervention was determined to be successful on based on the outcome and categorised as (i) positive, (ii) negative, (iii) neutral (no effect). For the classification of theories, each study must include at least one motivational or behavioural theory. Finally, for the analysis, we included the following categories in the framework: study sample, study length, data collection, analysis, study location and outcome.

266

A. S. Mustafa et al.

Fig. 1. Flow diagram of the literature review process.

4 Analysis The interventions were categorised based on the population size, study groups, research time frame, study location and the study outcome. The selected articles in this study predominantly used quantitative methods (77%) other articles employed a mixed approach. The majority of studies adopted the survey method for data collection. The number of participants in the studies ranges from 46 [38] to 8840 [50], with age groups between 17 and 82. Moreover, the intervention strategies were delivered over timescales ranging from two weeks to one year. We observed more publications in journals (n = 36) than in conference proceedings (n = 2) and dissertations (n = 1). These findings were significantly different from that of [6], which highlighted that most papers in their analysis were conference proceedings. Then, we examined the number of publications over the time frame and found that the studies were conducted between 1999 and 2020. Notably, a total of seventeen papers were published between 2017 and 2020, with three in 2017, four in 2018, five in 2019, and five in 2020. This confirms a growing trend in the number of publications in the context of hybrid theory-based interventions. The analysis highlighted that the majority of studies were conducted in China (n = 7), followed by Taiwan (n = 5) and the USA (n = 4) (see Fig. 2). 4.1 What Theories Were Targeted? This subsection summarises the main properties of the behaviour change interventions from the analysis. We identified 21 different theories featured in the studies (see Table 1). The Self-Determination Theory (SDT) is extensively used in 49% (n = 19) of the studies, followed by Theory of Planned Behaviour (TPB) in 41% (n = 16) of studies,

A Systematic Review of the Integration of Motivational and Behavioural Theories

267

Fig. 2. Number of studies per country.

and the Technology Adoption Model (TAM) in 28% (n = 11) of studies. Significantly, most results were positive in studies that integrated any of the three theories (SDT, TPB or TAM). This finding provides additional support for SDT’s suitability for integration with other theories. Table 1. Frequency of theories used in the studies. Theories

No. of studies

% of studies

Self-Determination Theory (SDT)

19

49%

Theory of Planned Behaviour (TPB)

16

41%

Technology Adoption Model (TAM)

11

28%

Task Technology Fit (TTF)

9

23%

Expectation Confirmation Model (ECM)

5

13%

Self-Efficacy Theory (SET)

4

10%

Social Cognitive Theory (SCT)

3

8%

Unified Theory of Acceptance and Use of Technology (UTAUT); Flow theory; Social capital theory

2

5%

Other theories (see Table 4)

1

3%

4.2 What Behaviours Were Targeted? Table 2 shows the behaviours targeted by the studies. Cugelman [25], underlines that gamification’s long-term effect is more significant than the short-term impact. Based on this, we found that eight health behaviours were targeted (n = 14) with mainly positive outcomes. However, studies that did not target health-related behaviour outcomes focused on online learning, information system, entrepreneurship, workplace, social network and gaming. As shown in Table 3, the most recurrently studied health area is Health, Fitness & Physical Activity (n = 13). The primary behavioural outcome here is to increase

268

A. S. Mustafa et al. Table 2. Targeted behaviours in the studies.

Behavioural contexts targeted

No. of studies (%)

Studies

Health, fitness & physical activity

14 (36)

[14, 15, 18, 32–36, 39, 40, 42, 50, 51, 57]

Online learning

10 (26%)

[28, 43, 45, 46, 52, 54–56, 58, 59]

Mobile shopping & banking

5 (23%)

[16, 31, 44, 49, 60]

ICT, information system & 2 (5%) KMS

[48, 53]

Entrepreneurship

2 (5%)

[26, 27]

Workplace

2 (5%)

[30, 37]

Social networking

2 (5%)

[29, 47]

Gameplay & online

2 (5%)

[38, 41]

physical activity. The remaining papers focused on Physical Education and Leisure, Myopia Prevention, Influenza Prevention, Exercise & Diet, Post-cardiac Rehabilitation, and Blood Donation, highlighting the growing number of studies in this domain. Table 3. Targeted health behaviours in the studies. Health behaviour targeted

No. studies (%)

Studies

Physical Activity (PA)

5 (37%)

[14, 15, 18, 34, 40]

Exercise and diet

3 (21%)

[39, 42, 50]

Physical Education (PE) and leisure

2 (14%)

[32, 33]

Seasonal influenza prevention

1 (7%)

[36]

Myopia prevention

1 (7%)

[35]

Post-cardiac rehabilitation

1 (7%)

[51]

Blood donation

1 (7%)

[57]

4.3 What Groups Were Targeted? We observed that the target categories of featured interventions varied. Most of the studies addressed a specific group, such as university students, programme analysts, online players, e-learning users, instructors, m-banking users and employees. On the other hand, only three studies targeted healthcare patients [34, 35, 51]. In our view, the result emphasises that it may be easier to implement hybrid theory-based gamification studies in a non-health context. The sampling designs used in the studies include of

A Systematic Review of the Integration of Motivational and Behavioural Theories

269

sampling, cohort groups, simple random sample and convenience sampling were used. In total, the majority of the studies used an online survey to recruit their participants. However, in specific cases, the focus group was predefined by the associated behaviour change, such as cardiac rehabilitation [51]. Notably, other studies recruited individuals interested in interventions through online networks. We identified most of the studies to be cross-sectional with only six longitudinal studies applied by scholars. Hence, there is a need for more longitudinal studies to determine the interventions’ long-term impact effectively. From our result, we also found that 30 (77%) of the papers show evidence of significant benefits of integrated theory-based interventions, while three (8%) show adverse effects and six (15%) neutral outcomes or no effect. Also, most empirical works (30 studies) showed positive impacts, which suggests that progress has been made in the context of hybrid-based interventions on behavioural change. However, just three papers [38, 42, 55] tested gamification and theory integration. In two of the studies [38, 42], experimental design (experimental and control groups) was employed and integrated SDT with other theories. The studies’ results reported two positive [42, 55] and one neutral [38] outcomes. Theory-based gamified approaches can effectively promote behavioural outcomes; thus, future experimental and longitudinal research is needed to understand the results better.

5 Discussion Notably, the majority of studies (77%) show significant beneficial outcomes. The analysis confirms that an integrated model can increase engagement and related outcomes, such as intention and behavioural change. We observed a growing interest in combining more than two theories (in five studies), whereas previous studies only combined two theories. Our study reveals the unequal distribution of theoretical use frequency with SDT and TPB as the most implemented theories. This can be attributed to the fact that the theories are selected most often because they have a clear conceptual framework, a positive outcome or that they satisfy other acceptance criteria. Although, this may not be the case, as some studies have shown negative or neutral effects of theory integration (SDT + TPB and SDT + SET). Our findings also indicate that integrated theory-driven gamification intervention for behaviour change are underutilised. Still, evidence demonstrates the potential for hybrid theory and game-based interventions to fit together to successfully predict behaviour outcome [22] (Table 4).

Intervention domains Entrepreneurship; Healthcare; Exercise & physical activity; information system

Theories integrated

SDT + TPB

Studies

[26, 27, 34–36, 39, 40, 42, 48, 57]

N = 3670 (Total) University students, office workers, Parents Australia [40, 57] Belgium [42] China [35, 48] Hong Kong [36] Malaysia [26] Yemen [27] USA [34]

Total sample sizes & characteristics

Table 4. Characteristics of included studies

(continued)

SDT and TPB support explanations of motivation of entrepreneurial behaviour [26] Satisfaction of SDT motivational factors significant in improving student’s intention [27] SDT explains more variance in TPB variables than TPB explains for SDT [34] Weak relationship between intentions and behaviour [34] Integrated model of SDT + TPB can be used to explain myopia-preventive behaviours [35] Intention significantly predicted reading distance [35] Facemask use positive related to intentions mediated by subjective norm, attitude, PBC [36] Indirect effect on exercise behaviour; both direct and indirect effect on diet behaviour [39] PA intentions strongest determinant of behaviour; intention fully mediated by TPB variables [40] PA behaviour positively predicted by intentions towards PA [42] Intervention intensity positively predicted desired changes in fat intake [42] AM strongly influenced IS discontinuation [48] Autonomous motivation predicted intention, and no effect of external regulation [57]

Outcomes

270 A. S. Mustafa et al.

Intervention domains Gameplay; physical activity; ICT training

Workplace; Online learning

Physical activity

Physical activity

Healthcare

Theories integrated

SDT + SET

TAM + TTF

TPB + EPPM

TPB + BPN

SCT + TAM + Social capital theory

Studies

[18, 38, 51, 53]

[37, 55, 58]

[14]

[15]

[16]

Table 4. (continued) Total sample sizes & characteristics

N = 365; Students Taiwan

N = 462; Athletes Australia

N = 336; Students Australia

N = 731 (Total) Prog analysts, MOOC users & instructors USA [37], China [58], North Cyprus [55]

N = 567 (Total) Cardiac-rehabilitation patients [51] Canada [18, 51] Thailand [53]

(continued)

Over 80% of all relationships proposed in the integrated model supported Integrated model demonstrated excellent fit and useful for predicting behavioural intention

Integrated model increased explanatory power for predicting exercise behaviour Intentions and PBC predicted intention to continue sports

Integrated model enhanced explanatory power for predicting exercise behaviour than only TPB

TAM + TTF had better explanation for variance in IT utilisation than TAM or TTF [37] PU, PEOU, TTF, social recognition, social influence and attitudes significantly influence continuance intention [55] PU, PEOU, TTF, social recognition, reputation, attitude and social influence strongly predicted continuance intention [58]

Individual and integrated models supported; SDT + SET more favourable over either SDT or SET [18] All psychological needs predicted self-determined motivation in integrated model [18] No major difference among the groups in relation to engagement and performance [38] SDT and SET partially supported but unable to predict physical activity change; [51] Higher Self-determined motivation will predict self-efficacy, satisfaction and usage intention [53]

Outcomes

A Systematic Review of the Integration of Motivational and Behavioural Theories 271

Intervention domains Online learning Social media Workplace

Mobile banking

Physical education

Physical activity Online internet

Online learning

Mobile data service

Theories integrated

Flow theory + ECM + ISSM

TAM + U&G theory

SDT + Social exchange theory

TAM + Trust theory

SDT + TBP + HMIM

SDT + Goal orientation theory

SCT + EDT

SDT + TTF

ECM + TPB

Studies

[28]

[29]

[30]

[31]

[32]

[33]

[41]

[43]

[44]

Table 4. (continued)

N = 207; Graduate Students South Korea

N = 414 Pakistan

N = 235; Internet users Taiwan

N = 723; 28 Schools Hungary

N = 274; College students Greece

N = 219; M-banking users Ethiopia

N = 453; Office workers China

N = 372 Students, UAE

N = 515; College students UAE

Total sample sizes & characteristics

(continued)

Integrated model has stronger explanatory power of continuance than ECM or TPB alone Subject norm, PBC User satisfaction, PF, PU, and PE have significant impact on continuance intention

TTF positively influenced behavioural intentions PC, perceived relatedness and social recognition strongly influence behavioural intentions

Continuance significantly related to SCT + EDT constructs Continuance intention predicted satisfaction, internet self-efficacy, and outcome expectations

Self-determined forms of behavioural regulation main predictor of intention

3 Basic psychological need satisfaction variables exclusively predicted autonomous motivation in PE

Attitude and trust mutually explain 50% variance in continuance intention to use m-banking

Need satisfaction mediated relationship between overall justice and intrinsic motivation

The integrated model strongly predicted users’ intention

PU, PE and PEOU have the strongest effect on continuous intention to use

Outcomes

272 A. S. Mustafa et al.

Intervention domains Online learning

Online learning

Social network Mobile shopping

Healthcare

Online learning

Online learning

Theories integrated

PAM + TTF

ECM + TPB + TAM + flow theory

TTF + Social capital theory

TAM + ECM

HBM + UTAUT

SDT + TAM

TAM + UTAUT

Studies

[45]

[46]

[47]

[49]

[50]

[52]

[54]

Table 4. (continued)

N = 305; University students The Netherlands

N = 174; UN staff USA

N = 8840; Students, workers China

N = 203; Mobile shoppers China

N = 315; Students, workers Taiwan

N = 363; e-learning students Taiwan

N = 135 Norway

Total sample sizes & characteristics

(continued)

The UTAUT model strongly predicts the perceived acceptance of MOOCs Attitude strongly effects behavioural intention

Three basic psychological needs have significant indirect effects on continuance intention Stronger influence of PU on continuance intention than perceived playfulness

Users’ risk perception negatively affected actual usage behaviour Actual usage behaviour positively affected weight-loss intention and behavioural intention

PU did not motivate all user groups Satisfaction and PEOU significantly influence different user groups

The fit between social characteristics and tech characteristics impacts users’ intentions

Continuance intention strongly affected by satisfaction and PBC (lesser predictor) PU, subjective norms, concentration and attitude moderately effect continuance intention

Both TTF and PAM variables explain continuance intention Users’ satisfaction influences the development of strong intention about IS continuance

Outcomes

A Systematic Review of the Integration of Motivational and Behavioural Theories 273

KMS

Online learning

M-banking

SCT + TTF

TPB + TTF

TAM + TTF + ECM

[56]

[59]

[60]

N = 43; Telecom utilities China

N = 870; Students Taiwan

N = 192; KMS users Taiwan

Total sample sizes & characteristics

Satisfaction, TTF, PU and perceived risk strong predictors of continuance intention

User attitudes towards subjective norms, and PBC indirectly impact utilisation Some features of utilisation determined by user perceptions and behavioural intentions

Integrated model explains about 50% of variance in KMS usage Task interdependence, TTF, self-efficacy and personal outcome significantly impact KMS usage

Outcomes

SDT: Self-Determination Theory, TAM: Technology Acceptance Model, TTF: Task Technology Fit, U&G: Uses and Gratifications Theory, TPB: Theory of Planned Behaviour, SET: Self Evaluation Theory, ECM: Expectation Confirmation Model, PAM: Post Acceptance Model, UTAUT: Unified Theory of Acceptance and Use of Technology, ISSM: Information System Success Model, HBM: Health Behavioural Model, SCT: Social Cognitive Theory, EDT: Expectancy Disconfirmation Theory, BPN: Basic Psychological Needs; HMIM: Hierarchical Model of Intrinsic Motivation, KMS: Knowledge Management System

Intervention domains

Theories integrated

Studies

Table 4. (continued)

274 A. S. Mustafa et al.

A Systematic Review of the Integration of Motivational and Behavioural Theories

275

6 Conclusion In this paper we reviewed studies that integrated several related theories in behaviour change interventions. To our knowledge, this is the first review conducted on integrated theory-based intervetions used to predict health behaviour outcome. The results revealed that SDT is the most implemented theory, followed by TPB. The majority of the papers reviewed showed a positive effect on behavioural outcomes, but these studies’ positive effect was also short-term. The study findings generally reinforce the use of behavioural change interventions based on theory integration. However, relatively few studies have tested integrated-theory driven behaviour change interventions, hence the need for more intervention research in line with [21]. This raises awareness of some limitations of theoretical-based research in behavioural sciences, thus highlighting the alternatives that theory integration offers. Since positive behavioural outcomes are often short-term, further integrated theory-based intervention research is required to predict and sustain long-term behavioural outcome in the health context. Acknowledgement. The authors would like to acknowledge the financial support of Universiti Tenaga Nasional under the Bold Research Grant (RJO10517844/012) and the Innovative Research Management Center (iRMC) UNITEN.

References 1. Trigueros, R., Aguilar-Parra, J.M., Cangas, A.J., Lopez-Liria, R., Alvarez, J.F.: Influence of physical education teachers on motivation, embarrassment and the intention of being physically active during adolescence. Int. J. Environ. Res. Public Health 16(13), 2295 (2019) 2. Sola, D., Couturier, J., Voyer, B.: Unlocking patient activation in chronic disease care. Br. J. Healthc. Manag. 21(5), 220–225 (2015) 3. Wu, X.Y., Han, L.H., Zhang, J.H., Luo, S., Hu, J.W., Sun, K.: The influence of physical activity, sedentary behavior on health-related quality of life among the general population of children and adolescents: a systematic review. PloS One 12(11), e0187668 (2017) 4. Deterding, S., Dixon, D., Khaled, R., Nacke, L.: From game design elements to gamefulness: defining “gamification”. In: Proceedings of the 15th International Academic MindTrek Conference: Envisioning Future Media Environments, pp. 9–15, September 2011 5. Landers, R.N., Armstrong, M.B., Collmus, A.B.: How to use game elements to enhance learning: applications of the theory of gamified learning. In: Serious Games and Edutainment Applications, pp. 457–483. Springer, Cham (2017) 6. Khalil, M., Wong, J., de Koning, B., Ebner, M., Paas, F.: Gamification in MOOCs: a review of the state of the art. In: 2018 IEEE Global Engineering Education Conference (EDUCON), pp. 1629–1638IEEE, April 2018 7. Antonaci, A., Klemke, R., Specht, M.: The effects of gamification in online learning environments: a systematic literature review. Informatics 6(3), 32 (2019) 8. Suh, A., Cheung, C.M., Ahuja, M., Wagner, C.: Gamification in the workplace: the central role of the aesthetic experience. J. Manag. IS 34(1), 268–305 (2017) 9. Yen, B.T., Mulley, C., Burke, M.: Gamification in transport interventions: another way to improve travel behavioural change. Cities 85, 140–149 (2019) 10. Bhattacherjee, A.: Understanding information systems continuance: an expectationconfirmation model. MIS Q. 25, 351–370 (2001)

276

A. S. Mustafa et al.

11. Sailer, M., Hense, J., Mandl, H., Klevers, M.: Fostering development of work competencies and motivation via gamification. In: Competence-Based Vocational and Professional Education, pp. 795–818. Springer, Cham (2017) 12. Greaves, C.J., Sheppard, K.E., Abraham, C., Hardeman, W., Roden, M., Evans, P.H., Schwarz, P.: Systematic review of reviews of intervention components associated with increased effectiveness in dietary and physical activity interventions. BMC Public Health 11(1), 1–12 (2011) 13. Davis, R., Campbell, R., Hildon, Z., Hobbs, L., Michie, S.: Theories of behaviour and behaviour change across the social and behavioural sciences: a scoping review. Health Psychol. Rev. 9(3), 323–344 (2015) 14. Richards, J.A., Johnson, M.P.: A case for theoretical integration: combining constructs from the theory of planned behavior and the extended parallel process model to predict exercise intentions. SAGE Open 4(2) (2014). https://doi.org/10.1177/2158244014534830 15. Gucciardi, D.F., Jackson, B.: Understanding sport continuation: an integration of the theories of planned behaviour and basic psychological needs. J. Sci. Med. Sport 18(1), 31–36 (2015) 16. Tsai, C.H.: Integrating social capital theory, social cognitive theory, and the technology acceptance model to explore a behavioral model of telehealth systems. Int. J. Environ. Res. Public Health 11(5), 4905–4925 (2014) 17. Schoeppe, S., Alley, S., Van Lippevelde, W., Bray, N.A., Williams, S.L., Duncan, M.J., Vandelanotte, C.: Efficacy of interventions that use apps to improve diet, physical activity and sedentary behaviour: a systematic review. Int. J. Behav. Nutr. Phys. Act. 13(1), 127 (2016) 18. Sweet, S.N., Fortier, M.S., Strachan, S.M., Blanchard, C.M.: Testing and integrating self-determination theory and self-efficacy theory in a physical activity context. Can. Psychol./Psychologie Canadienne 53(4), 319 (2012) 19. Noar, S.M., Zimmerman, R.S.: Health behavior theory and cumulative knowledge regarding health behaviors: are we moving in the right direction? Health Educ. Res. 20(3), 275–290 (2005) 20. Marteau, T., Dieppe, P., Foy, R., Kinmonth, A.L., Schneiderman, N.: Behavioural medicine: changing our behaviour. BMJ J. 332(7539), 437–438 (2006) 21. Hagger, M.S., Hamilton, K.: Changing behaviour using integrated theories. In: The Handbook of Behavior Change (2020) 22. Liebana-Cabanillas, F., Munoz-Leiva, F., Sanchez-Fernandez, J.: A global approach to the analysis of user behavior in mobile payment systems in the new electronic environment. Serv. Bus. 12(1), 25–64 (2018) 23. Hagger, M.S.: Theoretical integration in health psychology: unifying ideas and complementary explanations. Br. J. Health. Psychol. 14(2), 189–194 (2009) 24. David, M., Sutton, C.D.: Social Research: The Basics, vol. 74, no. 3. Sage, Thousand Oaks (2004) 25. Cugelman, B.: Gamification: what it is and why it matters to digital health behavior change developers. JMIR Serious Games 1(1), e3 (2013) 26. Al-Jubari, I., Hassan, A., Liñán, F.: Entrepreneurial intention among university students in Malaysia: integrating self-determination theory and the theory of planned behavior. Int. Entrep. Manag. J. 15(4), 1323–1342 (2019) 27. Al-Jubari, I.: College students’ entrepreneurial intention: testing an integrated model of SDT and TPB. Sage Open 9(2), 2158244019853467 (2019) 28. Al-Maroof, R.S., Salloum, S.A.: An Integrated model of continuous intention to use of google classroom. In: Recent Advances in Intelligent Systems and Smart Applications, pp. 311–335. Springer, Cham (2020) 29. Al-Maroof, R.S., Salloum, S.A., AlHamadand, A.Q.M., Shaalan, K.: A unified model for the use and acceptance of stickers in social media messaging. In: International Conference on Advanced Intelligent Systems and Informatics, pp. 370–381. Springer, October 2019

A Systematic Review of the Integration of Motivational and Behavioural Theories

277

30. Aryee, S., Walumbwa, F.O., Mondejar, R., Chu, C.W.: Accounting for the influence of overall justice on job performance: integrating self-determination and social exchange theories. J. Manag. Stud. 52(2), 231–252 (2015) 31. Asnakew, Z.S.: Customers’ continuance intention to use mobile banking: development and testing of an integrated model. Rev. Socionetwork Strateg. 14, 123–146 (2020) 32. Barkoukis, V., Hagger, M.S., Lambropoulos, G., Tsorbatzoudis, H.: Extending the transcontextual model in physical education and leisure-time contexts: examining the role of basic psychological need satisfaction. Br. J. Educ. Psychol. 80(4), 647–670 (2010) 33. Biddle, S., Soos, I., Chatzisarantis, N.: Predicting physical activity intentions using goal perspectives and self-determination theory approaches. Eur. Psychol. 4(2), 83 (1999) 34. Brooks, J.M., Iwanaga, K., Chiu, C.Y., Cotton, B.P., Deiches, J., Morrison, B., Moser, E., Chan, F.: Relationships between self-determination theory and theory of planned behavior applied to physical activity and exercise behavior in chronic pain. Psychol. Health Med. 22(7), 814–822 (2017) 35. Chan, D.K.C., Fung, Y.K., Xing, S., Hagger, M.S.: Myopia prevention, near work, and visual acuity of college students: integrating the theory of planned behavior and self-determination theory. J. Behav. Med. 37(3), 369–380 (2014) 36. Chung, P.K., Zhang, C.Q., Liu, J.D., Chan, D.K., Si, G., Hagger, M.S.: The process by which perceived autonomy support predicts motivation, intention, and behavior for seasonal influenza prevention in Hong Kong older adults. BMC Public Health 18(1), 1–9 (2018) 37. Dishaw, M.T., Strong, D.M.: Extending the technology acceptance model with task–technology fit constructs. Inf. Manag. 36(1), 9–21 (1999) 38. Jamshidifarsani, H., Tamayo-Serrano, P., Garbaya, S., Lim, T., Blazevic, P.: Integrating selfdetermination and self-efficacy in game design. In: International Conference on Games and Learning Alliance, pp. 178–190. Springer, Cham, December 2018 39. Hagger, M.S., Chatzisarantis, N.L., Harris, J.: From psychological need satisfaction to intentional behavior: testing a motivational sequence in two behavioral contexts. Pers. Soc. Psychol. Bull. 32(2), 131–148 (2006) 40. Hamilton, K., Cox, S., White, K.M.: Testing a model of physical activity among mothers and fathers of young children: integrating self-determined motivation, planning, and the theory of planned behavior. J. Sport Exerc. Psychol. 34(1), 124–145 (2012) 41. Hsu, M.H., Chiu, C.M., Ju, T.L.: Determinants of continued use of the WWW: an integration of two theoretical models. Ind. Manag. Data Syst. 104, 766–775 (2004) 42. Jacobs, N., Hagger, M.S., Streukens, S., De Bourdeaudhuij, I., Claes, N.: Testing an integrated model of the theory of planned behaviour and self-determination theory for different energy balance-related behaviours and intervention intensities. Br. J. Health. Psychol. 16(1), 113–134 (2011) 43. Khan, I.U., Hameed, Z., Yu, Y., Islam, T., Sheikh, Z., Khan, S.U.: Predicting the acceptance of MOOCs in a developing country: application of task-technology fit model, social motivation, and self-determination theory. Telematics Inform. 35(4), 964–978 (2018) 44. Kim, B.: An empirical investigation of mobile data service continuance: incorporating the theory of planned behavior into the expectation–confirmation model. Expert Syst. Appl. 37(10), 7033–7039 (2010) 45. Larsen, T.J., Sorebo, A.M., Sorebo, O.: The role of task-technology fit as users’ motivation to continue information system use. Comput. Hum. Behav. 25(3), 778–784 (2009) 46. Lee, M.C.: Explaining and predicting users’ continuance intention toward e-learning: an extension of the expectation–confirmation model. Comput. Educ. 54(2), 506–516 (2010) 47. Lu, H.P., Yang, Y.W.: Toward an understanding of the behavioral intention to use a social networking site: an extension of task-technology fit to social-technology fit. Comput. Hum. Behav. 34, 323–332 (2014)

278

A. S. Mustafa et al.

48. Luqman, A., Masood, A., Ali, A.: An SDT and TPB-based integrated approach to explore the role of autonomous and controlled motivations in “SNS discontinuance intention.” Comput. Hum. Behav. 85, 298–307 (2018) 49. Shang, D., Wu, W.: Understanding mobile shopping consumers’ continuance intention. Ind. Manag. Data Syst. 117, 213–227 (2017) 50. Wei, J., Vinnikova, A., Lu, L., Xu, J.: Understanding and predicting the adoption of fitness mobile apps: evidence from China. Health Commun. 1–12 (2020) 51. Sweet, S.N., Fortier, M.S., Strachan, S.M., Blanchard, C.M., Boulay, P.: Testing a longitudinal integrated self-efficacy and self-determination theory model for physical activity post-cardiac rehabilitation. Health Psychol. Res. 2(1), 1008 (2014) 52. Roca, J.C., Gagne, M.: Understanding e-learning continuance intention in the workplace: a self-determination theory perspective. Comput. Hum. Behav. 24(4), 1585–1604 (2008) 53. Techatassanasoontorn, A.A., Tanvisuth, A.: The integrated self-determination and selfefficacy theories of ICT training and use: the case of the socio-economically disadvantaged. GlobDev 2008, p. 19 (2008) 54. Kamp, C.V.D.: Acceptance of MOOCs by Dutch university students. Extending the unified theory of acceptance and use of technology (UTAUT) model with the technology acceptance model (TAM) (2019) 55. Vanduhe, V.Z., Nat, M., Hasan, H.F.: Continuance intentions to use gamification for training in higher education: integrating the technology acceptance model (TAM), social motivation, and task technology fit (TTF). IEEE Access 8, 21473–21484 (2020) 56. Lin, T.C., Huang, C.C.: Understanding knowledge management system usage antecedents: an integration of social cognitive theory and task technology fit. Inf. Manag. 45(6), 410–417 (2008) 57. Williams, L.A., Sun, J., Masser, B.: Integrating self-determination theory and the theory of planned behaviour to predict intention to donate blood. Transfus. Med. 29, 59–64 (2019) 58. Wu, B., Chen, X.: Continuance intention to use MOOCs: integrating the technology acceptance model (TAM) and task technology fit (TTF) model. Comput. Hum. Behav. 67, 221–232 (2017) 59. Yu, T.K., Yu, T.Y.: Modelling the factors that affect individuals’ utilisation of online learning systems: an empirical study combining the task technology fit model with the theory of planned behaviour. Br. J. Educ. Technol. 41(6), 1003–1017 (2010) 60. Yuan, S., Liu, Y., Yao, R., Liu, J.: An investigation of users’ continuance intention towards mobile banking in China. Inf. Dev. 32(1), 20–34 (2016)

Adopting React Personal Health Record (PHR) System in Yemen HealthCare Institutions Ziad Saif Alrobieh1(B) , Dhiaa Faisal Alshamy2 , and Maged Nasser3 1 Department of Communication and Computer Engineering, Alsaeed Faculty for Engineering

and Information Technology, Taiz University, Taiz, Yemen 2 Department of Networking and Distributed Systems, Faculty of Information Technology and Engineering, Taiz University, Taiz, Yemen 3 School of Computing, Universiti Teknologi Malaysia, 81310 Johor Bahru, Johor, Malaysia

Abstract. Health care is a critical sector of society that requires quality improvement of healthcare services, information technology IT systems have a great impact on improving the quality of these services, unfortunately, The effect of information culture on the implementation of information systems by healthcare providers in the developed countries is little known, considering the importance of information culture. Despite that many Yemeni healthcare facilities have been already using information systems to digitize the management of healthcare providing procedures, The patients’ health information including disease history and prescriptions is not fully recorded and additionally, there is no implementation of Personal Health Record PHR systems where patients can access, and control their health records from another place where their records are stored locally in the healthcare providers’ databases. The existing electronic health record systems are limited and do not exploit the available technology solutions and services. To explore the advantages of using PHR systems, multiple kinds of research are being studied and the proof of use has been cleared by these researchers’ conclusions, also a survey was made; to ensure people intention to using PHR systems, and to observe their opinion on what they need to be provided by the system and what interests them. Designing the web-based system was done after going through the available platforms to choose the most correct and suitable solutions to assure that the system meets the requirements needed. Adopting innovative and modern technological solutions such as PHR web applications is a fine way to improve patient safety and quality of care, increase efficiency, Decision supporting, and Increase patient and health workers’ satisfaction. The proposed solution ensured the patients’ satisfaction and safety by giving them access to their health records whenever and wherever they are through their device’s browser and helped the doctors to make the right decisions and speed up the healthcare providing process which reduced the damage caused by the current systems, saving human lives and preventing serious health issues. Keywords: Health records systems · PHR systems · Healthcare · Technological solutions

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 279–289, 2021. https://doi.org/10.1007/978-3-030-70713-2_27

280

Z. S. Alrobieh et al.

1 Introduction Healthcare is a remarkable sector in which we must be evolving every technological solution or invention we reach; we must try hard to give a better and innovative way of solving the problems in this particular sector. The way healthcare systems are managed is radically changed by technical growth, more specifically by the digital revolution [1]. Broadly, the introduction of the new innovations will allow governments to deliver value-added services to people and shows a range of developments in technology, firsthand health monitoring, and medical care [16]. Despite the remarkable development in the healthcare sector in the whole world, Yemen still has a lot of deprivations and obstacles related to this sector. For Yemeni public health, it has been stated that the information has still not been recognized as a culture. The potential of IT in the health sector is still not properly utilized [3]. As a result of a 4 year ago study on the quality of handwritten prescriptions in Sana’s, Yemen [2] which covers 2178 prescriptions from 23 randomly selected pharmacies with different geolocation considered, 99.12% of the analyzed prescriptions were considered as low-quality prescriptions which have writing errors related to physicians and patient information and to the prescribed medications where spelling and instruction of use are the most errors found. Thus, the current healthcare systems in Yemen as in many other countries don’t exploit the technology revolution and the available software technologies (e.g.: APIs, frameworks, services…etc.) which would make a majestic impact on the treatment process and caregiving procedures if they were implemented. For such systems to meet most of these requirements, they must use web-based solutions that support data sharing. Easy access and availability can be achieved by using appropriate platforms to improve interactivity and make the web app more user friendly and responsive and allow developers to create web-based mobile-friendly applications [17] while getting significant benefits from using specific APIs (i.e. symptoms, a large number of drugs containing information and conflicts) to make it easier, faster and safer to choose the right medicine and automatically provide suggestions and help with making more accurate decisions. The main aim of this paper is to encourage healthcare providers in Yemen to adopt PHR systems by showing them all the benefits that can be gained and additionally direct them with a design proposal along with the required technologies, this design will expose new, innovative technologies to ensure an appropriate design supported by health care APIs for instance which allow healthcare providers to access and use electronic applications and data in more innovative ways [18] than those used in current EHRs and unfortunately, in our country and many other countries, healthcare institutions still use desktop-based HealthCare applications, and even if they are; users -patient- don’t have access to their health data. The complexity and time cost of fully implementing healthcare applications in the health providing procedures increases while the user experience and ease of use are not taking into account at the beginning of building the system, i.e. doctors can’t use these systems to store patients symptoms and prescriptions, because it takes more time or requires full attention or specific conditions in the case of speechrecognition input methods these problems can be solved if the designer of the system takes an effort to make the application easier to use and more simple and time-efficient,

Adopting React Personal Health Record (PHR) System

281

it must benefit the doctors and helps them makes good decisions and prevents common errors, it also must speed up the process not slow it down. This can be achieved easily using modern mobile applications, to make use of the application from every device and from different and multiple places as health data need to be made available wherever or whenever, users can simply use their phones, tablets and any portable device’s browser with a decent internet connection. This allows individuals to access their PHRs via the Internet, using state-of-the-art security and privacy controls, at any time and from any location [12]. Recent web-based technologies opened the door for completely new possibilities for creating various medical information systems. Web-based applications are offering competitive benefits to old-style software-based systems, permitting businesses to consolidate and streamline their systems and processes and decrease costs [10]. Technology is becoming more and more advanced as we can now use IoT devices, sensors to measure and monitor human health metrics, send them into mobile or web applications, and access them through our portable devices and computers. The Personal Health Record (PHR) is an Internet-based set of tools that allows people to add, maintain, access, and coordinate their lifelong health information [12– 15], and make appropriate parts of their own medical and health-related information available to those who need it (specialists, doctors, nurses, family members, etc.) [12]. Whereas EMR is a patient’s health information inside a specific medical institution that is unshared in more than one. EHR has the same meaning of EMR but in addition of sharing it among more than one of an institution, but all the operations that are done on them -both EMR and EHR-, done by medical professionals or staff of the institution and no way to the patient to have hand in it as management or control. PHR comes to include the EHR concept with making the information more flexible and giving the patient the complete power to manage, control, and provide access to it. So PHRs patients are more comfortable with adding information to their health record and review all records at any time. Moreover, privacy is felt due to who can access what exactly in his records regardless of geographical distance. PHR abbreviates all that patients and health providers need to know from any place at any time and makes it easy to share with keeping privacy and giving the patient all his rights to deal with his health information.

2 Related Works This section will cover both related works on persuading healthcare providers to adopt web-based applications, and similar systems proposals. Since this paper aims to encourage healthcare providers to adopt web-based PHR systems, and to provide a proposed system design with the technologies which would be used, and the main idea for this proposed work allowing the patients in Yemen to have their medical records and health data on their own. The idea was to go over multiple researches and applications that discussed the implementation of PHR systems including the advantages and disadvantages of health information access for both patients and doctors. As a motivation for healthcare providers, implementing good electronic health services is an effective step toward making the patient more satisfied, where The quality of healthcare services in Yemen from the patient’s perspective has been studied and

282

Z. S. Alrobieh et al.

researched by Mr. Bashar Mohammed Al-Sofyani [8] in his thesis for the master degree of public health, he concluded that Satisfied patients due to providing good quality of care are more likely to comply with treatment and continue to use services. This will improve utilization and will finally lead to better general health indicators. Albokai et al. [4] improved the Quality of Healthcare by using the Information Technology System in the Hospitals of Yemen showed that it improves patient safety and quality of care, increases efficiency, Decision supporting, and Increase patient and health workers satisfaction. However, it was observed that multiple healthcare providers in Yemen have already implemented EHR systems but in limited services and without supporting the information access for patients. Although collecting information about patients is important and critical to construct a proper treatment plan for the patient. This is considered as one of the most important difficulties facing the general medical staff and the private doctors [4]. The adoption of PHRs and EHRs with patient access-support should be considered and “Late adopters of the electronic health record should move now” [9]. A Proposed PHR Architecture for Saudi Arabia Health Services concluded that such a system once approved for adoption in Saudi Arabia, will improve the health services and it will assist in disease prevention and emergency treatment intervention. They also hypothesized that increased patient engagement in their healthcare can improve the quality of the provided services and surely improving their health lifestyle [7]. A study aimed to elaborate the functional specifications of the pregnant woman PHR and to create and propose a prototype, although the study followed some functional principles i.e. (promoting information sharing among women, health professional’s hospitals and diagnostic services and promoting documentation of care), a specific design technology wasn’t provided [12]. P. Thummavet1 and S. Vasupongayya2 [13] propose a novel scheme for handling accesses to PHR information in emergencies, they focused on how to give emergency staff access to PHR information even when the owner is not able to give his/her consents using threshold cryptosystem, based on the owners’ PHR policy. The system consists of three levels of confidentiality (security, restriction, and exclusiveness), the PHR owner can define a confidentiality level to each record before it is uploaded to PHR server and emergency staff will have variety access ability either encryption key through a service provider (EmS) used for encryption or instantly if they are trusted users and pre-selected by the PHR owner. Although This scheme is efficient in case of security, somehow has a level of complexity for simple users. Muhammad H. Aboelfotoh et al. [14] proposed mobile-based system architecture that allows patients to use the online PHR systems that they are subscribed to and at the same time use their portable devices to provide direct data access to physicians using authenticated and integrated Backend infrastructure without fully interconnecting healthcare systems network, However, their proposal requires an existing online PHR system along with additional requirements i.e. (Smart health Card, Healthcare Provider (HCP) terminal) which increase cost and system complexity. Yeong-Tae Song et al. [15] proposed a PHR system that utilizes applications standards such as SNOMED CT, and HL7 CDA to achieve interoperability between different EHRs and PHR systems, a mobile application is used to collect medical data and store it in HL7 CDA format. Their model consists of four main models; Clinical Data Collection

Adopting React Personal Health Record (PHR) System

283

Module: the mobile application is used to collect medical data and generate CDA files that can be uploaded to a cloud-based management system, Cloud file Manager Module: this model used to store the CDA files for each individual, CDA Query Module: which uses XML parsing program to search nodes and extract codes and other values so that the Diagnosis Module can use them as input, Diagnosis Module: the extracted codes will be used to create the clinical decision logic, matches symptoms in the personal medical data to the diagnosis rules, they used Rule-based system CLIPS.

3 Advantages of Using and Improving the PHRs Systems Patients who have medical documents comprise advisory opinions, lab results, prescriptions, and MR, CT, Ultrasound images, and such on. in various formats and forms. That what makes the patient’s medical information is stored in different places according to which institution the patient goes to, which makes the ability to access it, by the patient himself and share it with others, is necessary to face changing places especially for people that travel a lot. What keeps effort, time, and cost that took in repeating diagnoses with repeating examinations and such like. Or the risk of taking incongruity medicine to end a patient’s life. Actually, there is no way to limit the need to use personal health records (PHR) but at the same time, we must ensure high performance, whether in-facility or speed and care about security. So, such advantages can be summarized as follows: • PHR provides a continual monitor for patients’ health status and acquaintances all necessary health information (medical history, medical examination data, physiological parameters, healthy lifestyle, etc.) at any time, anywhere, from any platform [8]. • Educating patients: Patient-accessible medical records improved recall and understanding of medical information by objective measurement in two randomized controlled trials. Among medical outpatients, smokers who received a copy of their most recent progress note were significantly more likely to identify smoking as a problem 2 weeks after their appointment, and this trend persisted at 6 months. 43 Older patients with chronic medical conditions also showed significant increases in their recall of medical problems and treatment plans that did not involve medications. • Empowering patients. • Improving doctor-patient communication. • Improving patient satisfaction. • Patient-accessible medical records are particularly helpful for patients who are concerned about what might be hidden in the chart. • Facilitating correction of errors: Patients found inaccuracies in the medical record in many of the studies. A descriptive study of medical inpatients found that half of the patients “made some addition or correction on a point of fact.” • Effects on documentation: Although both patients and staff had the impression that patient access to the records changed documentation patterns, little change was identified on objective analysis, and made the staff more accurate in what they wrote.

4 Methodology and System Design Several hospitals and healthcare institutions were visited to explore and identify existing EHR systems. Many hospital personnel was questioned and the level of satisfaction was

284

Z. S. Alrobieh et al.

observed with the current systems. In this work, each patient will have his or her health profile, accessible at all times and across mobile or digital devices, to support the existing healthcare systems with productivity, time-saving, and data sharing probability. We first have to measure the level of participants and willingness to use an online PHR system, and what would be the most important part of the system that they need, and will encourage them the most to use the system. And second, by going over multiple proposed PHR systems to collect ideas about what the system should implement. Having this information will help us to propose the most proper and suitable plan for constructing a PHR system for our society. 4.1 Determining Level of Interest Either From the patient’s perspective or the doctors, it’s hard for them to collect the health data, we can see that After making a questionnaire about the effectiveness of the proposed system idea and. The survey contained a group of questions that the respondents had to answer. The target of the survey was every individual at the age older than 15. A total of 131 random individuals were given a link to a Google form that must be submitted in 3 days. The included questions aimed to observe and study the level of excitement and interest to have such ability, it is also aimed to see how much they see the importance of this idea and how useful it would be. And was divided as follows: • Basic information (age-gender-education level- having a chronic disease): People of different ages will have different interests and needs of PHR system parts, As well as gender and education level, additionally the possibility of having a chronic disease will affect the willingness of people to use the system. Although Table 1 shows the level of interest for peoples of different genders and health states, in general, we can see that more people are encouraging this idea and interested in using a PHR system. • Online activities related to healthcare: • Search about some diseases: seeking information about a health issue or disease is critical while some sources may be tricky or misleading. • Search for information about a doctor or contact a doctor by email or specific application. • Used a PHR or health-related application before: having an experience will have a different impact on the level of participation. • Received a notification about a test result or browse his medical prescriptions with his mobile. • Feelings about the idea of having PHR Four levels of interest will be provided (highly interested, interested, not interested, highly not interested (thinks that the idea will have a negative impact). As it is shown in Table 1 most of the participants are interested in having and using a PHR system no matter how their health conditions would be.

Adopting React Personal Health Record (PHR) System

285

• Prioritizing system services: Prioritize the system aspects from most important to less important from everyone’s perspective most important service they wish the system will provide, and what they think it’s not that important. Figure 1 is shown the priority of health services from participants’ perspectives. • Concerns and limitations: What are the negative sides of implementing this system and what concerns them and probably prevents them from using the system and if there are limitations?

Table 1. Level of interest based on gender and chronic disease Basic information Gender Have chronic disease

Filter type

Participants number (%) Highly interested

Interested

Not interested

M

31 (41.9%)

28 (37.8%)

15 (20.27%)

F

17 (28.8%)

27 (45.76%)

15 (25.42%)

Yes

5 (55.5%)

3 (33.3%)

1 (11.1%)

No

39 (34.2%)

49 (42.98%)

26 (22.8%)

HEALTH SERVICES PRIORITY FOR INDIVIDUALS sharing health informa on with healthcare providers browsing health instruc ons and informa on childerens health monitoring reminders of medical tests results control and manage family health informa on doctor visits reserva on contac ng doctors adding new priscrip ons for each visit brows test results informa on from a trustable source

160 140 120 100 80

60

40

20

strickly important

0 important

not important

Fig. 1. Health services priority for individuals

Additionally, to gain a better insight into the positive impact on both patients and doctors, and to ensure the effectiveness of the proposed system, by going over multiple

286

Z. S. Alrobieh et al.

researches and articles that proved the usefulness of using PHR systems in general and using web-based systems in specific. 4.2 System Technology Selection To make the application be accessed from every user device from different and multiple places, as health data requires to be available no matter where or when this can be fulfilled easily with the use of mobile applications, users can simply use their phones, tablets, and any portable device’s browser with a decent internet connection. React-NodeJS Web Application The system designer must make an effort to facilitate the interactive, time-efficient, and interactive use of the application. React’s capability assists in designing a simpler user interface that facilitates the application and enhances the user experience. In this section, we will explain more about the React.js framework. React is a component-based library which is used to develop fast interactive UI’s (User Interfaces). It is currently one of the most popular JavaScript front-end libraries which has a strong foundation and a large community supporting it. As list advantages and giving some reasons why React-Js was the chosen technology, we can summarize that in the following points: • Easy creation of dynamic applications: because it requires less coding and offers more functionality, as opposed to JavaScript, where coding often gets complex very quickly. • Improved performance: Where react uses Virtual DOM (VDOM) thereby creating web applications faster. Virtual DOM compares the components’ previous states and updates only the items in the Real DOM that were changed, instead of updating all of the components again, as conventional web applications do. • Reusable components: Components are the building blocks of any React application, and a single app usually consists of multiple components. These components have their logic and controls, and such that can be reused throughout the application, which in turn dramatically reduces the application’s development time. • Unidirectional data flow: This means that when designing a React app, developers often nest child components within parent components. Since the data flows in a single direction, it becomes easier to debug errors and know where a problem occurs in an application at the moment in question. • Small learning curve: React is easy to learn, as it mostly combines basic HTML and JavaScript concepts with some beneficial additions. Still, as is the case with other tools and frameworks, you have to spend some time to get a proper understanding of React’s library. • JSX: JSX stands for JavaScript XML. It’s an XML/ HTML-like syntax used by React. • Virtual DOM: Manipulating real DOM is much slower than manipulating VDOM because nothing gets drawn on the screen. When the state of an object changes, VDOM changes only that object in the real DOM instead of updating all of the objects. • Performance: React uses VDOM, which makes the web applications run much faster than those developed with alternate front-end frameworks. React breaks a complex user interface into individual components, allowing multiple users to work on each component simultaneously, thereby speeding up the development time.

Adopting React Personal Health Record (PHR) System

287

• Extensions: React goes beyond simple UI design and has many extensions that offer complete application architecture support. It provides server-side rendering, which entails rendering a normally client-side only web application on the server, and then sends a fully rendered page to the client. It also employs Flux and Redux extensively in web application development. • One-way data-binding: that means unidirectional data flow as explained previously. • Debugging: React applications are easy to test due to a large developer community. 4.3 System Framework and Users’ Interface Using created auto-complete recommendations from multiple APIs, such as (medicine names, symptoms, prescriptions, and medical instructions), the system’s ease of use and efficiency will be ensured, which will give the PHR some kind of ease and aid in decisionmaking and reduce time costs while medical staff provides this information. APIs can also be used as a source to get valuable information about diseases and medicines; thus, it helps to get the correct information for the patients and prevent misleading, confusing,

This one of the most important part of the system where the prescripƟons for each visit will be recorded and valuable informaƟon and guidance will be stated to help the paƟent and give hem insights about the prescripƟon medicines like Ɵme to take and reminders and duraƟon of the prescripƟon, it will also provide some important instrucƟons the paƟent will have to consider.

Will contain a list of the diseases and medical condiƟons that the paƟents had along with specific dates of treatments and the period of each illness. MulƟple helpful properƟes will be provided i.e. a property called (Illness Sate) will indicate the state of that illness and show either the paƟent has fully recovered or not and other related properƟes according to the illness type.

Every time the patient visits the doctor, the visit details will be recorded such as medical tests ordered by the doctor or the medical diagnosis, and doctors notes about the paƟent condiƟon. It will also contain the list of prescripƟons given by the doctor and associated illness for the current condiƟon if the paƟent have a medical history of specific illness .

Fig. 2. Main PHR system components

The medical reports and tests results that the patient have taken during any treatment session, including the reports provided by the doctors for the patient visits and his medical state , the reports will also include any examinaƟon files as pictures or the other format.

288

Z. S. Alrobieh et al.

and sometimes wrong information from untrusted websites. Several advantages have been proposed by this system, and one of this advantage is when the patient leaves, he will be able to access all his recent treatment procedures and brows his medical reports and prescriptions through his browser anywhere in order to get any information he needs or instructions or medicines he has to take, along with other parts as it is shown in Fig. 2. The second advantage is, if he ends up visiting another doctor or had another medical condition and transferred to another health provider, the medical staff can easily find the information they needed in order to construct a proper treatment plan.

5 Conclusion Seeking health information access is not just claimed from the doctors but in fact, it has already been a legal right in many countries [6]. Adopting PHR system is highly required for the sack of a better HealthCare for every individual in our country, not only because other people in other countries have such system, but having such system in our country would help to increase people’s safety and gives them a great satisfaction about the services of the caregivers. Our study shows that Yemeni people are different from others country’s people and a high percentage of people feel that they need that kind of systems they can use to assets them into having a better health state and make it easy for them to get any information they need without worrying about whether it’s wrong or dangerous to take these medications in their prescriptions, they want to share their health information without going over all the healthcare places they went to, and they are ready to continue using the service that satisfies their needs. Thus, caregivers in our country must make a fast move towards implementing PHR systems. Our study proposed the system components they should put in mind when they decide that they will implement PHR system, that’s what people need, and that’s what they want, they can’t effort the high cost of expensive PHR systems and they fear that they may not be able to use a complex PHR system so it has to be simple and clear to them when they use the application, ease of use and interactivity must be fulfilled; that’s why we recommend using react as a framework to design the system. Our future work will focus on privacy and security when decentralized databases are used and blockchain technology using Hyperledger fabric is implemented. Interoperability is another issue that we can discuss to share healthcare data between multiple systems using scandalized records systems and design the system to support global standards like HL7 and other CDA.

References 1. Laurenza, E., et al.: The effect of digital technologies adoption in healthcare industry: a case based analysis. Bus. Process Manag. J. 24, 1124–1144 (2018) 2. Mohammed Al-Worafi, Y., Patel, R.P., Zaidi, S.T.R., et al.: Completeness and legibility of handwritten prescriptions in Sana’a Yemen. Med. Princ. Pract. 27(3), 290–292 (2018). https:// doi.org/10.1159/000487307 3. Mukred, A., Singh, D., Safie, N.: Investigating the impact of information culture on the adoption of information system in public health sector of developing countries. Int. J. Bus. Inf. Syst. 24(3), 261–284 (2017)

Adopting React Personal Health Record (PHR) System

289

4. Albokai, N., Liu, L., Alragawi, A., Albokai, A.: Improving the quality of healthcare by using information technology system in the hospitals of Yemen. Open J. Bus. Manag. 07, 728–754 (2019). https://doi.org/10.4236/ojbm.2019.72049 5. The impact of patient characteristics and the Internet usage on potential Personal Health Record (PHR) adoption in Primary Care 6. van Mens, H.J.T., Duijm, R.D., Nienhuis, R., de Keizer, N.F., Cornet, R.: Determinants and outcomes of patient access to medical records: systematic review of systematic reviews. Int. J. Med. Inform. 129, 226–233 (2019). https://doi.org/10.1016/j.ijmedinf.2019.05.014. Medline: 31445260 7. Mafawez, A., Qawqzeh, Y.: Proposed PHR architecture for Saudi Arabia health services. J. Eng. Appl. Sci. 4(1), 26–31 (2017) 8. Parkhomenko, A., Tyshchenko, I.: Research and Development of the API for Personal Health Record. CMIS (2019) 9. Rumball-Smith, J., Ross, K.: Bates DW late adopters of the electronic health record should move now BMJ Qual. Saf. 29, 238–240 (2020) 10. Lazakidou, A.: Web-based applications in healthcare. In: Lazakidou, A. (eds.) Web-Based Applications in Healthcare and Biomedicine. Annals of Information Systems, vol. 7. Springer, Boston (2010). https://doi.org/10.1007/978-1-4419-1274-9_9 11. Ariani A., Koesoema A.P., Soegijoko, S.: Innovative healthcare applications of ICT for developing countries. In: Qudrat-Ullah, H., Tsasis, P. (eds.) Innovative Healthcare Systems for the 21st Century. Understanding Complex Systems. Springer, Cham (2017). https://doi.org/10. 1007/978-3-319-55774-8_2 12. Duran, A., Galuscan, A., Muntean, C.: Proposed structure of personal health records for pregnant women. Med. Evol. XVI(1) (2010). Timis, oara 13. Thummavet, P., Vasupongayya, S.: A novel personal health record system for handling emergency situations. In: 2013 International Computer Science and Engineering Conference (ICSEC), Nakorn Pathom, pp. 266–271 (2013). https://doi.org/10.1109/ICSEC.2013. 6694791 14. Aboelfotoh, M.H., Martin, P., Hassanein, H.S.: A mobile-based architecture for integrating personal health record data. In: 2014 IEEE 16th International Conference on e-Health Networking, Applications and Services (Healthcom), Natal, pp. 269–274 (2014). https://doi.org/ 10.1109/HealthCom.2014.7001853 15. Song, Y., Hong, S., Pak, J.: Empowering patients using cloud based personal health record system. In: 2015 IEEE/ACIS 16th International Conference on Software Engineering, Artificial Intelligence, Networking and Parallel/Distributed Computing (SNPD), Takamatsu, pp. 1–6 (2015). https://doi.org/10.1109/SNPD.2015.7176216 16. Aceto, G., Persico, V., Pescapé, A.: The role of information and communication technologies in healthcare: taxonomies, perspectives, and challenges. J. Netw. Comput. Appl. 107, 125–154 (2018) 17. Shahzad, F.: Modern and responsive mobile-enabled web applications. Procedia Comput. Sci. 110, 410–415 (2017) 18. Zayas-Cabán, T., Chaney, K.J., Rucker, D.W.: National health information technology priorities for research: a policy and development agenda. J. Am. Med. Inform. Assoc. 27(4), 652–657 (2020)

Artificial Intelligence and Soft Computing

Application of Shuffled Frog-Leaping Algorithm for Optimal Software Project Scheduling and Staffing Ahmed O. Ameen1(B) , Hammed A. Mojeed1 , Abdulazeez T. Bolariwa1 , Abdullateef O. Balogun1,2 , Modinat A. Mabayoje1 , Fatima E. Usman-Hamzah1 , and Muyideen Abdulraheem1 1 Department of Computer Science, University of Ilorin, PMB 1515, Ilorin, Nigeria

{aminamed,mojeed.ha,balogun.ao1,mabayoje.ma,usman-hamza.fa, muyideen}@unilorin.edu.ng, [email protected] 2 Department of Computer and Information Sciences, Universiti Teknologi PETRONAS, 32610 Bandar Seri Iskandar, Perak, Malaysia

Abstract. Software Project Scheduling Problem is one of the most crucial issues in software development because it includes resources planning; cost estimates, staffing and cost control which if not properly planned affect the timely completion of the software project. Software project scheduling is a problem of scheduling the tasks (work packages) and employees in such a way that the overall project completion time is minimized without violating dependency constraints (tasks dependencies) and being consistent with resource constraints. This study adopts a Search Based Software Engineering approach that focuses on multi-objective optimization for a software project planning using the Shuffled Frog Leaping Algorithm, a memetic meta-heuristic algorithm. The objectives are optimal ordering of work packages without dependency violation and allocation of staff to the work packages such that only employee(s) with required competence(s) are allotted to a given work package. The study was carried out in four stages, namely: frog (solution) representation, definition of the fitness function, implementation of Shuffled Frog Leaping Algorithm and evaluation with a randomly generated Software Project Scheduling Problem. The study concludes that it is possible to find an efficient solution to a Software Project Scheduling Problem by implementing the SFLA than any other traditional computing means which are tedious, error prone and costly. Keywords: Shuffled Frog-Leaping Algorithm · Software Project Scheduling Problem · Software project planning · Search Based Software Engineering

1 Introduction Software development for organizations is a very complex task as it deals with managing people, technologies and business processes [1]. In software development process, effective planning is important because failure to plan and/or poor planning can result in © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 293–303, 2021. https://doi.org/10.1007/978-3-030-70713-2_28

294

A. O. Ameen et al.

unnecessary delays and overhead costs [2]. Due to this uncertainty incurred in planning software project, given timing and budget constraints are often unacceptable; which in turn leads to business critical failures. Software development companies often struggle to deliver projects timely, within budget and with required quality. Possible causes of this problem are poor project scheduling and ineffective team staffing [3]. Therefore, software engineering projects require good software project management techniques to ensure that projects are completed on schedule and within budget [4]. In order to achieve proper planning and management of software project, tasks need to be optimally scheduled and resources be effectively allocated. Scheduling is setting a sequence of time-dependent functions to execute a set of dependent tasks that constitute a project [5]. Dependency of tasks in terms of priority and precedence is very crucial to software project scheduling. There are priorities constraints between tasks in projects, but in addition to these constraints there may be another kind of constraints between tasks based on resource allocation [5]. Apart from considering priority and precedence limitations, scheduling should be carried out in a way to be consistent with resource constraints. Good allocations (team staffing) are very crucial for software projects, since humans are their main resources [8, 20]. The importance of effective software project scheduling cannot be overemphasized when managing the development of medium to large scale projects as it is required to carry out projects that can meet the deadline and budget [8]. Software Project Scheduling Problem (SPSP) is a kind of optimization problem that seeks to find optimal schedule for a software project so that the precedence and resource constraints are satisfied and ensuring that project cost and duration are minimized [3]. This problem has been found to be Non-deterministic Polynomial (NP)-hard [9, 19]. To solve this problem, meta-heuristic evolutionary algorithms such as Genetic Algorithm [10], Ant Colony Optimization [9], Shuffled Frog Leaping (SFL) algorithm [4] and Differential Evolutionary Algorithm [5, 6] have been successfully applied. Majority of these studies however, consider only task scheduling for the formulation of the problem [21–23]. There is a need for studies that combines tasks scheduling and staffing (allocation of jobs to developers) in software development project planning problem. In this work, a memetic approach based on Shuffled Frog-Leaping Algorithm (SFLA) is presented for optimal project scheduling and staffing when the objectives are combined. 1.1 The SFLA Algorithm Shuffled Frog Leaping Algorithm (SFLA) is a novel memetic meta-heuristic first proposed by Eusuff and Lansey [11] for solving combinatorial optimization problems and it was first used to solve problem of water resource in network distribution [12]. The SFLA has been designed as a meta-heuristic to perform an informed search using a heuristic function [13]. The SFLA is a fusion of deterministic and random approaches. The deterministic strategy allows the algorithm to use response surface information effectively to guide the heuristic search as in Particle Swarm Optimization (PSO). The random approach ensures the flexibility and robustness of the search pattern. The SFLA does not specify the individuals belonging to it population rather, it uses an abstract model, called a virtual population [13].

Application of Shuffled Frog-Leaping Algorithm

295

The search begins with a randomly selected population of P frogs (i.e. solutions). The population is partitioned into several m memeplexes (parallel communities) that can evolve independently to search the solution space in different directions. The individual frogs contain ideas (memes) that can be influenced by the ideas of other frogs within each memeplex and evolve through an optimization process refeered to as memetic evolution [14]. Memetic evolution enhance the quality of worst frog Xw and guide its performance towards a goal. To ensure that the evolution process is competitive, it is required that frogs with better memes (ideas) contribute more to the development of new ideas than frogs with poor ideas. During thi evolution step, the frogs may change their memes using the information from the memeplex best frog Xb or the global best frog Xg of the entire population [13]. Accordingly, the position of the frog with the worst fitness is adjusted using Eqs. 1–2. Change in Frog position: Di = rand () . (Xb − Xw )

(1)

Xnew = Xw + Di ; (Dmax >= Di >= −Dmax )

(2)

New position:

Where rand() is a random number between 0 and 1; and Dmax is the maximum allowed change in a frog’s position. If this process yields a better frog (solution), it replaces the worst frog. Elsewise, the calculations in Eqs. (1) and (2) are repeated with respect to the global best frog (that is Xg replaces Xb). If no improvement becomes possible in this latter case, then a new solution (frog) with any arbitrary fitness is randomly generated to replace the worst frog [14]. The calculations then continue for a specific number of evolutionary iterations within each memeplex. After a number of memetic evolutionary steps, ideas are pass among memeplexes in a shuffling process (global search). The local search and the shuffling processes continue until convergence criteria are satisfied. The algorithm has been tested on several combinatorial problems and found to be efficient in finding global solutions [14]. The core parameters of SFLA are: population size, P, number of memeplexes, m, and number of evolutionary iterations in each memeplex, q [3].

2 Related Works Considering the application of SFLA, Elbeltagi, Hegazy and Grierson [14] compared the searching mechanism of the Genetic Algorithm (GA) with that of the SFLA and the experimental results of the comparison show that the SFLA have better performance than the GA in solving some problems of continuous functions. Their work also proposed an improved SFLA, introduced a new parameter called search-acceleration factor (C) to the original formulation of the SFLA, analyzed the positive role of the new parameter and solved discrete and continuous optimization problems. Nejad, Jahani and Sarlak [15] applied SFLA to Economic Load Dispatch (ELD) problem in power system. Their objective was to find the optimal combination of power generations that minimizes the total generation cost while satisfying an equality constraint and inequality constraints. Two

296

A. O. Ameen et al.

representative systems (IEEE 30 bus and 57 bus) were used to test their proposed SFLA algorithm in comparison with the GA based method for the solution of the ELD problem. The result proved that the SFLA technique was faster than the GA technique. Also, Liping, Weiwei, Yefeng and Yixian [16] introduced the SFLA to solve an uncapacitated Single Level Lot Sizing (SLLS) problem and gained ideal results. Gerasimou et al. [17] investigated the application of a Particle Swarm Optimization (PSO) algorithm to software project scheduling and effective team staffing. The study aims to create optimal project schedules by specifying the best sequence for executing a project’s tasks to minimize the total project duration and seeks to form skillful and productive working teams with the best utilization of developer skills. A combination of Constriction-PSO and Binary-PSO variations were employed to solve the problem. Results from empirical experiments showed that PSO was able to generate feasible solutions with feasibility rate of approximately 100% and hit rate of virtually 100% in all of considered problems. However, as the complexity and size of the problems increase a progressive decrease in these percentages is observed reaching as low as 30%. This shows that the employed algorithm still encounters difficulties in producing optimal solution as project complexity increases. Chen and Zhang [18] developed an approach based on an event-based scheduler (EBS) and an ant colony optimization (ACO) algorithm for optimal project scheduling and staffing. The model employed the event-based scheduler to simplify the restricted flexibility of human resource allocation. The project plan was model as task list and employee allocation matrix, then Ant Colony Optimization (ACO) algorithm was applied to solve the problem. Experimental results showed that the representation scheme with the EBS is effective, and the proposed algorithm manages to yield better plans with lower costs and more stable workload assignments compared with other existing approaches such as the Tabu Search (TS) algorithm for the multiskill scheduling Problem, the knowledge-based GA (KGA) and the time-line-based GA. The study however considered not the employee experience in the formulation. Stylianou and Andreou [7] proposed a procedure for software project managers to support their project scheduling and team staffing activities by adopting a genetic algorithm approach as an optimisation technique in order to construct a project’s optimal schedule and to assign the most experienced employees to tasks. Experimental results obtained revealed that the genetic algorithm is capable of finding optimal solutions for projects of varying sizes when using either one of the objective functions. However, when the objective functions were combined, the genetic algorithm presents difficulties in reaching optimal solutions especially when having preference to assign the most experienced employees over the project’s duration. This study presents SFLA as a memetic meta-hueristic algorithm to tackle this shortcoming. Recent works have focused on combining task scheduling and team allocation/ resource assigning based on multiple skills (as also adopted in this study) using different optimization approaches. Lin, Zhu, and Gao [24] proposed a genetic programming hyper-heuristic algorithm for minimizing makespan in multi-skill resource constrained project scheduling problem (MS–RCPSP). Comparisons with existing algorithms such

Application of Shuffled Frog-Leaping Algorithm

297

as HACO, GRASP and DEGR showed that the proposed algorithm performed considerably better with regards to solution quality and convergence rate. The same multiskill formulation was also employed by Li et al. [25] with focus on skill evolution and cooperation effectiveness in project scheduling. Van Den Eeckhout, Maenhout and Vanhoucke [26] applied a heuristic procedures based on iterated local search to an integrated personnel staffing problem and the project scheduling problem formulation such that the demand for staff and the scheduling of the resources is determined simultaneously as proposed in this study. However, their objective is to determine the personnel budget that minimizes project cost rather than combining minimized completion time and cost objectives. Recently, an optimization procedure for large scale resource constrained multi-objective project scheduling problem based on cooperative coevolution was proposed by Shen, Guo and Li [27]. Duration and cost are considered together as objectives with employees’ satisfaction. Experimental results on 15 randomly generated large-scale instances with up to 2048 decision variables indicated the high scalability of the proposed approach with regards to convergence ability.

3 Methodology To model the problem, Design Structure Matrix (DSM) which enforces the dependencies among tasks was used. It is represented as a jagged array of two-dimension where row indices represent WP ids. Using an hypothetical software project consisting of seven WPs, an example of a modeled DSM is shown in Fig. 1.The DSM indicates that WP1 does not depend on any task before it can actually start, WP1 must finish before WP2 can start, WP1 and WP2 must finish before WP3 and WP4 can start, WP1, WP2, WP3 and WP4 must finish before WP5 can start, WP4 and WP5 must finish before WP6 can finish and WP5 must finish before WP7 can start. For a software project scheduling problem, the number of WPs is usually less than or equal to 2n – 1, with n representing the number of employees required to complete the project.

WP 1 WP 2 WP 3 WP 4 WP 5 WP 6 WP 7

1 1 1 1 4 5

2 2 2 5

3

4

Fig. 1. DSM model representation of dependencies constraints

The staff allocation is modeled using binary representation of an integer number × having a value in the interval 1 to 2n – 1, where n equals the employee involved in the project. The value of each bit in the binary equivalence denotes an employee involvement in the current task. A value of 1 means the corresponding employee is

298

A. O. Ameen et al.

allocated for the given WP and 0 means the employee is not allocated. Starting from the left, the first bit denotes employee1 s involvement in the task, the next bit represents employee2 s involvement and so on. Assuming that four employees are available for the project represent by the DSM in Fig. 1, the employee assignment of any of the WPs is the binary equivalence of a number between 1 and 15. An example of employee assignment under this representation is presented in Table 1. Associated with each employee is skill set represented as a linear array of skill types. Also, for each WP, the required competence(s) is defined which is represented as an n-array of skill types. The total required competence of a WP is the sum of all the inherent skill set possessed by the team of employees assigned to the WP. Table 1. Employee assignment representation Work package Employee assignment Binary equivalence Remarks 1

2

0010

Task assigned to only employee 3

2

11

1011

Task assigned to employees 1, 3 and 4

3

15

1111

Task assigned to employees 1, 2, 3 and 4

4

5

0101

Task assigned to employees 2 and 4

5

4

0100

Task assigned to only employee2

6

13

1101

Task assigned to employees 1, 2 and 4

7

7

0111

Task assigned to employees 2, 3 and 4

3.1 Frog Representation A frog represents a feasible solution to project scheduling and staffing problem. It is encoded as an n × 2 array where each row consists of a WP id and an integer number representing employee assignment. The row index indicates the position of the WP in the WPs ordering. For example, row index 0 indicates position (POS) 1, and the associated WP start first before any other WP. A typical frog schema is shown in Fig. 2. 3.2 SFLA Design Shuffled Frog Leaping Algorithm (SFLA) works generally as follows: At first, a virtual or random population of frogs is created (where p is the population size). Subsequently, the fitness of the individual frogs is evaluated. Afterwards, the frogs are sorted in descending order of their fitness (that is the fittest to the worst). Thereafter, the frogs are partitioned into m memeplexes. Then, a local search is performed within each memeplex. During

Application of Shuffled Frog-Leaping Algorithm

1 2 4 3 5 7 6

299

1 5 6 7 2 3 7

Fig. 2. A frog representation

each intra-memeplex local search, the best frog and the worst frog are identified as Xb and Xw respectively and the global best frog is identified as Xg. Then, a process is applied to improve only the worst frog, excluding other frogs. Consequently, in this approach the position of the worst frog (Xw) is adjusted using Eqs. 3–7. chunkLength = 0.5 × frog_size

(3)

Start = rand() × (frog_size − chunkLength

(4)

I = chunkLength + start; start = 0

(8)

Where: V = number of dependency violations, M = number of skill mismatches and n = number of WPs for frog A.

4 Results and Discussion The modified SFLA was tested on a randomly generated software project scheduling and staffing problem consisting of seven (7) WPs, twelve (12) dependencies, five (5) employees. The population size is varied as 100, 150, 200 and 300. The number of memeplexes is also varied as 5, 10 for each cases of the population size. This variation of population size and memeplex size is necessary because there are no generally acceptable criteria for choosing population size and number of memeplexes for a given problem. These parameters together with the number of evolutionary iterations greatly influence the performance of the algorithm. The number of evolutionary iterations per memeplex is set to 2N, where N is the number of frogs in each memeplex as proposed [4]. Owing to the fact that SFLA works on a virtual or randomly generated population and tries to improve the frogs based on the convergence criteria set, its result and how well the improvement of frogs is done is always time varied. Hence, there is need to run the algorithm a number of instances and then average the results to have a better evaluation of the performance of the algorithm on a given problem and how well the feasible solutions to the problem (frog) are improved before finally selecting the best solution. Table 2 presents the results of experiments carried out on the random problem using SFLA with varied population size and number of memeplexes, and pure random search.

Application of Shuffled Frog-Leaping Algorithm

301

A total of twenty (20) independent runs as proposed in [28] were performed on each case of the variation and the results were averaged. The same experiments were also carried out with pure random search and for a comparison with the proposed approach. All algorithms are implemented in Java. Table 2. Experimental results of SFLA and random search Population size

Number of memeplex (es)

100 150 200 300

Average fitness (SFLA)

Average fitness (random) 4.12

5

0.27

10

0.49

5

0.23

10

0.32

5

0.16

10

0.28

5

0.15

10

0.22

5.11 4.79 5.14

It can be deduced that the algorithm worked better on a larger population size and for the same population size when the number of memeplex was varied, the lower the number of memeplex, the better the improvement of the whole population, the average fitness of the individual frogs in the population and the selected best solution. The SFLA approach was also compared with a pure random approach of generating feasible solutions (frog) based on a set threshold (maximum of 3 dependency violation can be made) and the proposed SFLA proved better when compared with the results of the random search approach. Figure 4 presents this comparison in a line graph.

6

Average Fitness

5 4 3

Ramdom Search

2

SFLA

1 0 100

150 200 Populaon size

300

Fig. 4. Average fitness comparison of SFLA and random search

302

A. O. Ameen et al.

From Fig. 4 it is observed that SFLA significantly outperformed random search in all population sizes with difference of up to 4.92 average fitness. This result revealed the effectiveness of SFLA in handling project scheduling and staffing problem under our formulation.

5 Conclusion In this work, a good data structure that enforces dependency constraints among Work Packages (WPs) was successfully adopted. The study was able to find a mathematical representation with easy implementation for staff allocation. This enables the adoption of a good data structure in representing a frog (solution) that will cater for both work package ordering and staff allocation. The study adopts the power of SFLA to find the near-optimal solution for randomly generated Software Project Scheduling Problem (SPSP) and a comparison was made with a purely random approach. The SFLA approach in project planning provides a new, effective and efficient perspective to recent software projects scheduling. The result analysis of the study shows that it performs reasonably well in project scheduling. In the future, we plan to include more objectives, carry out empirical studies with real world project scheduling standard problem instances and compare results with existing studies.

References 1. Kang, K., Hahn, J.: Learning and forgetting curves in software development: does type of knowledge matter? In: ICIS 2009 Proceedings, p. 194 (2009) 2. Mojeed, H.A., Bajeh, A.O., Balogun, A.O., Adeleke, H.O.: Memetic approach for multiobjective overtime planning in software engineering projects. J. Eng. Sci. Technol. 14(6), 3213–3233 (2019) 3. Patil, N., Sawanti, K., Warade, P., Shinde, Y.: Survey paper for software project scheduling and staffing problem. Int. J. Adv. Res. Comput. Commun. Eng. 7, 5675–5677 (2014) 4. Oladele, R.O., Mojeed, H.A.: A shuffled frog-leaping algorithm for optimal software project planning! Afr. J. Comput. ICT 7(1), 147–152 (2014) 5. Amiri, M., Barbin, J.P.: New approach for solving software project scheduling problem using differential evolution algorithm! Int. J. Found. Comput. Sci. Technol. 5(1), 1–5 (2015) 6. Eshraghi, A.: A new approach for solving resource constrained project scheduling problems using differential evolution algorithm. Int. J. Ind. Eng. Comput. 7(2), 205–216 (2016) 7. Stylianou, C.S., Andreou, A.S.: Intelligent software project scheduling and team staffing with genetic algorithm. In: IFIP Advances in Information and Communication Technology (IFIPAICT), vol. 364. Springer, Heidelberg (2011) 8. Shen, X., Minku, L.L., Bahsoon, R., Yao, X.: Dynamic software project scheduling through a proactive-rescheduling method. IEE Trans. Softw. Eng. 42(7), 658–686 (2016) 9. Vitekar, K.N., Dhanawe, S.A., Hanchate, D.B.: Review of solving software project scheduling problem with ant colony optimization. Int. J. Adv. Res. Electr. Electron. Instrum. Eng. 2(4), 1177–1186 (2013) 10. Karova, M., Petkova, J., Smarkov, V.: A genetic algorithm for project planning problem. In: Proceedings International Scientific Conference Computer Science 2008, pp. 647–651 (2008) 11. Eusuff, M.M., Lansey, K.E.: Optimization of water distribution network design using the shuffled frog leaping algorithm. J. Water Resour. Plan. Manag. 129(3), 210–225 (2003)

Application of Shuffled Frog-Leaping Algorithm

303

12. Mai, G., Li, Y.: An improved shuffled frog leaping algorithm and its application. In: Proceedings of International Conference on Advances in Mechanical Engineering and Industrial Informatics, China (2015) 13. Eusuff, M., Lansey, K., Pasha, F.: Shuffled frog leaping algorithm: a memetic meta-heuristic for discrete optimization. Eng. Optim. 38(2), 129–154 (2006) 14. Elbeltagi, E., Hegazy, T., Grierson, D.: A modified shuffled frog-leaping optimization algorithm: applications to project management. Struct. Infrastruct. Eng. 3(1), 53–60 (2007) 15. Nejad, H.C., Jahani, R., Sarlak, G.: Applying shuffled frog-leaping algorithm for economic load dispatch of power system. Am. J. Sci. Res. 20, 82–89 (2011) 16. Liping, Z., Weiwei, W., Yefeng, X., Yixian, C.: Application of shuffled frog leaping algorithm to uncapacitated SLLS problem. AASRI Procedia 1, 226–231 (2012) 17. Gerasimou, S., Stylianou, C., Andreou, A.S.: An investigation of optimal project scheduling and team staffing in software development using particle swarm optimization. ICEIS 2, 168– 171 (2012) 18. Chen, W.N., Zhang, J.: Ant colony optimization for software project scheduling and staffing with an event-based scheduler. IEEE Trans. Softw. Eng. 39(1), 1–17 (2013) 19. Weisstein, E.W.: NP-Hard Problem (2017). https://mathworld.wolfram.com/NP-HardPr oblem.html 20. Wysocki, R.K.: Effective Project Management: Traditional, Agile, Extreme, 5th edn., pp. 167– 171. Wiley Publishing, Indianapolis (2009) 21. Marler, R.T., Arora, J.S.: Survey of multi-objective optimization methods for engineering. Struct. Multidiscip. Optim. 26, 369–395 (2004) 22. Krasnogor, N., Aragon, A., Pacheco, J.: Metaheuristic procedures for training neural networks. Operations Research/Computer Science Interfaces Series, vol. 36, pp. 225–248 (2006) 23. Rezende, A.V., Silva, L., Britto, A., Amaral, R.: Software project scheduling problem in the context of search-based software engineering: a systematic review. J. Syst. Softw. 155, 43–56 (2019) 24. Lin, J., Zhu, L., Gao, K.: A genetic programming hyper-heuristic approach for the multi-skill resource constrained project scheduling problem. Expert Syst. Appl. 140, 112915 (2020) 25. Li, Q., Sun, Q., Tao, S., Gao, X.: Multi-skill project scheduling with skill evolution and cooperation effectiveness. Eng. Constr. Archit. Manag. 27, 2023–2045 (2019) 26. Van Den Eeckhout, M., Maenhout, B., Vanhoucke, M.: A heuristic procedure to solve the project staffing problem with discrete time/resource trade-offs and personnel scheduling constraints. Comput. Oper. Res. 101, 144–161 (2019) 27. Shen, X., Guo, Y., Li, A.: Cooperative coevolution with an improved resource allocation for large-scale multi-objective software project scheduling. Appl. Soft Comput. 88, 106059 (2020) 28. Harman, M., Mansouri, S.A., Zhang, Y.: Search based software engineering. A comprehensive analysis and review of trends techniques and applications. Technical report TR-09-03. Department of computer science, King’s College, London (2009)

A Long Short Term Memory and a Discrete Wavelet Transform to Predict the Stock Price Mu’tasem Jarrah1(B) and Naomie Salim2 1 King Abdulaziz University, Jeddah, Kingdom of Saudi Arabia

[email protected] 2 Universiti Teknologi Malaysia - UTM, Johor Bahru, Malaysia

[email protected]

Abstract. Financial Analysis is a challenging task in the present-day world, where investment value and quality are paramount. This research work introduces the use of a prediction technique that uses a combination of Discrete Wavelet Transform (DWT) and Long Short-Term Memory (LSTM) to predict stock prices in the Saudi stock market for the subsequent seven days. A time series model is used where comprises the historical closing values of several stocks listed on the Saudi stock exchange. This model is called the Discrete Long Short-Term Memory (DLSTM) which comprises memory elements that preserve data for extended periods. The function determined the historical closing price of the stock market and then employed Autoregressive Integrated Moving Average (ARIMA) for analysis. The DLSTM-based experimental model had a prediction accuracy of 97.54%, while that of ARIMA was 97.29%. The results indicate that DLSTM is an effective tool for predicting the prices in the stock market. The results highlight the importance of deep learning and the concurrent use of several information sources to predict stock price levels Keywords: Long Short Term Memory · Deep learning · Prediction · Stock market

1 Introduction Stock price forecasts have been an area of significant interest since the past several decades [1]; however, such predictions are challenging because of the dynamic and complex nature of the environment [2]. Forecasting the trend and the prices in the stock market are considered crucial in the financial and investment domains. Several researchers have studied and suggested techniques to forecast market price to reap profits during trading by employing several methods such as statistical analysis and technical analysis. Stock trends are difficult to predict since there are a lot of uncertain factors and noise that affect prices. Numerous aspects may influence the market price on any specific day; changes to the national economy, sentiment, value of the product, political aspects, and weather are some such aspects [3]. Researchers have assessed and worked on stock price © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 304–313, 2021. https://doi.org/10.1007/978-3-030-70713-2_29

A Long Short Term Memory and a Discrete Wavelet Transform

305

trends to understand the factors that have the most significant effect on prices. This study uses LSTM and DWT to predict the prices of the stocks for the subsequent seven days. The sentiment of the customers is an aspect that affects the stock markets. Additionally, financial developments and the sentiment of the stock buyers who understand that their perspective concerning products or services offered by a firm significantly affect price volatility, which will be addressed in future studies. The present research compares traditional methods like ARIMA to the proposed DLSTM model. Numerous options have been evaluated for modelling, and the selected model has been assessed considering many possibilities while considering different model configurations. The present study concerns four critical aspects for the proposed models; these aspects are Discrete Wavelet Transform (DWT) used in combination with the Long Short-Term Memory (LSTM) deep-learning framework. The DWT technique facilitates noise elimination specific to the financial time series data using unsupervised techniques. The paper is structured as specified: Sect. 2 details the works specific to stock market forecast; Sect. 3 discusses the proposed model and its characteristics. Section 4 lists the results and comprises a discussion of the results, while Sect. 5 concludes the research and contains recommendations for additional research.

2 Literature Survey The model proposed in this research is an improved version of the model used in [4]. The present study builds on the existing model to make the regression analysis of the stock market more accurate. Data collection comprises the first step in the proposed model. In contrast, the second step consists of data cleansing and transformation, which are necessary steps to have the data required for analysis. Labelling is the critical phase which comprises the determination of data polarity of individual opinion as being positive, negative, or neutral. The fourth step comprises classification, whereby stock patterns are identified by employing the hybrid Naïve Bayes Classifiers (NBCs). The last step determines model performance. The Hybrid Naïve Bayes Classifiers (HNBCs) is the machine learning technique suggested in this study to perform the classification of the stock market sentiment. The results are important for firms, investors, and academicians since the results may be used to plan further action according to the sentiment of the individuals associated with the stock market. The results achieved using the proposed technique have been significant, whereby the accuracy is 90.38%. Batra & Daudpota (2018) suggested a novel machine learning technique concerning sentiment analysis to determine perspective (neutral, positive, or negative) used a piece of text addressing a person, product, enterprise, or another entity. Sentiment analysis helps determine the mood of the individuals whose perspective potentially affects stock price; hence, this technique can help predict real-world stock movement [5].

306

M. Jarrah and N. Salim

In a related study by Fischer and Krauss (2018), Long short-term memory (LSTM) networks were used as an advanced technique for learning sequences. Such networks are used predominantly with financial time series data; however, they are intrinsically applicable in this domain. We implement LSTM-based networks for forecasting the outof-sample directional trends of the stocks listed on the S&P 500. Additionally, characteristics determining profitability are highlighted, thereby providing information regarding the intricate working of artificial neural networks. One pattern associated with trading stocks is that they are associated with a high degree of volatility but a reversal in the short-term [6]. Minh, Sadeghi-Niaraki, Huy, Min, & Moon, (2018) suggested a new technique to forecast stock-price direction using sentiment dictionary and finance-related developments. Financial news is an established important aspect that leads to a change in stock prices. Nevertheless, previously conducted research has emphasised on the analysis of the superficial aspects, while disregarding the structural association between the words making a sentence. Numerous studies concerning sentiment analysis have attempted to determine the correlation between news events and the reaction of the investors. However, the lingual dataset was typically used to build the sentiment dataset. The lingual dataset is not related to the financial domain and led to inadequate performance [7]. Chou & Nguyen (2018) suggest an intelligent prediction method based on time series data that employs the sliding-window metaheuristic optimisation for forecasting stock prices one step in advance. The proposed system comprises a standalone application based on a graphical user interface. The hybrid system designed as part of the research demonstrated excellent forecast performance, which leads to higher profits concerning investment. The suggested framework comprises a powerful prediction method for severely non-linear time series data where traditional models may be unable to detect the patterns accurately [8].

3 Prediction Using DWT and LSTM 3.1 Discrete Wavelet Transform (DWT) The DWT technique has powerful feature extraction capability; therefore, it is employed in several domains like signal processing and financial time series. The primary aspect of the wavelet transform is that it allows the analysis of the frequency elements of the financial time series data concurrently as opposed to Fourier transform. Therefore, the wavelet transform technique facilitates a better understanding of financial time series comprising significant irregularities. This study uses the Haar function as the wavelet basis because it helps with decomposing the financial time series into the constituent time and frequency domains and also leads to significantly reduced processing time [9]. In the context of the Haar functionbased wavelet transform, O(n) denotes the time complexity of the process; here, n denotes the time series size. The expression for the continuous wavelet transform (CWT) is specified below: √ (1) ∅_(a, τ )(t) = 1/ a ∅((t − τ )/a)

A Long Short Term Memory and a Discrete Wavelet Transform

307

Where a and τ denote the scale and translation factors, respectively, and φ(t) represents the basis wavelet function [10]. 3.2 Long Short-Term Memory A long short-term memory (LSTM) element [11] or network is a sophisticated variant of the basic recurrent neural network. It may potentially be used as a constituent building element for improved series analysis using the recurrent network. An LSTM block is fundamentally a recurrent network since it comprises recurring connections like those found in traditional recurrent networks. The LSTM is formulated to resemble a recurrent neural network so that it can be used for processing long-term associations with better accuracy compared to traditional Recurrent Neural Networks. According to [12], the concurrent use of LSTM and RNN has shown better performance compared to simple RNN and Deep Neural Networks (DNN) in the context of speech recognition or stock price movement. Conventional DNNs are restricted in the sense that the modelling can be based only on a fixed-size sliding window where the network does not have a dependency on the time steps processed previously; hence, it is not the first choice for appropriate modelling of stock-specific data (refer to Fig. 1). The study uses data collected from the Saudi Stock Market (TADAWUL) to train and test the proposed model. After the data is gathered, it is normalised, and the training and testing sets are created. The training set is employed to train the formulated DLSTM framework so that it can predict the stock data for the subsequent seven days. After training, the model is tested using the data from several companies, and the outputs are compared to the actual stock prices using graphical plots, which are depicted in Fig. 2.

Fig. 1. Long short term memory cell

4 Experimental Setup This section comprises three sub-sections: the first specifies the characteristics of the dataset employed for experimentation, the second sub-section discusses the prediction method and accuracy measurement, while the last section comprises the results of the experiment.

308

M. Jarrah and N. Salim Table 1. Shows the formula for each component at time step t.

Component

Formula     Input gate it = σ Wi . ht−1 , xt + bi     Forget gate ft = σ Wf . ht−1 , xt + bf     Cell candidate gt = σ Wg . ht−1 , xt + bg     Output gate ot = σ Wo . ht−1 , xt + bo Cell state Hidden state

Ct = ft ∗ Ct−1 + it ∗ gt ht = ot ∗ tanh(Ct )

Purpose Control level of cell state update Control level of cell state reset (forget) Add information to cell state Control level of cell state added to hidden state Transfer data from one step to the next Used for predictions (Output)

- Where tanh denotes the state activation function [13].

DWT Tadawul Fi-

Data Processing

nance Date Split Data

Testing

Build Model Validation

Prediction (Next 7 Days)

Fig. 2. Architecture of the proposed model (DLSTM)

4.1 Data Description The present study uses the historical data concerning the stocks listed on the Saudi stock market (Tadawul). The entity was approved to operate in Saudi Arabia as the Securities Exchange (also referred to as the Exchange), and it records the daily open/low/high/close prices and the volume relating to all the stocks traded on the market. Tadawul comprises 1300 records for each of the 146 stocks that were listed between 2011/01/01 and 2016/03/31. Of the 190,000 series gathered from the database, 130,000 were employed for training the model, while the remaining 60,000 were used for validating the model. In the context of the suggested model specific to this research, academicians have preferred to use as input the closing price for six different companies from different industries. It should be noted that all companies had a noteworthy deviation in the outcomes. The data specific to these companies had a marginal error rate and were, therefore, considered as errorless. 4.2 Prediction Procedure Numerous tests were used during the research to regulate the parameters to fine-tune the results. The number of training periods is the first aspect of the LSTM model that needs to be tuned. The next variable is the batch size, which determines the update frequency of the network weights. The number of neurons is the third aspect, which modifies the learning ability of the system. Furthermore, in this study, the Adam optimised method

A Long Short Term Memory and a Discrete Wavelet Transform

309

is used for the LSTM because it builds upon the stochastic gradient descent algorithm, which is widely accepted in the deep learning domain. A higher number of neurons is typically associated with the system having an increased capacity of learning the problem structure, though the training duration gets extended. Higher learning capability also leads to the problem of overfitting of training data. The test parameters and the averages indicating prediction accuracy (MAE, MSE, RMSE) are listed in Table 1 for all the cases. Table 2. Experiments details Experiment

Epochs

Batch size

Neurons

MAE

MSE

RMSE

1

100

8

4

0.46

0.26

0.507

2

150

4

4

0.28

0.09

0.301

3

150

4

5

2.13

5.79

2.406

4

200

4

5

0.24

0.19

0.424

5

200

4

6

2.29

5.86

2.421

Using the table, it may be observed that test 4 produces the best results. Table 2 lists the data specific to the chosen sample that comprises four companies from the Saudi market and the S&P 500. Additionally, historical stock data was obtained from the website https://macrotrends.dpdcart.com/cart/deliver?purchase_id=12487241&salt=4526fa b69067075ba5560b21f1850513b192ef77. The data was processed using the ARIMA and DLSTM processes. The results of the tests performed on the Saudi companies and S&P 500 stocks are specified in Table 3 and Table 4 and depicted using Fig. 3, 4, 5 and Fig. 6 and Fig. 7 on respectively. The present study used a sample comprising of six randomly selected companies. Concerning the prediction process, the stock closing price was considered a significant parameter because it is associated with the opening price for the subsequent day. During the prediction process, the dataset pertaining to the Saudi stock market was split into two for every firm on the dataset. These two datasets are called the training and testing sets. Of the total number of entries, 1306 entries comprise the training set, while the remaining would be employed for model testing. Additionally, the S&P 500 dataset is also split into the training and testing sets. In this case, the first 1313 records are used for training and the others for testing. Models will be formulated using the training dataset, while the testing dataset would be used to validate the models by predicting the outputs. The time step for every dataset will move by one. A model is employed to forecast the time step, and the value expected for the testing set will be fed to the model to help predict the next step. Such a setting is similar to the real-world scenario where the updated stock market view can be accessed every day and be used for forecasting the outcomes for the subsequent day. Lastly, the predictions made using the testing set are collected, and the error value is calculated to ascertain the predictive power of the model.

310

M. Jarrah and N. Salim

The Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE) are the metrics used to tune the model, where higher errors are corrected to facilitate better results that are in line with the real data.

5 Results The researchers used Python 3.6 running on Windows 10 operating system to validate the model. Several Python libraries were used: firstNumpy considers a homogenous multidimensional array as the primary object. It may be understood as the element table (typically containing numeric values of the same type). The Pandas package provided with Python provides flexible, quick, and expressive mechanisms that are formulated to provide easy-to-understand “labelled” and “relational” data. Sklearnit is a library that provides for efficient and straightforward data assessment. Furthermore, the Keras library may be executed using Theano, TensorFlow, or the MS Cognitive Toolkit. This library provides for swift testing using deep neural networks and focuses on extensibility, modularity, and user-friendliness. Finally, the Matplotlibit library is the Python plotting library which works with its mathematical extension called NumPy. Table 3, along with Fig. 3, 4, 5 and 6 depict a sample stock on which model testing was performed using a seven-day timeframe. Table 3 contains details about the actual and predicted stock prices along with information concerning prediction accuracy and error indicators, namely, MAE, MSE, and RMSE. Additionally, Fig. 7 depicts the accuracy summary obtained using the application of the suggested DLSTM framework. Table 3. Predictions result for the next 7 days Company Name

Day 1 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7 MAE

MSE

RMSE

Actual Saudi Arabian Mining Co. DLSTM

30.16 29.86 29.50 29.07 28.67 28.54 28.52 30.16 30.06 29.66 29.28 28.85 28.45 28.45

0.131 0.022

0.150

ARIMA

30.24 30.09 29.65 29.32 28.83 28.45 28.52

0.137 0.025

0.159

Actual

49.15 49.14 49.11 49.07 49.02 49.00 48.98

Yanbu Cement Co.

Sabic

Saudi Indian

DLSTM

49.12 49.22 49.14 49.08 49.05 49.00 48.99

0.017 0.000

0.021

ARIMA

49.14 49.16 49.13 49.11 49.06 49.00 48.99

0.026 0.001

0.035

Actual

77.27 76.86 76.29 75.60 74.91 74.73 74.75

DLSTM

77.29 77.34 76.56 75.90 75.12 74.44 74.58

0.205 0.059

0.243

ARIMA

77.33 77.24 76.44 75.99 75.09 74.48 74.79

0.249 0.079

0.281

Actual

11.01 11.76 12.70 13.76 14.81 15.16 15.21

DLSTM

10.93 11.06 12.12 13.26 14.45 15.52 15.54

0.413 0.205

0.453

ARIMA

10.91 11.04 11.94 12.94 14.04 15.09 15.31

0.475 0.338

0.581

A Long Short Term Memory and a Discrete Wavelet Transform

Fig. 3. Prediction for Saudi Mining Co.

311

Fig. 4. Prediction for Yanbu Cement Co.

Fig. 5. Prediction for Sabic.

Fig. 6. Prediction for Saudi Indian.

In the next step, the final test consisting of the ARIMA and DLSTM techniques is performed on the S&P 500 dataset. Fig. 7 and Table 4 provide information specific to the experimental outcomes for the 7-day window. Table 4. Predictions result for the next 7 days (S&P index) Index Name S&P Actual 500 DLSTM Index ARIMA

Day 1 Day 2

Day 3

Day 4

Day 5

Day 6

Day 7

MAE MSE RMSE

2049.80 2036.71 2035.94 2037.05 2055.01 2063.95 2059.74 2050.96 2038.52 2037.74 2038.86 2055.85 2064.23 2060.29

1.18

1.75

2051.65 2048.93 2036.36 2035.95 2038.19 2055.79 2065.44 6.61 76.58

1.32 8.75

312

M. Jarrah and N. Salim

Fig. 7. Summary for prediction next 7 days (S&P index)

6 Conclusions and Future Work Several researchers have worked on the subject of stock market price prediction and have created two prediction mechanisms. The first mechanism forecasts the direction of the movement in the stock market and individual stock prices. At the same time, the other mechanism helps to predict future values. The predictions generated using the framework help provide financial assistance to users who can make better-informed decisions concerning stock market investments. This study proposes a new hybrid framework that is based on the LSTM and DWT combination, which relies on a technical dataset for stock price prediction. The proposed framework is capable of integrating the data collected from the stock market using the DWT and LSTM and then perform a simple optimisation process. Additionally, the integrated technique may be employed for creating better techniques that can address the risks better and provide better assistance to investors. It is suggested that researchers pursuing such studies considered several additional aspects to enhance prediction accuracy. The Umrah, Hajj period, and the Ramadan celebration month are celebrated events. Such factors could be studies for any potential effect on the prediction accuracy of stock price movements on the Saudi stock market.

Abbreviations LSTM DWT RNN DLSTM

Long Short Term Memory Discrete Wavelet Transform Recurrent Neural Networks Discrete Long Short Term Memory

References 1. Li, R., DianZheng F., Zeyu, Z.: An analysis of the correlation between internet public opinion and stock market. Paper presented at the 2017 4th International Conference on Information Science and Control Engineering (ICISCE) (2017)

A Long Short Term Memory and a Discrete Wavelet Transform

313

2. Nabipour, M., Nayyeri, P., Jabani, H., Mosavi, A., Salwana, E.: Deep learning for stock market prediction. Entropy 22(8), 840 (2020) 3. Jarrah, M., Salim, N.: A recurrent neural network and a discrete wavelet transform to predict the Saudi stock price trends. Int. J. Adv. Comput. Sci. Appl. 10(4), 155–162 (2019) 4. Batra, R., Daudpota, S.M.: Integrating stocktwits with sentiment analysis for better prediction of stock price movement. Paper presented at the 2018 International Conference on Computing, Mathematics and Engineering Technologies (iCoMET) (2018) 5. Bruce, L.M., Koger, C.H., Li, J.: Dimensionality reduction of hyperspectral data using discrete wavelet transform feature extraction. IEEE Trans. Geosci. Remote Sens. 40(10), 2331–2338 (2002) 6. Chou, J.-S., Nguyen, T.-K.: Forward forecast of stock price using sliding-window metaheuristic-optimized machine-learning regression. IEEE Trans. Industr. Inf. 14(7), 3132– 3142 (2018) 7. Fischer, T., Krauss, C.: Deep learning with long short-term memory networks for financial market predictions. Eur. J. Oper. Res. 270(2), 654–669 (2018) 8. Li, R., Fu, D., Zheng, Z.: An analysis of the correlation between internet public opinion and stock market. Paper presented at the 2017 4th International Conference on Information Science and Control Engineering (ICISCE) (2017) 9. Minh, D.L., Sadeghi-Niaraki, A., Huy, H.D., Min, K., Moon, H.: Deep learning approach for short-term stock trends prediction based on two-stream gated recurrent unit network. IEEE Access 6, 55392–55404 (2018) 10. Mithani, F., Machchhar, S., Jasdanwala, F.: A modified BPN approach for stock market prediction. Paper presented at the 2016 IEEE International Conference on Computational Intelligence and Computing Research (ICCIC) (2016) 11. Zhu, L.-F., Ke, L.-L., Zhu, X.-Q., Xiang, Y., Wang, Y.-S.: Crack identification of functionally graded beams using continuous wavelet transform. Compos. Struct. 210, 473–485 (2019)

Effective Web Service Classification Using a Hybrid of Ontology Generation and Machine Learning Algorithm Murtoza Monzur(B) , Radziah Mohamad, and Nor Azizah Saadon School of Computing, Faculty of Engineering, Universiti Teknologi Malaysia UTM, 81310 Skudai, Johor, Malaysia [email protected]

Abstract. Efficient and fast service discovery becomes an extremely challenging task due to the proliferation and availability of functionally-similar web services. Service classification or service grouping is a popular and widely applied technique to classify services into several groups according to similarity, in order to ease up and expedite the discovery process. Existing research on web service classification uses several techniques, approaches and frameworks for web service classification. This study focused on a hybrid service classification approach based on a combination of ontology generation and machine learning algorithm, in order to gain more speed and accuracy during the classification process. Ontology generation is applied to capture the similarity between complicated words. Then, two machine learning classification algorithms, namely, Support Vector Machines (SVMs) and Naive Bayes (NB), were applied for classifying services according to their functionality. The experimental results showed significant improvement in terms of accuracy, precision and recall. The hybrid approach of ontology generation and NB algorithm achieved an accuracy of 94.50%, a precision of 93.00% and a recall of 95.00%. Therefore, a hybrid approach of ontology generation and NB has the potential to pave the way for efficient and accurate service classification and discovery. Keywords: Web service discovery · Web service description language (WSDL) · Service classification · Ontology · Machine learning · Support Vector Machines (SVMs) · Naive Bayes (NB)

1 Introduction With the expansion of service-oriented architectures, web services have turned into a distinguished technology for providing superior solutions for the interoperability of various types of systems. Web services are the compilation of related application functions and freely coupled software components that can be distributed and utilized on the web [1]. This enables various applications from various sources to communicate with one another, utilizing standard protocols in real-time with minimal human cooperation. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 314–323, 2021. https://doi.org/10.1007/978-3-030-70713-2_30

Effective Web Service Classification Using a Hybrid of Ontology Generation

315

The fundamental advantage of utilizing web services is property interoperability, which makes web services progressively famous compared to related technologies [2]. Web services have shown significant improvement in a distributed computing paradigm. They are currently considered the most suitable way to publish and describe business processes. The popularity of web service applications affects the number of web services on the web. The proliferation of functionally similar web services makes it difficult to identify which context needs to be considered, and how to classify during the discovery process to achieve overall user satisfaction [3]. Overall, web service discovery has become a very significant and challenging task. Web service classification to identify functionally-similar web services has become a major approach to the efficient discovery of suitable web services. Web service classification is a process of distributing web services into several classes according to their functions and contexts, such that the similarity between services within one class remains high, while dissimilarity remains low [4]. Domain experts typically execute this classification process manually. With the exponential growth of web services on the internet, however, arranging, classifying, and handling web services manually has become impractical, as this requires intense human effort. Moreover, due to the vast number of categories in web service registries, it is an error-prone task [5]. Classification accuracy can be improved by using an ontology that captures the similarity between conceptually complicated words. Moreover, combining machine learning algorithms together with the ontology, can be an effective approach to classify web services. Therefore, this work proposes a web service classification approach based on ontology generation and machine learning algorithms in order to capture the similarity and classify functionally similar services. The remainder of this paper is organized as follows. Section 2 provides an overview of previous works related to different web service classification techniques, approaches and frameworks. Section 3 describes the detailed description of the proposed approach. Section 4 explains the experimental results. Finally, Sect. 5 concludes this paper and discusses potential future work.

2 Related Work Work in the web service classification area has recently gained significant attention based on the popularity of web services and the possible advantages that can be achieved from automated web services classification. Existing classification approaches use several techniques, approaches and frameworks to compute the similarity between web services. These include Specificity Aware Ontology Generation [6], Search Space Reduction Approach by applying Modified Negative Selection (M-NSA) algorithm [4], Hybrid Term Similarity (HTS) method and Context-Aware Similarity (CAS) method that uses Support Vector Machines (SVMs) for similarity calculations [1], and K-means++ method by extracting the feature vector of the service function description from the WSDL file [7]. Most existing approaches consider the functional properties for classifying web services. Most of these works prefer a hybrid approach for classifying web services. The difficulties in RESTful web service discovery is brought about by the absence of WSDL-like documents to provide a standard definition or portrayal of the RESTful web service

316

M. Monzur et al.

[4]. The work by [4] focused on enhancing the accuracy of the classification prior to discovery, in order to achieve less computation time and more accurate solutions. Any ontology generation method is not considered by this work. The work by [7] considered users’ QoS records for user classification. For web service classification, the work used extracted features from the WSDL file. But the number of extracted features was relatively low, since the WSDL file contains a huge amount of information about a web service. Therefore, considering only a few features for classification can affect the quality of the classes. The work by [1, 8, 9] and [6] considered several ontology generation methods for capturing the similarity between web services by extracting several features from the WSDL file, including domain-related information that can be very useful for achieving accurate classification. But the problem is that all of those works used agglomerative algorithms for identifying the classes. There are several drawbacks of using an agglomerative algorithm [10]. • It can never undo any previous steps. For example, the algorithm used to separate or classify two points, and later if the work is not considered as a good one, the algorithm does not provide the option to undo that step. • The time complexity of the agglomerative algorithm can bring about long computation times. • The correct number of classes determination is very difficult by the dendrogram for a large number of datasets. A combination of SLS and SVMs classification algorithms shows significant improvement compared to the SVMs classification algorithm alone, in terms of accuracy [11]. The most popular machine learning algorithms for classification are Support Vector Machines (SVMs) and Naive Bayes (NB), as they provide an embedded feature selection method. Feature selection is automatically selecting a subset of the most appropriate features for a problem from an original feature set, to be included in a model. The embedded feature selection method primarily works with learning algorithms. During model creation, feature selection is performed without splitting the data into training and testing sets. The combination of an ontology generation method and machine learning algorithms such as SVMs or NB can be one of the approaches for classifying web services in a better manner, due to the faster computational speed and accuracy of the machine learning algorithm, as well as the ability of the ontology generation method to capture the similarity among complex terms. A combination of ontology generation and machine learning algorithm can achieve superior accuracy. The work by [12] describes an ontology as a set of representational primitives and specifications with existing interrelationships for a particular domain. A web ontology is used to describe complex items on the web, and can define the rich concepts and knowledge about information interpretation. It comprises hierarchical definitions of principal concepts in a domain, along with descriptions of properties for each concept [13]. Concepts in ontology construction can be modelled as classes or sub-classes, depending on the hierarchy. According to Table 1, most of the works consider a hybrid approach for classifying web services. The ontology generation method is also widely used for capturing the hidden semantic meanings of complicated words.

Effective Web Service Classification Using a Hybrid of Ontology Generation

317

Table 1. Existing works on service classification. Author

Algorithm

Result

[4]

K-means, Modified NSA

Accuracy: 90.1%

[6]

Ontology learning, Agglomerative

Precision: 92.62%; Recall: 92.82%; F-measure: 92.63%

[9]

Ontology learning, Agglomerative

Precision: 92.75%; Recall: 94.12%; F-measure: 93.43%

[1]

Ontology learning, Agglomerative

Precision: 89.61%; Recall: 80.23%; F-measure: 84.66%

[8]

Ontology learning, Agglomerative

Average Precision: 90.8%; Average Recall: 91.22%; Average F-Measure: 90.6%

[11]

SVMs and SLS

Accuracy: 84.86%

[14]

Naive Bayes

Accuracy: 90%

3 Proposed Classification Approach In order to accomplish the proposed classification approach, a few works need to be done in a flow, and each work in this flow performs a significant role in the proposed approach. Figure 1 illustrates the workflow of the proposed approach.

Fig. 1. Overview of the proposed approach.

3.1 Dataset Collection Data collection and pre-processing are among the most critical activities of any machine learning model. There are several web service repositories available on the internet, such as webservicelist.com, woogle.com, seekda.com, and programmableWeb.com. Among

318

M. Monzur et al.

these repositories, ProgrammableWeb.com provides all the necessary information about RESTful web services of various categories. A web crawler tool is used to acquire all the important features for a web service such as serviceAPI name, serviceAPIHref, serviceTags, serviceDescription, serviceCategory, etc. A total of 12,920 web services details were collected based on some particular domains including agriculture, entertainment, communication, finance, education, food, healthcare, simulation, travel, security and media. 3.2 Data Pre-processing Data pre-processing is also important for better performance. Cleaning and replacing all of the null values with accurate or approximate information is an important step in pre-processing data. The WEKA® machine learning workbench was used for feature selection and pre-processing purposes. Feature selection is the process of removing irrelevant or redundant features without losing any important information. The purpose of feature selection is to enhance the capability of an algorithm by minimizing redundancy and optimizing relevant data. In addition, it reduced the necessary storage space and processing time. 3.3 Designing Context Ontology The most critical part of the ontology generation is distinguishing the semantically meaningful concepts and connections that exist among concepts. After the pre-processing process, the subsequent stage identified the TF–IDF values of all tokenized words, and organized the words in climbing TF–IDF series. The words are ranked by awarding the maximum rank to the word with the maximum TF–IDF value. A threshold value T is then defined. Similarity scores are calculated with the help of similarity filters referred to as proper equivalent, feature equivalent, feature-&-feature equivalent, joint equivalent, relative equivalent, annex equivalent and using Eq. (1) and Eq. (2), along with Table 2 [8]. For a proper equivalent, the similarity score calculated as 1. Sim(Ci, Cj) = Wm + We × ESim(Ci, Cj)

(1)

ESim(Ci, Cj) = − log(d (Ci, Cj)/2D)

(2)

3.4 A Hybrid Approach of Ontology and SVMs Classifier Support Vector Machines (SVMs) classification algorithm offers kernel trick, which is used during the classification process for handling nonlinear data. The algorithm is designed in such a way so that the hyperplane used to separate two data point always follows the largest amount of margin rule. The hyperplane construction was performed in an iterative manner. The primary goal was to minimize the error in classes. The SVMs kernel transformed the data point into the required form, such as a low dimensional input data into a high dimensional input data by adding more dimensions to it (Fig. 2).

Effective Web Service Classification Using a Hybrid of Ontology Generation

319

Table 2. Assigned values for Wm and We [8]. Matching filter

Weight Wm

We

0.89

0.11

Feature-&-feature equivalent 0.86

0.14

Feature equivalent Joint equivalent

0.80

0.20

Relative equivalent

0.75

0.25

Annex equivalent

0.63

0.37

Fig. 2. The algorithm of a hybrid approach using SVMs classifier.

3.5 A Hybrid Approach of Ontology and NB Classifier Naive Bayes (NB) classification algorithm used the Bayes Probability Theorem for classifying the data into several classes. NB is considered as the most straightforward and fastest classification algorithm, and is also suitable for large datasets. During the classification process, every feature is considered as an independent feature. This assumption is also called class conditional independence. This assumption simplified the computation process, and consequently, the classification process becomes fast, accurate and more reliable (Fig. 3).

320

M. Monzur et al.

Fig. 3. The algorithm of a hybrid approach using NB classifier.

3.6 Training and Testing Training and testing procedures are used for evaluation. It is a process where the accuracy, data quality, and necessary output all occur. A total of 80% of the data is used for training from the vast data collection gathered, and the remaining 20% of the data is used for testing. 3.7 Evaluation Metrics The performance measurement is used to ensure the performance or usefulness of the proposed approach, and analyzes the significance of the proposed approach. This work considered the evaluation metrics of accuracy, precision and recall as the performance measurement criteria. Precision (P) is defined as the number of true positives (TP) divided by the number of true positives and the number of false positives (FP) [15]. P(%) = TP/ (TP + FP) × 100

(3)

Recall (R) is defined as the number of true positives (TP) divided by the number of true positives (TP) and the number of false negatives (FN) [15]. R(%) = TP/(TP + FN ) × 100

(4)

Accuracy (A) is defined as the overall proportion accuracy of classification that is classified correctly [15]. A(%) = (TP + TN )/(TP + FP + FN + TN ) × 100

(5)

Effective Web Service Classification Using a Hybrid of Ontology Generation

321

4 Result Analysis During the classification process, particular domains were considered for classification purposes because of the high level of relativeness between those domains. Among the features, the selected features were serviceAPI name, serviceAPIHref, serviceTags and serviceDescription. Based on these features, web services were classified. Among all these features, the serviceTags and serviceDescription were considered as the most important feature, as they described the functionalities provided by a service (Fig. 4).

Feature Importance Service Descripon Service Tags Service Href Service Name 0%

10%

20%

30%

40%

50%

60%

70%

Fig. 4. Feature importance between several features.

Table 3 presents the results obtained using the proposed hybrid classification approach. Table 3. The comparison of the proposed approach with similar approaches. Algorithm

Accuracy

Precision

Recall

Ontology learning, Agglomerative

Not presented

92.62%

92.82%

K-means, Modified NSA

90.1%

Not presented

Not presented

Ontology learning, Agglomerative

Not presented

92.75%

94.12%

Ontology learning, Agglomerative

Not presented

89.61%

80.23%

Ontology learning, Agglomerative

Not presented

90.8%

91.22%

SVMs and SLS

84.86%

Not presented

Not presented

Naive Bayes

90%

Not presented

Not presented

Ontology generation, SVMs

90.59%

89.75%

91.23%

Ontology generation, NB

94.50%

93.00%

95.00%

322

M. Monzur et al.

The results obtained by applying the proposed hybrid classification approach showed significant improvements in terms of accuracy compared with the similar hybrid approaches in prior work. The accuracy helps to deal with the proliferation of functionally similar services by classifying them according to their domain and functions. There are several reasons for achieving better accuracy compared to other works. The service description played a significant role during the classification process. The proposed hybrid approach used an ontology generation method for capturing the similarities between complex terms. Consequently, most of the similar words considered as their base form, and were classified under the same class. Moreover, most of the information and important features were in the form of natural language. The algorithms perform better when provided details are converted to a machine-readable format. The Label Encoder function is used inside the algorithm to transform and fit the natural language into a machine-readable format. Although the hybrid of ontology generation and SVMs classification algorithm provided slightly better accuracy compared to the other hybrid approaches, the precision and recall value was not up to the mark. This is due to longer computational time. The SVMs classifier took more time to train, and as a result, the performance slightly decreased. The hybrid of ontology generation and NB classification algorithm performed better in all three comparison criteria (accuracy, precision and recall). The performance improved due to the use of the Gaussian Classifier, which is suitable for a large chunk of a dataset. The algorithm performed better with the help of the Gaussian Classifier, and provided high accuracy and speed on the large dataset. Thus, the NB classifier performed better in the case of text analysis with low computation time and provided better accuracy, precision and recall values.

5 Future Work This study only considered service classification, not service discovery. However, the proposed hybrid approach can be studied more deeply to expand it towards service discovery. More domains can be considered rather than just considering only a few domains. Furthermore, Artificial Neural Networks (ANNs) and Deep Learning (DL) can be applied to the dataset to classify web services. The accuracy is predicted to be more precise, and the approach will be increasingly solid to utilize in a framework. Acknowledgements. We would like to thank the Ministry of Education (MOE) Malaysia for sponsoring the research through the Fundamental Research Grant Scheme (FRGS) with vote number 5F080 and Universiti Teknologi Malaysia for providing the facilities and supporting the research. In addition, we would like to extend our gratitude to the lab members of Software Engineering Research Group (SERG), School of Computing, Faculty of Engineering, Universiti Teknologi Malaysia for their invaluable ideas and support throughout this study.

References 1. Rupasingha, R.A.H.M., Paik, I., Kumara, B.T.G.S.: Calculating web service similarity using ontology learning with machine learning. In: 2015 IEEE International Conference on Computational Intelligence and Computing Research (ICCIC), Madurai, pp. 1–8 (2015)

Effective Web Service Classification Using a Hybrid of Ontology Generation

323

2. Sambasivam, G., Amudhavel, J., Vengattaraman, T., Dhavachelvan, P.: An QoS based multifaceted matchmaking framework for web services discovery. Future Comput. Inform. J. 3, 371–383 (2018) 3. Cao, Z., Liu, H., Zhang, X.: An efficient algorithm of context-clustered microservice discovery. In: CASE 2018 Proceedings of the 2nd International Conference on Computer Science and Application Engineering, Hohhot, China (2018) 4. Garba, S., Mohamad, R., Saadon, N.A.: Search space reduction approach for self-adaptive web service discovery in dynamic mobile environment. In: Saeed, F., Mohammed, F., Gazem, N. (eds.) Emerging Trends in Intelligent Computing and Informatics. IRICT 2019. Advances in Intelligent Systems and Computing, vol. 1073. Springer, Cham (2020) 5. Raj, M., Pragasam, S.: QoS based classification using K-Nearest Neighbor algorithm for effective web service selection. In: IEEE International Conference on Electrical, Computer and Communication Technologies (ICECCT), Coimbatore, pp. 1–4 (2015) 6. Rupasingha, R.A.H.M., Paik, I., Kumara, B.T.G.S.: Specificity-aware ontology generation for improving web service clustering. IEICE Trans. Inf. Syst. E101-D(8), 2035–2043 (2018) 7. Wen, T., Bao, J., Ding, F.: QoS-aware web service recommendation model based on users and services clustering. In: ICITEE 2018 Proceedings of the International Conference on Information Technology and Electrical Engineering 2018, Xiamen, Fujian, China (2018) 8. Kumara, B.T.G.S., Paik, I., Chen, W.: Web-service clustering with a hybrid of ontology learning and information-retrieval-based term similarity. In: 2013 IEEE 20th International Conference on Web Service, Santa Clara, CA, pp. 340–347 (2013) 9. Rupasingha, R.A.H.M., Paik, I., Kumara, B.T.G.S., Siriweera, T.H.A.S.: Domain-aware web service clustering based on ontology generation by text mining. In: 2016 IEEE 7th Annual Information Technology, Electronics and Mobile Communication Conference (IEMCON), Vancouver, BC, pp. 1–7 (2016) 10. Sasirekha, K., Baby, P.: Agglomerative hierarchical clustering algorithm- a review. Int. J. Sci. Res. Publ. 3(3), 1 (2013) 11. Laachemi, A., Boughaci, D.: A stochastic local search combined with support vector machine for web services classification. In: International Conference on Advanced Aspects of Software Engineering (ICAASE), Constantine, pp. 9–16 (2016) 12. Mohd-Hamka, N., Mohamad, R.: OntoUji–ontology to evaluate domain ontology for semantic web services description. Jurnal Teknologi. 69(6), 21–26 (2014) 13. Mohamad, R., Zeshan, F.: Medical ontology in the dynamic healthcare environment. Procedia Comput. Sci. 10, 340–348 (2012) 14. Liu, J., Tian, Z., Liu, P., Jiang, J., Li, Z.: An approach of semantic web service classification based on Naive Bayes. In: IEEE International Conference on Services Computing (SCC), San Francisco, CA, pp. 356–362 (2016) 15. Sokolova, M., Lapalme, G.: A systematic analysis of performance measures for classification tasks. Inf. Process. Manag. 45(4), 427–437 (2009)

Binary Cuckoo Optimisation Algorithm and Information Theory for Filter-Based Feature Selection Ali Muhammad Usman1,2(B) , Umi Kalsom Yusof1 , and Syibrah Naim3 1 School of Computer Sciences, Universiti Sains Malaysia, 11800 Pulau Pinang, Malaysia

[email protected] 2 Department of Computer Sciences, Federal College of Education (Technical), Gombe, Nigeria 3 Technology Department, Endicott College of International Studies (ECIS),

Woosong University, Daejeon, Korea [email protected]

Abstract. Dimensionality reduction is among the data mining process that is used to reduce the noise and complexity of the features. Feature selection (FS) is a typical dimensionality reduction that is used to reduce the unwanted features from the datasets. FS can be either filter or wrapper. Filters lack interaction among selected subsets of features which in turns affect the classification performance of the chosen subsets of features. This study proposes two ideas of information theory entropy (E) as well as mutual information (MI). Both of them were used together with binary cuckoo optimisation algorithm BCOA (BCOA-E and BCOAMI) to reduce both the error rate and computational complexity on four different datasets. A support vector machine classifier was used to measure the error rates. The results are in favour of BCOA-E in terms of accuracy. In contrast, BCOAMI is computationally faster than BCOA-E. Comparison with other approaches found in the literature shows that the proposed methods performed better in terms of accuracy, number of selected features and execution time. Keywords: Feature selection · Filter-based · Binary Cuckoo optimisation · Information theory

1 Introduction In the various fields of health care, online education, bioinformatics, and social media, among others, data have now become abundant. Since then, the exponential growth of data has become a significant problem for successful data management. As such, data mining and machine learning approaches must be implemented to uncover secret information from these vast data pools [13, 14]. Classification is amongst the methods of data mining that are used to classify each instance into a set of groups. Feature space the only problem downgrading a classifier’s efficiency. Except there is an earlier understanding of the best features, it is otherwise difficult to find the most useful and appropriate features, especially when the size of the feature is large [23]. The term feature © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 324–338, 2021. https://doi.org/10.1007/978-3-030-70713-2_31

Binary Cuckoo Optimisation Algorithm and Information Theory

325

selection (FS) is, therefore, introduced to pick the most essential and appropriate features from these enormous volumes of data. FS’s two main problems are how to search for the best subsets and then evaluate the best one generated [14, 31]. Most current algorithms cannot properly discover the enormous space of an FS without deprived of being stuck in some local optima [6, 10, 29]. Evolutionary algorithms(EAs) are now being used as search methods to elucidate FS problems; nevertheless, several of them still grieve from early convergence. Cuckoo Optimization Algorithm (COA) introduced by [22] is one of the EAs mentioned in [13, 14, 26] that have qualified search operators and can contribute to the search space realisation of the most promising area and converge more rapidly than many other EAs. It depends on the method of the FS to assess or evaluate the best subsets of the generated features. To determine the accuracy or error rate of the chosen subset of features, the wrapper method of the FS uses a classification algorithm and selects the subsets with better accuracy. However, these processes are highly computationally expensive, particularly on high-dimensional datasets [20, 25]. Filter methods, alternatively, are computationally fast and can scale speedily to large dimensional datasets. A lack of feature dependence or relationship between the selected features is one of its significant downsides [13, 31]. Therefore, this study will address the issue of the feature dependency among selected subsets of features. Information theory is a practical approach that can able to measure the relevance within two or more features together with their class label in feature ranking. The most frequently used ones are mutual information (MI) and entropy [4]. Researchers are now using the concepts of both MI and entropy to find the significance and redundancy of the selected features by combining them with various EAs. For instance, [3] used both entropy as well as MI as a fitness evaluation measure in Binary Particle Swarm Optimisation (BPSO). In the work of Mlakar et al. in [17], MI is being used along with PSO. Besides, the Particle Swarm Optimisation (PSO) was used to enhance crowding features and clustering to obtain the best subset of features. Recently, [9] used differential evolution (DE) for feature ranking with the help of MI, Relief-F and Fisher scores. The results obtained surpass both the single and multi-objective approaches presented. All these previous work testified that the concept of information theory is successful in addressing the problems of FS. Thus, in this paper, the enhanced version of the COA, the Binary COA (BCOA) developed by [16] that is suitable for handling FS is proposed as a search technique together with MI and information gained based entropy as the filter evaluation measures. The remainder of the paper is standardised as follows: Sect. 2 describes BCOA, MI, as well as entropy. Section 3 is the proposed filter-based BCOA (BCOAMI and BCOAE) along with the experimentation. Section 4 describes the results and discussion. Lastly, in Sect. 5, the conclusions were offered as well as further research directions.

2 Background This section describes all the ingredients that are used to carry out this study. It includes the BCOA, MI, and gain ratio based entropy.

326

A. M. Usman et al.

2.1 Binary Cuckoo Optimisation Algorithm Binary COA was proposed in [16] since the original COA is meant to solve an only continuous optimisation problem. The BCOA is the most suitable in solving FS problems than its COA counterpart. To calculate the X G and X CP of the habitat in the COA in [22] we use: YNH = XCP + rand (X G − XCP )

(1)

To create a new habitat X NH suitable for discrete binary problems, a sigmoid function (Sig) in the Eq. 2 was used. The reason is to map X NH into the range [0, 1]. Then Eq. 3 will alter the values in the habitat as 0 or 1. Whereby rand in Eq. 3 is a random number, that is generated randomly. Sig =

1 1 + e−X NH

If Sig > rand Then XNH = 1 Else XNH = 0

(2) (3)

2.2 Information Gain Based Entropy The information gain entropy is calculated based on Eq. 4. The higher values of the entropy signify the same probability of occurrence of each variable in contrast to the low entropy that means the different possibility of event of an incident for each variable.  H (X ) = − Pxi log2 Pxi (4) i

X is the random variable and P(xi ) = Pr{X = P(xi ), xi ∈ X } is the mass probability density of X. 2.3 Mutual Information Mutual information (MI) is the measure of relationship or dependence between two arbitrary variables by providing a means to assess the relevance of the subset of the features. The MI between two features X and Y is defined as [9, 27]:    P(xi , yj )   I (X ; Y ) = − P(xi , yj )log2 P (5) i,j P(xi )P yj Equation 5 shows that the I(X; Y) will be large if the two features X and Y are so much related. Else, I(X; Y) = 0 if X and Y are not related at all.

Binary Cuckoo Optimisation Algorithm and Information Theory

327

2.4 Some Related Works The performance of K nearest neighbour (KNN) and SVM based on current filters is presented by Freeman et al. in [8]. The results have shown that MI can develop a better subset of functionality for SVM and KNN. Also, MI is capable of evolving useful subsets of functionality for the two classification algorithms. The idea of maximum relevance and minimum redundancy within the MI was presented by Peng et al. in [21]. The objective was to find the subset of functionality with reduced redundancy and to improve the relevance with the class label. Based on that, researchers now use it to obtain the relationship or dependency between two pairs of features. But, due to the use of sequential search, it can quickly get trap in the local optima. Estevez et al. use genetic Algorithm (GA) in [5] to remedy the constraint of sequential search. Besides, a normalise FS based MI (NMIFS) was proposed because MI favoured characteristics with higher values. The NMIFS is an improvement of the MIFS, MIFS-U and mRMR methods offered in work Battiti in [2]. However, it is also limited to only one pair of features, and yet a non-optimal set of features are likely to be chosen. This motivates many researchers to use other optimisation algorithms, that can search for the best optimal subset of features with best classification performance. For example, Cervante et al., in [3] used a binary PSO together with entropy and MI as evaluation criteria. The results obtained on the four datasets showed that BPSO with mutual information could evolve a set of features with a fewer number of features. Whereas, BPSO with entropy has more classification accuracy using a DT compared to BPSO with MI. Moreover, Moghadasian and Hosseini in [18] used MI and entropy are used as evaluation criteria on some six high dimensional datasets. An artificial neural network was used to measure the classification accuracy and cuckoo search as the search technique. The experimental results displayed that around 90% of the main features was minimised and yet achieved better classification accuracy than using full-length features. In the work of Mlakar et al. in [17], the concept of MI is being used along with PSO. Besides, the PSO is to enhance crowding features and clustering to obtain the best subset of features. Recently, Huda et al. in [11] use a group-based PSO by updating the Pbest along with the Gbest to get the relevance features while ignoring the redundant features. Moslehi et al. in [19] proposed a hybrid filter-wrapper FS by combining GA and PSO along with Artificial Neural Network on five different datasets. Liu et al. in [14] presented a survey paper on the optimisation algorithm that has been used for FS. Out of the numerous algorithms, they conclude that there is still a chance to use other algorithms that are not fully explored in the FS domain. Recently, Usman et al. in [28] presented a comparative analysis among some nature-inspired algorithms for feature selection on some medical datasets. The results obtained showed that binary flower pollination algorithm performed better than the standard flower pollination algorithm in terms of both the number of selected features and classification accuracy. Moreover, the proposed BPFA performed better than harmony search and particle swarm optimisation that uses rough set and quick reduct, respectively. Recently Usman et al. in [30] use the concepts of BCOA for filter-based FS but its limited gain ratio based entropy. Other optimisation algorithms are now becoming popular in dealing with FS problems. For example, Mafarja et al. in [15] hybridised Whale Optimisation Algorithm (WOA) together with Simulated Annealing (SA) to solve FS problems. The datasets

328

A. M. Usman et al.

used coincided with the datasets used in this study. Hence is used for comparison despite the fact that its wrapper-based approach. Similarly, Samy et al. in [24] introduced a new binary WOA for FS based on whales’ behaviour. The Optimum-Path Forest technique is used as an objective function. The results obtained was tested on five colour images datasets. It’s found that the process is much faster than the other classification techniques. In another perspective, Arora et al. in [1] presented two binary variants of the Butterfly Optimization Algorithm (BOA). Among which, two transfer functions are used to map the continuous search space to a discrete one. Twenty-one datasets are used in the experiments. The superior performance of the proposed binary variants is proved in the experiments. Moreover, Jain et al. in [12] offer an enhanced binary version of Gravitational Search Algorithm (GSA) is presented, which is based on the law of gravity and attraction of masses to address this problem of feature selection in medical data. The speed of a random forest classifier is combined with the optimisation behaviour of the GSA. A substantial improvement was recorded in terms of the prediction accuracy. Furthermore, Hichem et al. in [10] presented a new binary Grasshopper Optimisation Algorithm for FS. Whereby, the binarisation of continuous space transforms the continuous values of the continuous space into binary values 0 or 1 in the binary space was realised. Lately, Tahir et al. in [25] presented a novel Binary Chaotic GA for FS in healthcare. To conclude, Fahad et al. in [6] introduced symmetric uncertainty based Ant Colony Optimisation Algorithm for streaming FS in high dimensional medical datasets. The review of the related works shows that optimisation algorithms are becoming more relevant in dealing with different kinds of FS problems. They are used explicitly as search techniques, to search for the most relevant subsets of features. On the other hand, the concepts of information theory play a vital role as a filter evaluation measure, specifically in the filter-based approach.

3 Proposed Approaches In this section, the two filter evaluation measures are being used together with BCOA to form BCOA-MI and BCOA-E. The detail is explained below. 3.1 BCOA Based MI for FS The MI is used to measure the relationship between two pair of features along with their target class. As such, it is used to measure the relevance and redundancy between two couple of features during the feature interaction between them. Based on that, BCOAMI is proposed containing both the relevance and redundancy as the fitness evaluation measure that guides the BCOA to hunt for the subset of features. It is indicated in the Eq. 6: Fmi = −β(Rel mi + Red mi ) − Red mi Rel mi (X ; C) = max

 i

I (x; c)and Red mi (X ; Y ) = min

(6)  1   I xi ; yj |m| i,j

Binary Cuckoo Optimisation Algorithm and Information Theory

329

c and X represent the target class and the discrete binary feature subsets, respectively. The Rel mi uses a pairwise method to calculate the MI between every feature and its target class, that ultimately determine the relevancy of the chosen feature subsets to the target class. Red mi evaluates the MI shared by each pair of the selected features, which means that there is redundancy inside the selected features. Thus, Eq. 6 F mi is s a maximisation function because it maximises the relevancy Relmi and simultaneously minimises the Red mi of the selected features. 3.2 BCOA Based Information Gain Entropy for FS Unlike the F mi that is considered as two-way relevance and redundancy, in FS, Feature interaction may happen in more than two ways; we may have a group of features interactions. Therefore, BCOA-E is proposed to consider a group of features during feature interaction. Hence, the fitness function is clearly defined, as shown in the Eq. 7. FE = −β(Rel E + Red mi ) − Red E Rel E (X ; C) = maxIG

 i

I (x; c)and Red E (X ; Y ) = min

(7) 1  IG(x{X /x}) |m| i,j

Also, RelE evaluates information gain of c given the information on the features in X, and this indicates the relevancy between the selected subset of features as well as the target class. On the other hand, RedE assesses the combined entropy of all the given features in X, and this shows that there is redundancy inside the chosen subsets of features. Therefore, Eq. 7 FE is also considered as a maximisation function that maximises relevancy RelE and concurrently minimises the redundancy RedE among the selected subset of features.

330

A. M. Usman et al.

3.3 Relevance and Redundancy Weighted Values in BCOA-MI and βCOA-E It can be discerned that both Eq. 6 and Eq. 7 has a β1 and β2 respectively. The essence of the β values is to see which one can significantly improve the relevance and consequently reduced redundancy. Based on that, we sum up the relevance and redundancy, then multiplied it with the values and deduct it from the outcome. The reason is that; relevance is needed the most than the redundancy for the optimal result as reported by Hancer et al. in [10]. The weighted values used by Cervante et al. in [3] are adopted in this study. 3.4 Experimental Design Table 1 depicts the datasets used in this study, and it can be found in Frank and Asuncion [7]. From the table, four datasets are used in the experiments with WaveformEW having the highest number of features and instances while Lympography is having the least. The initial and maximum population of the BCOA are set to twenty and thirty; for the thirty different runs. SVM was used to measure the classification accuracy. The datasets are divided into a training set (70%) and testing set (30%). Besides, ten-fold cross-validation was used on each dataset. Table 1. Experimental detests. S/N Detests

Features Instances

1

Lymphography 18

148

2

SpectEW

22

267

3

KrvskpEW

36

3196

4

WaveformEW

40

5000

4 Discussion of Results Table 2, 3, 4 and 5 show the results of the proposed methods. Firstly, BCOA-MI and BCOA-E results are displayed in Table 2. From the Table “Ave Size”, “Ave Acc”, “Best Acc”, “Time” and “All” represent the average number of selected features, ave age accuracy, best accuracy, time, and all features, respectively. 4.1 Results of BCOA-MI and BCOA-E Table 2 shows the results of BCOA-MI along with BCOA-E without any weight function. It can be observed from the results that BCOA-MI performed much better on the average size features selected in all the datasets where around 75% of the total features is reduced. In contrast to the BCOA-E which performed much better in terms of accuracy. Similarly, a less computational time was recorded in the BCOA-MI, and this is due to the pair number of features it deals with compared to BCOA-E that used a group of features. The results clearly showed that both BCOA-MI and BCOA-E could significantly minimise the feature size and attain an improved or similar performance than using the full features.

Binary Cuckoo Optimisation Algorithm and Information Theory

331

Table 2. Experimental results of the proposed algorithms (BCOA-MI) and (BCOA-E). Detests

Approach

Ave-size

Ave-Acc (Best Acc)

Lymphography

All

18

0.875

BCOA-MI

3

0.840 (0.850)

1.68

BCOA-E

4.8

0.855 (0.859)

52.08

All

22

0.851

BCOA-MI

4

0.881 (0.884)

1.85

BCOA-E

4.2

0.888 (0.904)

54.21

All

36

0.892

BCOA-MI

4.2

0.920 (0.945)

56.11

BCOA-E

13.9

0.980 (0.984)

1649.60

All

40

0.771

BCOA-MI

17.5

0.660 (0.660)

172.62

BCOA-E

20.2

0.760 (0.760)

5100.90

SpectEW

KrvskpEW

WaveformEW

Time

4.2 Results of BCOA-MI and BCOA-E with BWeighted Values From Table 3, it can be seen that the higher the β1 value in BOCA-MI, the better the accuracy in the entire datasets. Therefore, the relevance is more significant than the redundancy, which consequently leads to higher accuracy on the higher values of the β1. But looking at the WaveformEW dataset in the table when β1 = 0:9 and 0:8 the different between the best values is not much they are 0.778 and 0.779 respectively. Moreover, the feature size got reduced to around 70%. On the other hand, the higher the β2 value in BCOAE depicted in Table 4, the higher the number of the selected feature. The number of features reduced by almost 40% compared to the full-length features. Also, the accuracy increases as the β2 increase in the majority of the datasets. Comparison between BCOA-MI with β1 in Table 3 along with BCOA-E with β2 in Table 4, one can notice that: (i. β1 is worse than β2 in terms of accuracy (ii. β2 is worse than β1 in terms of the number of selected features and (iii. β1 is computationally less expensive compared to β2. Employing both β1 and β2 values within the filter evaluation measures could significantly reduce the number of features and obtained appropriate classification accuracy than using the full-length features. The “Std” in both Table 3 and Table 4 represent the standard deviation in all the thirty different runs. 4.3 Average Fitness of BCOA-MI and BCOA-E Table 5 shows that the proposed BCOA-E converged earlier with least fitness value than the BCOA-MI on all the four datasets. Although, BCOA-MI recorded the highest fitness value it mostly obtained the best classification performance in terms number of selected

332

A. M. Usman et al. Table 3. Results of the BCOA-MI with different weights of β1. Detests

β1

Ave-size

Ave-Acc (Best Acc)

Std

Lymphography

0.9

7.8

0.860 (0.888)

0.013

1.69

0.8

5.2

0.840 (0.850)

0.013

1.69

0.7

4.9

0.834 (0.834)

0.000

1.69

0.6

4.1

0.800 (0.800)

0.000

1.68

0.5

3

0.780 (0.799)

0.001

1.68

0.9

9.2

0.888 (0.894)

0.012

1.87

0.8

7.8

0.871 (0.885)

0.012

1.87

0.7

5.6

0.844 (0.855)

0.011

1.86

0.6

4.2

0.833 (0.840)

0.011

1.86

SpectEW

KrvskpEW

WaveformEW

Time

0.5

4

0.830 (0.830)

0.000

1.85

0.9

17.2

0.942 (0.946)

0.001

59.55

0.8

16.7

0.935 (0.940)

0.002

57.45

0.7

15.2

0.930 (0.937)

0.002

57.11

0.6

14.2

0.924 (0.925)

0.001

56.13

0.5

12.2

0.920 (0.923)

0.001

56.11

0.9

21.4

0.775 (0.778)

0.001

179.9

0.8

20.2

0.770 (0.779)

0.004

175.7

0.7

19.2

0.760 (0.774)

0.003

174.0

0.6

18.4

0.688 (0.727)

0.000

172.6

0.5

17.5

0.660 (0.660)

0.000

172.6

features and computational time compared to its BCOA-E counterpart as shown earlier in Table 2, Table 3 and Table 4. Also, the values of the standard deviation in the table are within the required standard limit in all the iterations. 4.4 Convergence Trends of BCOA-MI and BCOA-E Figure 1 shows the convergence of the proposed BCOA-MI and BCOA-E. At the top of the chart is the name of the dataset, while the fitness and number of iterations are represented on the x-axis and y-axis respectively. From at the curve on each graph, it can be observed that BCOA-E is at the bottom compared to the BCOA-MI, this means that BCOA-E converges to the best fitness compare to the BCOA-MI. Perhaps, it can be due to interaction among a group of features in the BCOA-E. On the other hand, BCOA-MI has limited feature interaction, since it interacts with only pair features at a time.

Binary Cuckoo Optimisation Algorithm and Information Theory Table 4. Results of the BCOA-E with different weights of β2. Detests

β2

Ave-size

Ave-Acc (Best Acc)

Std

Time

Lymphography

0.9

12.6

0.890 (0.890)

0.000

52.39

0.8

10.5

0.880 (0.888)

0.000

52.39

0.7

8.9

0.874 (0.879)

0.001

52.39

0.6

6.4

0.860 (0.872)

0.001

52.08

0.5

5.1

0.855 (0.859)

0.001

51.46

0.9

10.2

0.899 (0.914)

0.004

54.79

0.8

8.7

0.891 (0.895)

0.001

54.79

0.7

6.8

0.884 (0.889)

0.001

54.5

0.6

5.2

0.871 (0.880)

0.002

54.5

SpectEW

KrvskpEW

WaveformEW

0.5

5

0.862 (0.869)

0.001

54.21

0.9

19.2

0.972 (0.976)

0.001

1750.8

0.8

18.4

0.965 (0.980)

0.005

1689

0.7

16.3

0.950 (0.977)

0.005

1679

0.6

15.4

0.944 (0.945)

0.001

1650.2

0.5

13.9

0.929 (0.933)

0.001

1649.6

0.9

26.5

0.822 (0.888)

0.004

5315.2

0.8

24.3

0.790 (0.819)

0.003

5190.5

0.7

21.2

0.770 (0.785)

0.003

5140.2

0.6

20.1

0.768 (0.769)

0.001

5100.9

0.5

19.2

0.760 (0.760)

0.000

5100.9

Table 5. Average fitness for BCOA-MI and BCOA-E. Detests

BCOA-MI Fitness

BCO-E StdDev Fitness StdDev

Lymphography 0.158

0.001

0.131

0.001

SpectEW

0.179

0.001

0.137

0.002

KrvskpEW

0.064

0.000

0.055

0.002

WaveformEW

0.279

0.000

0.274

0.000

333

334

A. M. Usman et al.

Fig. 1. Convergence trends of BCOA-MI and BCOA-E.

4.5 Comparison with Other Existing Approaches The results obtained are compared with the existing work, that works with similar datasets, for example, BPSO-MI and BPSO-E in the work Cervante et al. in [3] together with WOA-SA in work Mafarja et al. in [16]. The detailed of the comparison is depicted in Table 6. The proposed results were compared with the existing works in terms of numbers of selected features, classification accuracy and the time it takes to finish its execution on each dataset during the thirty independent runs. In all aspect, our proposed methods performed better than the BPSOMI and BPSOE. Whereas, in terms of the computational time, our approaches performed better than WOA-SA excepts on Lymphography dataset whereWOA-SA recorded the least time. In terms of accuracy, WOA-SA achieved the best accuracy in two of the datasets, while our proposed methods achieved the best accuracy on the remaining two datasets. Our proposed methods recorded the least number of features in all datasets compared to the other approaches. Therefore, one can conclude that the proposed methods performed better than the existing works in terms of the number of selected features, computational time as well as classification accuracy.

Binary Cuckoo Optimisation Algorithm and Information Theory

335

Table 6. Comparison of the proposed algorithms with other existing approaches. Detests

Approach

Ave-Size

Ave-Acc (Best Acc)

Lymphography

All

18

0.875

BCOA-MI

3

0.780 (0.799)

0.001

1.66

BCOA-E

5.1

0.855 (0.859)

0.001

52.08

All

18

0.755

BPSO-MI

3

0.711 (0.711)

0.000

3.89

BPSO-E

6.3

0.740 (0.778)

0.017

61.45

WOA-SA

7.2

0.890

All

22

0.851

BCOA-MI

4

0.830 (0.830)

0.000

1.85

BCOA-E

4.2

0.862 (0.869)

0.001

54.21 2.13

SpectEW

KrvskpEW

WaveformEW

Std-Acc

Time

1.66

All

22

0.809

BPSO-MI

3.1

0.783 (0.794)

0.002

BPSO-E

4.5

0.812 (828)

0.010

WOA-SA

6

0.880

All

36

0.892

BCOA-MI

4.2

0.920 (0.945)

0.001

56.11

BCOA-E

13.9

0.980 (0.984)

0.001

649.60

All

36

0.985

BPSO-MI

4.7

0.797 (0.902)

0.027

76.23

BPSO-E

15.7

0.970 (0.977)

0.011

203.67

WOA-SA

12.8

0.980

641.0

641.01

All

40

0.771

BCOA-MI

17.5

0.660 (0.660)

0.000

172.62

BCOA-E

20.2

0.760 (0.760)

0.000

5100.90

62.89 313.38

All

40

0.696

BPSO-MI

19.4

0.620 (0.649)

0.011

1497.9

BPSO-E

20.9

0.688(0.698)

0.002

6102.76

WOA-SA

20.6

0.770

1770.48

5 Conclusions and Future Works The aim of this paper has been achieved by developing two filter-based evaluation measures based on entropy and MI, together with BCOA. The results demonstrated that BCOA-MI is capable of evaluating the relevance and redundancy of the pair features. In comparison, BCO-E shows its priority in assessing both the relevance and redundancy

336

A. M. Usman et al.

when dealing with a group of features. In either case, weighted values are employed. And it is found that the higher the values, the higher the number of features and the accuracy. BCOA-MI recorded the least accuracy compared with BCOA-E. Perhaps, it might be due to the feature interaction among a group of features by the BCOA-E. On the other hand, BCOA-E is computationally expensive compared with the BCOAMI. BCOA-MI interacts with only pair features that make it computationally faster. Apart from using different newer optimisation algorithms to solve similar problems for competitive results, in the future, we will investigate the use of the nondominated sorting mechanism together with BCOA to solve the conflicting issues in FS rather than using the weighted values. Acknowledgement. This document is the results of the research project funded by the Universiti Sains Malaysia via Research University Grant (RUI) (1001/PKOMP/8014084) together with Woosong University, Korea.

List of Acronyms FS MI COA BCOA BCOA-MI BCOA-E WOA-SA BPSOMI BPSOE

Feature Selection Mutual Information Cuckoo Optimisation Algorithm Binary Cuckoo Optimisation Algorithm Binary Cuckoo Optimisation Algorithm Mutual Information Binary Cuckoo Optimisation Algorithm Entropy Wolf Optimisation Algorithm Simulated Annealing Binary Particle Swarm Optimisation Mutual Information Binary Particle Swarm Optimisation Entropy

References 1. Arora, S., Anand, P.: Binary butterfly optimisation approaches for feature selection. Expert Syst. Appl. 116, 147–160 (2019) 2. Battiti, R.: Using mutual information for selecting features in supervised neural net learning. IEEE Trans. Neural Netw. 5(4), 537–550 (1994) 3. Cervante, L., Xue, B., Zhang, M., Shang, L.: Binary particle swarm optimisation for feature selection: a filter based approach. In: 2012 IEEE Congress on Evolutionary Computation (CEC), pp. 1–8. IEEE (2012) 4. Dash, M., Liu, H.: Feature selection for classification. Intell. Data Anal. 1(3), 131–156 (1997) 5. Estevez, P.A., Tesmer, M., Perez, C.A., Zurada, J.M.: Normalised mutual information feature selection. IEEE Trans. Neural Netw. 20(2), 189–201 (2009) 6. Fahad, L.G., Tahir, S.F., Shahzad, W., Hassan, M., Alquhayz, H., Hassan, R.: Ant colony optimisation-based streaming feature selection: an application to the medical image diagnosis. Sci. Program. 2020 (2020) 7. Frank, A., Asuncion, A.: UCI Machine Learning Repository, vol. 213, p. 2. School of Information and Computer Science, University of California (2010). https://archive.ics.uci.edu/ml. irvine,ca

Binary Cuckoo Optimisation Algorithm and Information Theory

337

8. Freeman, C., Kuli´c, D., Basir, O.: An evaluation of classifier-specific filter measure performance for feature selection. Pattern Recogn. 48(5), 1812–1826 (2015) 9. Hancer, E., Xue, B., Zhang, M.: Differential evolution for filter feature selection based on information theory and feature ranking. Knowl. Based Syst. 140, 103–119 (2018) 10. Hancer, E., Xue, B., Zhang, M., Karaboga, D., Akay, B.: Pareto front feature selection based on artificial bee colony optimisation. Inf. Sci. 422, 462–479 (2018) 11. Hichem, H., Elkamel, M., Rafik, M., Mesaaoud, M.T., Ouahiba, C.: A new binary grasshopper optimisation algorithm for feature selection problem. J. King Saud Univ. Comput. Inf. Sci. (2019) 12. Huda, R.K., Banka, H.: A group evaluation based binary pso algorithm for feature selection in high dimensional data. Evol. Intell., 15 (2020) 13. Jain, R., Sawhney, R., Mathur, P.: Feature selection for cryotherapy and immunotherapy treatment methods based on gravitational search algorithm. In: 2018 International Conference on Current Trends towards Converging Technologies (ICCTCT), pp. 1–7. IEEE (2018) 14. Li, J., Cheng, K., Wang, S., Morstatter, F., Trevino, R.P., Tang, J., Liu, H.: Feature selection: a data perspective. ACM Comput. Surv. (CSUR) 50(6), 94 (2017) 15. Liu, W., Wang, J.: A brief survey on nature-inspired metaheuristics for feature selection in classification in this decade. In: 2019 IEEE 16th International Conference on Networking, Sensing and Control (ICNSC), pp. 424–429. IEEE (2019) 16. Mafarja, M.M., Mirjalili, S.: Hybrid whale optimisation algorithm with simulated annealing for feature selection. Neurocomputing 260, 302–312 (2017) 17. Mahmoudi, S., Rajabioun, R., Lotfi, S.: Binary cuckoo optimisation algorithm. Nature, pp. 1–7 (2013) 18. Mlakar, U., Fister, I., Brest, J.: Hybrid multi-objective PSO for filter-based feature selection. In: 23rd International Conference on Soft Computing. pp. 113–123. Springer, Cham (2017) 19. Moghadasian, M., Hosseini, S.P.: Binary cuckoo optimisation algorithm for feature selection in high-dimensional datasets. In: International Conference on Innovative Engineering Technologies, ICIET2014, pp. 18–21 (2014) 20. Moslehi, F., Haeri, A.: A novel hybrid wrapper–filter approach based on genetic algorithm, particle swarm optimisation for feature subset selection. J. Ambient Intell. Humanized Comput. 11(3), 1105–1127 (2020) 21. Nogueira, S., Sechidis, K., Brown, G.: On the stability of feature selection algorithms. J. Mach. Learn. Res. 18(1), 6345–6398 (2017) 22. Peng, H., Long, F., Ding, C.: Feature selection based on mutual information criteria of maxdependency, max-relevance, and min-redundancy. IEEE Trans. Pattern Anal. Mach. Intell. 27(8), 1226–1238 (2005) 23. Rajabioun, R.: Cuckoo optimisation algorithm. Appl. Soft Comput. 11(8), 5508–5518 (2011) 24. Russell, S.J., Russell, P.: Artificial Intelligence: A Modern Approach. Pearson Education Limited, Malaysia (2016) 25. Samy, A., Hosny, K.M., Zaied, A.N.H.: An efficient binary whale optimisation algorithm with optimum path forest for feature selection. Int. J. Comput. Appl. Technol. 63(1–2), 41–54 (2020) 26. Tahir, M., Tubaishat, A., Al-Obeidat, F., Shah, B., Halim, Z., Waqas, M.: A novel binary chaotic genetic algorithm for feature selection and its utility in affective computing and healthcare. Neural Comput. Appl., 1–22 (2020) 27. Tavana, M., Shahdi-Pashaki, S., Teymourian, E., Santos-Arteaga, F.J., Komaki, M.: A discrete cuckoo optimisation algorithm for consolidation in cloud computing. Comput. Ind. Eng. 115, 495–511 (2018) 28. Tsanas, A., Little, M.A., McSharry, P.E.: A simple filter benchmark for feature selection. J. Mach. Learn. Res. 1, 1–24 (2010)

338

A. M. Usman et al.

29. Usman, A.M., Abdullah, A.U., Adamu, A., Ahmed, M.M.: Comparative evaluation of naturebased optimisation algorithms for feature selection on some medical datasets. i-manager’s comparative evaluation of naturebased optimisation algorithms for feature selection on some medical datasets. i-manager’s J. Image Process. 5(4), 9 (2018) 30. Usman, A.M., Yusof, U.K., Naim, S.: Cuckoo inspired algorithms for feature selection in heart disease prediction. Int. J. Adv. Intell. Inf. 4(2), 95–106 (2018) 31. Usman, A.M., Yusof, U.K., Naim, S.: Filter-based multi-objective feature selection using NSGA III and Cuckoo optimisation algorithm. IEEE Access 8, 76333–76356 (2020) 32. Xue, B., Zhang, M., Browne, W.N., Yao, X.: A survey on evolutionary computation approaches to feature selection. IEEE Trans. Evol. Comput. 20(4), 606–626 (2016)

Optimized Text Classification Using Correlated Based Improved Genetic Algorithm Thabit Sabbah(B) Al Quds Open University (QOU), Ramallah, Palestine [email protected]

Abstract. Text Classification (TC) is one of the basic processes in many Information Retrieval systems. Still, the performance of TC is a subject of improvement, and many approaches were proposes to achieve this aim. This work proposes an Improved Genetic algorithm (IGA) inspired by Genetic Engineering to enhance TC performance. In IGA, chromosome generation process were re-designed to diminish the effect of correlated genes. The Support Vector Machine (SVM) classifier were utilized based on the “Sport Text” popular dataset to evaluate the proposed approach. Empirical classification results were improved using IGA as compared to normal GA optimization. The proposed Improved Genetic Algorithm (IGA) improved the correct rates of TC by 1.39% in average. Keywords: Text classification · Improved Genetic Algorithm · Genetic engineering · Feature correlation

1 Introduction 1.1 Text Classification Text Classification and Text Categorization are common terms in the field of text analysis, however, the slightly difference between the two terms is related to the output while almost the process is similar. Text Classification places each portion of text whether it is a tweet, a paragraph, or even a document into a predefined class which is also known as label [1]. However, in Text categorization, usually, the classes are not predefined, and it is the responsibility of the classifier to determine how many and what are the categories to be generated as well as the placement of text portions among these categories [2]. In Text Classification, usually, an initial preprocessing stage is applied, and then the documents are represented numerically in Vector Space Model on which classifying algorithms such as Support Vector Machine can complete the task. For the aim of numerical representation of text, many different methods can be applied, however, most of these methods are based on assigning a numerical value (known as weight) for each term in the text [3].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 339–350, 2021. https://doi.org/10.1007/978-3-030-70713-2_32

340

T. Sabbah

The performance of TC is an open research problem, since there is no optimal solution has been provided, although many approaches were proposed in the last few decades. This research deals with Text Classification problem where the output classes are predefined. Although the aim is to enhance the performance of TC, however, the major contribution in this work is proposing and utilizing an Improved correlation based Genetic Algorithm (IGA), hence, the classical text classification approach is applied. The application of the proposed Improved Genetic Algorithm (IGA) for text classification requires the classification process to be fitted into the general approach of genetic optimization, which is based on the natural selection of effective features that lead to the higher classification performance, as discussed in the subsequent section. 1.2 Genetic Optimization Genetic Algorithm (GA) is an optimization method, which was initiated by [4]. GA is inspired from natural selection process and motivated by the saying “survival of the fittest”. During the last few periods, GA were employed in resolving optimization problems [5] in numerous applications and research domains such as classification, clustering, spam detection [6–8] and more. GA imitates the process of natural evolution by leading the random search in the space to find the optimal solution of the optimization problem. The random searches are guided through maintaining and merging the “effective” factors of a good solution to yield improved solutions. Commonly, GA consists of some stages that include many steps, following are brief descriptions of the general GA stages and steps as depicted in Fig. 1. New Start

Population

Initial Initialization Stage Population

Sequent population Generation

Evaluation

Mutation Crossover Selection

End

Yes

Termination

Elitism

No

Current Population

Fig. 1. General stages of genetic algorithm.

Initialization GA optimization process starts with the random generation of a population, were a population consists of individuals that are also known as chromosomes that represent a group of possible solutions of the problem. Usually in GA, solution is representable in various ways, such as binary string or numerical values [9]. In this work the binary string of zeros and ones representation is applied, in which the binary one indicates that

Optimized Text Classification Using Correlated Based IGA

341

the corresponding feature is involved in the solution. However, the sequence of binary zeros and ones string, which forms the chromosome usually, are referred as the genes. Fitness Evaluation and Ranking In this stage, the generated population is assessed; the fitness of each individual in the current population is calculated. The fitness value of an individual is determined by the fitness function, which is the function that measures the quality of the solution (i.e. how close is a solution to the best solution of the problem), and then the individuals are ranked based on their fitness values. In a classification problem, the accuracy or the correct rate and other performance measures can acts as the fitness function. Hence, this work considers the classification correct rate as the fitness value. After ranking, the algorithm terminates if termination condition is satisfied, otherwise, the algorithm iterates while generating successive populations through the stages of Selection and Sequent population generation that includes the processes of elitism, crossover, and mutation. Sequent Generation Production If a new population is to be generated (i.e. next generation) after the Selection process, GA usually divides the individuals into three major parts. These parts are: one: those individuals who will survive in the new generation as they are (i.e. without any modification), two: the individuals that will involve breeding to generate new individuals, and three: the individuals will be mutated. The first part is determined using the “Elitism” or the elitist selection strategy [10], in which the individuals with high fitness values are usually selected to remain and survive in the sequent generation; in order to guarantee that the new generation will at least be effective as the current generation. The remaining individuals of the current population are included in part two for breeding through the crossover process, while a random selection process controls the third part in which the individuals will be mutated. In crossover procedure, two individuals (known as parents) are chosen to produce the offspring(s). The frequently used techniques for parents’ selection include Roulette wheel, Tournament, and Stochastic uniform selection [11, 12]. Moreover, there are several frequent techniques which are employable to perform the crossover operation [12, 13] such as the single point, the two points, and the scattered methods. Figure 2 illustrates these methods.

Fig. 2. Popular crossover and mutation methods.

342

T. Sabbah

In single point and two points methods Fig. 2(c-a) and (c-b) of the crossover methods respectively, points (or lines) of crossover are placed on parental chromosomes. The genes located before the crossover point from the first parent and the genes located after the point from the second parent to be concatenated into the new offspring in the single point method. While, in the two points crossover method, two random points (or lines) of crossover are placed on parental chromosomes, and then, the genes located between the two crossover points from the second parent, and the genes located before the first point and after the second point from the first parent chromosome are selected (and vice versa) to produce the new offspring(s). However, in the scattered crossover (see Fig. 2(cc) crossover methods column) which is also known as uniform crossover operation, a binary randomly vector that matches in length the parental chromosomes is generated. Then, the genes corresponding the locations of ones in the binary vector from the first parental chromosome and those from the second parental chromosome corresponding to the zeros locations are selected to produce the new offspring. Mutation is another method used for new chromosomes production; the new individuals are produced through changing of current individuals genes. The mutation process aims to provide genetic diversity to the sequent population and enables the GA to discover a wider search space with the anticipation to increase the possibility of producing new individuals with best fitness, as well as to prevent GA from falling off into a local optimum. Various procedures are frequently applied to perform mutation operation such as Gaussian, flip bit, and interchange methods as shown in Fig. 2 (Mutation Methods Column). Based on different approaches, generally, the parents’ genes are inverted (either binary 1 turned into 0 or 0 turned into 1) to yield the new offspring. In the interchange mutation, shown in Fig. 2(m-a), a single gene is selected randomly from the chosen individual, then the value of selected gene is inverted (either 0 turned to 1 or 1 turned to 0) to produce the new offspring. Similarly, the flip bit mutation in Fig. 2(m-b), selects multiple genes to be inverted based on a generated random mutation vector whose length is same as parent chromosome’s length, the values of the genes from the chosen individual located corresponding to the 1 s locations in the mutation vector are inverted in order to produce the new offspring. However, in gaussian mutation as in Fig. 2(m-c), a random number (or mutation) from a Gaussian distribution is selected and then added to the values of all genes of the selected individual to produce the new offspring. It should be noted that Gaussian mutation could be utilized with the number chromosome representations while the flip bit and interchange are utilizable with binary representations. After the execution of the elitism, crossover, and mutation operations, the new individual chromosomes are aggregated (gathered) to form the new population (sequent generation), then the fitness and ranking process (described in Fitness Evaluation and Ranking) is performed before the check for termination. At termination, the top ranked individual among the last population is considered to the optimized solution. The proposed IGA involves a new mutation method which is applied to all newly generated chromosomes during sequent population generation stage in which the genes are inverted based on their correlation such the genes are injected or ejected from the chromosome to diminish the effect of the appearance of such correlated features in the a chromosomes.

Optimized Text Classification Using Correlated Based IGA

343

2 Related Works Since its proposal in 1975 as a nature inspired approach, GA was utilized in various domains for optimization. However, over the years, many works were proposed to enhance this algorithm. This work, presents the proposed Improved GA in which the improvement is performed in the step of new offspring(s) generation. Table 1 reviews some of the works proposed enhancements on GA in the last few years. Table 1. Review of major GA enhancement proposals Ref.

Proposed enhancement

[14]

Arithmetic crossover and mutation operators with variable length chromosomes

[7]

Mutation based on clustering points of extremism to overcome the limitations of k-means algorithm

[15]

Enhancement on memory updating based on environment schemes reaction for constrained knapsack problems in dynamic environments

[16]

Multi-hop Path Finding fitness function were proposed to extend network’s lifetime

[17]

Multi coefficients’ weighting based Elitism

[18]

Solution vector representation based on feature (sentence) index for text summarization

[19]

Decision trees based on Random forest were generated for each population; best fitness tree is selected for the classifier

[6]

Parents’ chromosomes selection for crossover operation based on cumulative term weights to generate new offspring

[20]

Recovery method from the uniformity (i.e. GA fails to produce better fitness values) by a migration test and step

[8]

Mature convergence approach by metropolis simulated annealing process after classical crossover and mutation operations

[21]

Feature’s subsets size controlling in the fitness function

These approaches list in Table 1 varies in various aspects such as: the domain of application, the stage in which the enhancement were proposed, and the representation of features, as well as the methods used in sequent population generation (i.e. elitism, crossover, and mutation). Although some of these approaches were domain specific, however few works were specified to the TC domain. In the domain of TC, GA has been proposed in many works to play different roles in the process of TC. [6] utilized GA for TC in the stage of feature selection, such that the crossover operation was based on term and document frequencies, while the mutation operation was based on the performance of classifier on the original parents. The thoughts behind this work is to perform these operations on useful information instead of random selection of features.

344

T. Sabbah

Earlier, GA was utilized by [22] for an informative feature subspace selection, where the GA was applied to reduce the dimensionality of the sub-feature space selected using the Information Gain (IG) feature selection method. Similar to [22–25] employed the GA for sub-feature space selection on top of the feature space generated by different traditional feature selection method such as IG, DF, MI, and CHI. Later, [26] has employed GA as a learning technique to improve text categorization by the automatic generation of categorizing rules.

3 Proposed Approach 3.1 Proposed IGA As mentioned earlier, the proposed IGA involves an extra mutation step that is applied to all newly generated chromosomes in which the genes are inverted based on their correlation such that the genes are injected or ejected from the chromosome to diminish the effect of the appearance or absence of correlated features in the a chromosomes. This approach is motivated by the idea of genetic engineering, in which the genes that are responsible of a feature(s) are identified and then manipulated to enhance the new generations. In this work, the correlated features (genes) are identified and then treated by either injection into or discharge from the chromosome to enhance the performance. The hypothesis behind this work is that the existence of the correlated features in a chromosome based on the random selection of features or because of crossover and mutation processes (which is also based on randomization) during chromosomes generation affects the performance of these chromosomes as well as the GA as a result. Therefore, this work proposes an extra mutation process (based on the correlation between features (genes)), such that the highly positively or negatively correlated features in a chromosome are firstly identified and then manipulated. Figure 3 shows the steps of the proposed Improved Genetic Algorithm (IGA). Start Initialization Stage

New Population

Initial

Sequent population Generation

Population Correlation based Mutation

Correlation Matrix Calculation

Mutation

Evaluation

Correlation Matrix

Crossover Elitism

Selection

End

Yes

Termination

No

Current Population

Fig. 3. Improved Genetic Algorithm (IGA).

Optimized Text Classification Using Correlated Based IGA

345

Figure 3 shows the extra step named “Correlation Based Mutation (CBM)” that is located before the “Evaluation” and “Selection” operations. The CBM process requires the calculation of “Correlation Matrix” of the feature space, which is calculated once by the “Correlation Matrix Calculation” process. This process performs the controlled mutation such that if two highly correlated variables (features) appears in a chromosome then one of these features is ejected (removed), however, if the two highly correlated variables (features) are not include in the chromosome then one of these features is injected (planted) into the chromosome. The ejection/injection process is achieved by inverting the binary value that represents that feature. The example in Fig. 4 illustrate the CBM.

Before CBM

V1

V2

V3

V4

..

..

..

Vn

1

1

0

0

1

..

..

0

• (V1, V2): variables of highest posive correlaon • (V4,Vn): variables of highest negave correlaon

Aer CBM

(a)

1

0

0

0

1

..

CBM ..

1

(b) Fig. 4. CBM operation illustration

Consider the calculated correlation matrix between feature space variables as in Fig. 4(a), in which the correlation value between V1 and V2 equals 0.5, while the correlation between V4 and Vn equals to −1. Now for any generated chromosome, for example the (Before CBM) chromosome in Fig. 4(b) if it includes both V1 and V2 , then one of them will be ejected as a result of CBM, while if both are absent from the chromosome, one of them will be injected. Moreover, if the chromosome includes V4 and Vn , then one of them will be ejected, while if both are absent from the chromosome, one of them will be injected. As showed in Fig. 4(b), the binary representation of (Before CBM) chromosome shows that V1 and V2 are included in this generated chromosome, however, these two features have the highest positive correlation. As a result, the CBM operation will eject one of these genes (features) by turning one of their binary representation 1 into 0 as showed in (After CBM) chromosome. Moreover, the example shows that none of V4 and Vn is included in the (Before CBM) chromosome; therefore, CBM injected one of these features into the chromosome by turning its binary representation 0 into 1. The shaded cells in the (After CBM) chromosome are the inverted binary values. 3.2 IGA Based Text Classification The employment of IGA for text classification follows the general TC classification approach, in which the Vector Space Model [27] of the dataset is split into two parts (i.e. training and testing), the IGA is applied as the feature selection method during the training phase where the classification correct rate is the fitness value. At the end

346

T. Sabbah

of training phase, the features included by the chromosome that achieved the highest fitness value are considered for testing phase.

4 Experimental Environment To test the performance of the proposed approach, the “Sports articles” dataset [28] is employed. Table 2 shows the utilized dataset specification which was downloaded from the Machine Learning Repository known as UCI1 . Table 2. Dataset specifications Specification

Value

Dataset characteristics

Multivariate, text

Number of labels

2 (subjective, objective)

Number of instances

1000 (365 subjective, 635 objective)

Attribute characteristics Integer Number of attributes

57

Based on the Yarpiz2 “Implementation of Binary GA in MATLAB”, the proposed IGA was run out using MATLAB R2016a software, under a 64bit Windows10 environment on a Core i7 2.1 GHz with 16 GB Ram laptop computer. The Support Vector Machine (SVM) classifier were employed under the default configuration, and the performance were measured based on the classification correct rate. The experiment is conducted 10 times, and the average of best scores in each iteration was recorded. Finally, the results were benchmarked against the basic GA performance under the same experimental parameters. Table 3 shows the IGA experimental parameters. Table 3. Experimental parameters Parameter

Value

Population size

20

Number of generations(iterations)

50

Crossover percentage

0.8

Selection

Roulette wheel (selection pressure = 8)

Mutation rate (mu)

0.02

1 https://archive.ics.uci.edu/ml/index.php. 2 www.yarpiz.com.

Optimized Text Classification Using Correlated Based IGA

347

In this experiment, the case of max positive and negative highly correlated variables were treated, such that the two variables with max positive correlation and the two variables with max negative correlation were treated so that only one variable of each pair enforced to appear in each generated chromosome.

5 Results and Discussions Figure 5 pictures the results of the proposed IGA compared to the basic GA performance on “Sports articles” dataset. It is seen in Fig. 5 that the correct rate of IGA based classification is increasing with the number of iterations in general, and outperforming the results of Basic GA classification. The correct rate of classification based on IGA was enhanced from 84.7% up to 86.2% through the iterations, while the Baseline GA enhances it from 84.4% up to 86.1%. The results based on IGA has an average improvement by 1.39% as compared to the Baseline GA. Moreover, it is noticeable from Fig. 5 that the performance of IGA initial population also outperform the corresponding population of Baseline GA, which also indicates that the proposed improvement is effective at initialization stage. 86.5%

Correct Rate

86.0%

85.5%

85.0% Baseline (GA) IGA

84.5%

84.0% 0

5

10

15

20

25 Iteraons

30

35

40

45

50

Fig. 5. Comparison of IGA and Basic GA results on “Sports articles” dataset

Table 4 shows the achieved classification results through the iterations and the improvements percentage. The described results in Fig. 5 and Table 4, supports the claim that the appearance of correlated features in chromosomes affects the performance of GA that represents the hypothesis behind this work. The proposed technique in this work ensures that a generated chromosome should include only one feature out of each pair of the positively and negatively highly correlated features, which explains the improvements achieved by the proposed IGA.

84.6

84.7

84.8

84.8

84.9

84.9

84.9

84.9

84.9

84.9

84.9

85

85

85.2

85.2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

Average

84.4

2

Baseline (GA)(%)

85.5

85.5

85.5

85.5

85.5

85.3

85.1

85

85

84.9

84.9

84.9

84.9

84.8

84.8

84.7

IGA (%)

Classification results

1

Iteration

1.29

1.29

1.29

1.29

1.25

1.01

0.76

0.68

0.64

0.56

0.56

0.56

0.56

0.42

0.4

0.32

Improvement (%)

85.7 85.7

33

85.7

85.6

85.6

85.5

85.4

85.4

85.4

85.4

85.4

85.4

85.4

85.4

85.3

85.2

85.2

85.8

85.8

85.8

85.8

85.8

85.8

85.7

85.7

85.7

85.6

85.6

85.6

85.6

85.6

85.5

85.5

85.5

Baseline (GA) IGA (%) (%)

Classification results

32

31

30

29

28

27

26

25

24

23

22

21

20

19

18

17

Iteration

1.61

1.61

1.61

1.61

1.61

1.61

1.53

1.53

1.53

1.45

1.45

1.45

1.45

1.45

1.31

1.29

1.29

Improvement (%)

50

49

48

47

46

45

44

43

42

41

40

39

38

37

36

35

34

Iteration

Table 4. Classification results of IGA and Basic GA.

85.4

86.1

86.1

86.1

86

86

86

86

86

86

86

85.8

85.8

85.8

85.8

85.7

85.7

85.7

85.6

86.2

86.2

86.2

86.2

86.1

86.1

86.1

85.9

85.9

85.9

85.9

85.8

85.8

85.8

85.8

85.8

85.8

Baseline (GA) IGA (%) (%)

Classification results

1.39

2.13

2.13

2.11

2.05

2

2

2

1.76

1.72

1.72

1.72

1.65

1.65

1.65

1.61

1.61

1.61

Improvement (%)

348 T. Sabbah

Optimized Text Classification Using Correlated Based IGA

349

6 Conclusions This work presented a proposed improvement of GA named IGA inspired from the domain of Genetic Engineering, in which the improvement involved a controlled based mutation step that is applied to all generated chromosomes after either the initialization or sequent population generation stages. The controlled mutation treats highly correlated gens (feature) of a chromosome, such that only one gene of each highly correlated genes pairs is enforced to appear in the chromosome. The SVM based classification correct rate performance using the popular “Sports articles” dataset showed an improvement of the proposed method based results as compared to the baseline GA results, which supports the initial hypothesis behind this work and promote further research on the topic in various directions.

References 1. Jiang, S., Pang, G., Wu, M., Kuang, L.: An improved K-nearest-neighbor algorithm for text categorization. Exp. Syst. Appl. 39(1), 1503–1509 (2012). https://doi.org/10.1016/j.eswa. 2011.08.040 2. Hartmann, J., Huppertz, J., Schamp, C., Heitmann, M.: Comparing automated text classification methods. Int. J. Res. Mark. 36(1), 20–38 (2019). https://doi.org/10.1016/j.ijresmar.2018. 09.009 3. Sabbah, T., et al.: Modified frequency-based term weighting schemes for text classification. Appl. Soft Comput. 58, 193–206 (2017). https://doi.org/10.1016/j.asoc.2017.04.069 4. Holland, J.H.: Adaptation in Natural and Artificial Systems Ann Arbor. The University of Michigan Press 1, 975 (1975) 5. Thengade, A., Dondal, R.: Genetic algorithm-survey paper. In: MPGI National Multi Conference 2012, pp. 7–8. Citeseer (2012) 6. Ghareb, A.S., Bakar, A.A., Hamdan, A.R.: Hybrid feature selection based on enhanced genetic algorithm for text categorization. Exp. Syst. Appl. 49, 31–47 (2016). https://doi.org/10.1016/ j.eswa.2015.12.004 7. El-Shorbagy, M.A., Ayoub, A.Y., Mousa, A.A., El-Desoky, I.M.: An enhanced genetic algorithm with new mutation for cluster analysis. Comput. Stat. 34(3), 1355–1392 (2019). https:// doi.org/10.1007/s00180-019-00871-5 8. Salehi, S., Selamat, A., Bostanian, M.: Enhanced genetic algorithm for spam detection in email. In: 2011 IEEE 2nd International Conference on Software Engineering and Service Science, 15–17 July 2011, pp. 594–597 (2011) 9. Garzelli, A., Capobianco, L., Nencini, F.: 9 - Fusion of multispectral and panchromatic images as an optimisation problem. In: Stathaki, T. (ed.) Image Fusion, pp. 223–250. Academic Press, Oxford (2008) 10. Baluja, S., Caruana, R.: Removing the Genetics from the Standard Genetic Algorithm. In: Prieditis, A., Russell, S. (eds.) Machine Learning Proceedings 1995, pp. 38–46. Morgan Kaufmann, San Francisco (CA) (1995) 11. Yadav, S.L., Sohal, A.: Comparative study of different selection techniques in genetic algorithm. Int. J. Eng. Sci. Math. 6(3), 174–180 (2017) 12. Mirjalili, S.: Genetic algorithm. In: Evolutionary Algorithms and Neural Networks: Theory and Applications, pp. 43–55. Springer, Cham (2019) 13. Sivanandam, S., Deepa, S.: Introduction to Genetic Algorithms. Springer, Heidelberg (2007)

350

T. Sabbah

14. Nazarahari, M., Khanmirza, E., Doostie, S.: Multi-objective multi-robot path planning in continuous environment using an enhanced genetic algorithm. Exp. Syst. Appl. 115, 106–120 (2019). https://doi.org/10.1016/j.eswa.2018.08.008 15. Qian, S., Liu, Y., Ye, Y., Xu, G.: An enhanced genetic algorithm for constrained knapsack problems in dynamic environments. Nat. Comput. 18(4), 913–932 (2019). https://doi.org/10. 1007/s11047-018-09725-3 16. Al-Shalabi, M., Anbar, M., Wan, T.-C., Alqattan, Z.: Energy efficient multi-hop path in wireless sensor networks using an enhanced genetic algorithm. Inf. Sci. 500, 259–273 (2019). https://doi.org/10.1016/j.ins.2019.05.094 17. Wan, J., Chu, P., Jiao, Y., Li, Y.: Improvement of machine learning enhanced genetic algorithm for nonlinear beam dynamics optimization. Nucl. Instrum. Meth. Phys. Res. Sect. A 946, 162683 (2019). https://doi.org/10.1016/j.nima.2019.162683 18. Anh, B.T.M., My, N.T., Trang, N.T.T.: Enhanced genetic algorithm for single document extractive summarization. In: 10th International Symposium on Information and Communication Technology, Hanoi, Ha Long Bay, Viet Nam 2019, pp. 370–376. Association for Computing Machinery (2019) 19. Saidi, R., Bouaguel, W., Essoussi, N.: Hybrid feature selection method based on the Genetic Algorithm and Pearson Correlation coefficient. In: Hassanien, A.E. (ed.) Machine Learning Paradigms: Theory and Application, pp. 3–24. Springer, Cham (2019) 20. Tsai, C.-F., Chen, Z.-Y., Ke, S.-W.: Evolutionary instance selection for text classification. J. Syst. Softw. 90, 104–113 (2014). https://doi.org/10.1016/j.jss.2013.12.034 21. Tan, F., Fu, X., Zhang, Y., Bourgeois, A.G.: A genetic algorithm-based method for feature subset selection. Soft. Comput. 12(2), 111–120 (2008). https://doi.org/10.1007/s00500-0070193-8 22. U˘guz, H.: A two-stage feature selection method for text categorization by using information gain, principal component analysis and genetic algorithm. Knowl. Based Syst. 24(7), 1024– 1032 (2011). https://doi.org/10.1016/j.knosys.2011.04.014 23. Gunal, S.: Hybrid feature selection for text classification. Turkish J. Electr. Eng. Comput. Sci. 20(2), 1296–1311 (2012) 24. Lei, S.: A feature selection method based on information gain and genetic algorithm. In: 2012 International Conference on Computer Science and Electronics Engineering 2012, pp. 355– 358. IEEE (2012) 25. Uysal, A.K., Gunal, S.: A novel probabilistic feature selection method for text classification. Knowl. Based Syst. 36, 226–235 (2012) 26. Afif, M.H., Ghareb, A.S., Saif, A., Bakar, A., Bazighifan, O.: Genetic algorithm rule based categorization method for textual data mining. Decis. Sci. Lett. 9(1), 37–50 (2020) 27. Sabbah, T., Selamat, A., Selamat, M.H., Ibrahim, R., Fujita, H.: Hybridized term-weighting method for Dark Web classification. Neurocomputing 173(Part 3), 1908–1926 (2016). https:// doi.org/10.1016/j.neucom.2015.09.063 28. Dua, D., Graff, C.: UCI Machine Learning Repository. University of California, School of Information and Computer Science, Irvine, CA, USA (2019)

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions for a Stand-Alone Hybrid Energy System Abbas Q. Mohammed1,2(B) , Kassim A. Al-Anbarri1 , and Rafid M. Hannun3 1 Faculty of Engineering, Electrical Engineering Department, Mustansiriyah University,

Baghdad, Iraq 2 Construction and Projects Department, University of Thi-Qar, Nassriyah, Thi-Qar, Iraq 3 Mechanical Engineering, College of Engineering, University of Thi-Qar,

Nassriyah, Thi-Qar, Iraq

Abstract. This article proposes a new algorithm called Nomadic People Optimizer (NPO) to find the optimal sizing of a hybrid energy system (HES), consisting of photovoltaic cell (PV), battery storage (BS), and diesel generator (DG). The HES supply the electricity to an academic building located in Thi-Qar Province, which is located in southern Iraq on latitude 31.06º and longitude 46.26º. The objectives of this algorithm are to reduce the total cost during the life cycle of the project, and this is an economic aspect that in turn reduces energy costs, the second goal is to reduce emissions of carbon dioxide. While continuing to supply the electrical load with electricity throughout the life cycle of the project for 25 years. The results show that optimal sizing of the HES achieved by 1875 number of the PV,687 number of the BS, and single DG. Keywords: Renewable energy · Solar energy · Nomadic people optimizer · Optimization

1 Introduction Electric power is one of the most sought after commodities of the human race. More than 70% of the world’s energy demand comes from fossil fuels burning, like crude oil, coal, carbon gas, and natural gas [1]. As the economies and world populations expand, energy demand rises, resulting in a rise in fossil fuel usage. Conventional fuel supplies are therefore limited and quickly depleting. Also, fossil fuels are responsible for pollution, including greenhouse gasses (GHGs) that lead to global warming [2]. Global energy demand is projected to grow by 56% from 2010 to 2040, rising CO2 emissions from 31.2 billion tons in 2010 to 45.5 billion tons in 2040. Moreover, in the coming decades, fossil-based oil, coal, and gas reserves will rapidly deplete [3]. These current and expected circumstances push scientists to pursue a strategy that involves improving energy-efficient systems [4, 5]. And replacing fossil-fuel power generation units with those that use renewable energy sources (RES) [6]. It is therefore important © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 351–363, 2021. https://doi.org/10.1007/978-3-030-70713-2_33

352

A. Q. Mohammed et al.

to use environmentally friendly energy sources for environmental improvement [7]. Clean energy sources such as solar, wind, hydro, and geothermal energy; are important in this sense since they are environmentally friendly [8]. However, for many reasons solar power may be an ideal choice in the future world for several reasons: firstly, the solar energy is the highest renewable source produced the electricity [9]. Studies have shown that global energy demand can be met satisfactorily with solar energy, since it is plentiful and readily available and have less the cost of energy [10]. Secondly, it is a promising global energy source because it is not exhaustible, offering high and increased production efficiencies than other energy sources [11]. Even if the RES is appealing with its environmentally friendly and easy to replenish solutions to the energy crises, they are unpredictable and cannot be anticipated. It is important to note that this issue can be alleviated by incorporating various the RES that will serve as a hybrid renewable energy system (HRES) and by integrating with other storage sources such as batteries, and the DGs to form the HES [12]. The design of HES is a very complex problem that has a high number of parameters, and thus classic design techniques can produce unsatisfactory results [13]. In some studies, the HOMER program is used to determine the HES configuration. The HOMER software, which optimizes HES, uses an enumerative technique in searching for the optimal design. The enumerative technique ensures the optimal solution possible, but an incredibly high Processor time can be needed. In recent years, the academic community, and the industry have been paying more attention to the optimization algorithms. These algorithms were applied to several problems and obtained incredibly good results [14]. The problem of these algorithms are the inability to balance between the local search and the global search, also have a control parameter that makes it more complicated. So, this study employs a new algorithm called “Nomadic People Optimizer (NPO)”, which relies on the pattern of the life of nomads. The NPO simulates the life pattern of the nomads during their search for sources of life (such as grass, and water for their animals). The algorithm also captures how the nomads have lived for several years, and how they have been continuously migrating from place to place in search of comfort. This algorithm has a peculiar ability to achieve the right balance between exploration and exploitation and does not rely on any control parameters to control the search process [15–18].

2 The Problem of Optimal Sizing 2.1 Construction of the Proposed HES The HES in this paper consists of four components (PV, BS, DG), in addition to the inverter that converts DC power to AC power, where the PV and BS are connected to the DC bus, while DG and the AC load connected to the AC bus. Fig. 1 shows a block diagram of the proposed HES.

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions

353

Fig. 1. Block diagram of the proposed HES

2.2 Data Collection The output power of the PV dependent on the solar radiation and the temperature, so it is data was collected from the weather forecast in Thi-Qar for every hour of the year. Thi-Qar Province is characterized by high solar radiation. The annual solar radiation rate that fell on the Thi-Qar Province from the years (1961–1991) is 4.92 kW/m2 /day [19]. Figure 2 and Fig. 3 shows the solar radiation and the temperature for the first ten days of July to clarify. The load data collected from Thi-Qar Electricity Distribution Directorate, where the peak load equal to 180 kW, an average load equal to 96.2688 kW. Fig. 4 shows the load demand in the first ten days of July to clarify.

Fig. 2. The solar radiation for the first ten days of (July)

354

A. Q. Mohammed et al.

Fig. 3. The air temperature for the first ten days of (July)

Fig. 4. The load demand for the first ten days of (July)

2.3 Modeling of the Proposed HES PV System Model By using Eq. 1, the output power of the PV can be calculated at a time (t) [20]   S × (1 + K(TPV − TPV−STC ) PPV (t) = Prat × DF × SSTC

(1)

Where: PPV (t) is the output power of the PV in time (t) [W], Prat is rated power of the PV [W], S is the solar radiation at a time (t) [W/m2 ], SSTC is the solar radiation at standard test conditions(STC) [W/m2 ], TPV is cell temperature [°C], TPV−STC is the cell temperature of a PV at SOC [°C], K is the temperature coefficient of the maximum power of the PV, and DF is a module derating factor.

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions

Cell temperature can determine by using Eq. 2 [21]:    NOCT − 20 TPV = Tair + S × 800

355

(2)

Where: NOCT is the normal operating cell temperature of the PV [°C], Tair is the air temperature [°C]. The datasheet of the PV that is used in this studying shown in Table 1. Table 1. PV datasheet Type

Poly Crystalline

SSTC

1000 W/m2

Prat

355 W

TPV −STC

25 ◦ C

K

−0.38%

Life cycle

25 years

DF

0.94%

Maintenance cost

0$

NOCT

45◦ C

Capital cost

220 $

BS System Model The BS is charged  only when there is a surplus of energy that generated by RESs  EEX/DE (t) > 0 , and the level of charge (LOC) of the BS is less than the maximum (LOC(t) < LOCmax ). During charging, the LOC of the BS at the time (t) is given by using Eq. 3 [12]:   ED(t) × effBS (3) LOC(t) = LOC(t − 1) × (1 − Sdisc ) + Eren (t) − effINV EEX/DE (t) = Eren (t) −

ED(t) effINV

(4)

Where: LOC is a level of charge the BS in a present hour (W), LOC(t −1) is a level of charge BS in a previous hour (W), Sdisc is the rate of self-discharge for the BS in time (t) (%), Eren (t) is the energy that generated the RESs in time (t) (W), eff BS is the efficiency of BS (%), eff INV is the efficiency of the inverter (%), LOCmax is the maximum state charge of BS (W), ED(t) is the load demand of the energy in time (t) (W), and EEX/DE (t) is a surplus or a deficit of the RESs in time (t) (Wh). The discharge of only when there is an energy deficit that   the BS is happening generated by RESs EEX/DE (t) < 0 and the LOC of the BS is greater than the minimum (LOC(t) > LOCmin ). During discharging, the LOC of the BS at a time (t) is given by using Eq. 5 [12]:   ED(t) (5) − Eren (t) LOC(t) = LOC(t − 1) × (1 − Sdisc ) + effINV Where: LOCmin is the minimum level of charge BS (W).

356

A. Q. Mohammed et al. Table 2. BS datasheet Type

12 V Monoblock

DOD

0.4 × CMax

CMax

2400 W

Sdisc

0.01488%/h

eff BS

>= 90%

Maintenance cost

0$

LOCMax

2400 W

Capital cost

250 $

LOCmin

960 W

The datasheet of the BS that used in this studying shown in Table 2. CMax is the nominal capacity of the BS (W), (DOD) is the maximum allowable discharge depth of the BS (%).The life cycle of the BS dependent on DOD. Table 3 shows the relation DOD of the BS with the life cycle of the BS. Table 3. The life cycle of the BS DOD

Life cycle

At 30%

3600 cycle

At 40%

2600 cycle

At 50%

2000 cycle

At 60%

1500 cycle

Ideal float condition 10 ears

Inverter Model The number of inverters required for the HES is calculated by using Eq. 6 [22] NINV =

PH−Max PINV−Max

(6)

Where: PG−Max is represent the maximum power generated by the components that are connected with the inverter (W), PInv−Max is the maximum power of the inverter (INV)(W). The datasheet of the INV that is used in this studying shown in Table 4. DG System Model The DG is needed to supply the load continuously if the energy provided by the RESs cannot meet demand and the LOC at a minimum. The DG is independent of climate for supplying power but their operation has harmful effects such as CO2 diffusion that pollutes the environment. Also, it has high costs for maintenance. The consumption of the fuel F(t) in liters/h of the DG is linked to the rated power for the DG, and the average output power of the DG in one hour can be calculated by using Eq. 7 [12]: F(t) = (0.246 × PADG ) + (0.08415 × PRDG )

(7)

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions

357

Table 4. INV datasheet Type

Bi-directional inverter

Life Span

25 years

PINV−Max

10000 W

Maintenance cost

20 $/year

effInv

>= 90%

Capital cost

3367 $

Where: F(t) is the consumption of the fuel in time (t) (l/h), PRDG is the rated power of the DG (kW). PADG is the average output power of the DG in time (t) (kW), 0.246, and 0.08415 are constant factors in l/kWh. In this study, the DG covers a deficit of the RESs without charge of the BS. In this study, the DG does not charge the BS. The datasheet of the DG that is used in this studying shown in Table 5. Table 5. DG datasheet Type

Perkins

Life span

15000 h

PRdg (Kw)

200 kW

Maintenance cost

0.309 $/h

PRdg (KVA)

250 A

Capital cost

29200 $

3 Nomadic People Optimizer (NPO) In this study, a new swarm-based metaheuristic (Nomadic People Optimizer (NPO)) is used to simulate the lifestyle of nomads as they travel in search of the life sources such as water, and the grass for their livestock. In this research, the creation of the NPO was influenced by the Bedouins and their lifestyles. The nomads are from the Sheik family and the rest of the families are considered normal. The Sheik as a clan leader is responsible for deciding where and when families can travel to ensure their safety, and the sheik also decides the pattern in which the ordinary families may be put around the Sheik’s house. The family tents are typically semicircularly distributed around the Sheik’s tent. The Sheik selects the families to find a new appropriate place; selected families are forced to travel randomly in various directions and distances in search of the best place to move. The nomads spend all their time traveling with their animals to find a better place to support their time [15]. 3.1 The Objective of the Proposed Algorithm The major purpose of the optimization is to find the optimal sizing of a stand-alone HES (PV system, DG system, and BS system) to minimize the total system cost (CT ), this is an economic aspect, which in turn reduces the  cost of energy (COE). The second objective is minimizing the total CO2 emissions ECO2 T , with continuous provide the load by

358

A. Q. Mohammed et al.

the electricity (the reliability as constrained) through the life cycle of the project for 25 years. A typical system configuration N is a row vector of three elements (n1 to n3 ), where each element represents the required number of subsystem components in the HES. The row vector X is represented by using Eq. 8: N = [n1 n2 n 3 ]

(8)

Where: n1 is the number of modules required for the PV,n2 is the number of modules required for the BS, n3 is the number of modules required for DG. The following will explain these objectives:

4 Total System Cost CT The CT is one of the NPO’s objectives; it represents the total system cost for 25 years. The implemented CT in this study considered the total cost of capital of components system (PV system, BS system, INV system, and the DG system), as well as the total replacement cost of the system component through the same period; the total maintenance cost of the system components for 25 years was also considered, together with the total cost of fuel of the DG through 25 years. The CT was calculated by using Eq. 9: CT = CTCap + CTRep + CTMaint + CTFuel

(9)

Where: CTCap is the total capital cost of the system components ($), CTRep is the total replacement cost of the system component ($), CTMaint is the total maintenance cost of the system components ($), CTFuel is the total fuel cost of the DG ($), through the life cycle of the project (25) years. Minimizing CT leads to minimizing the cost of energy (COE) that is calculated by using Eq. 10 [23]. CT COE = n 1 ED(t)

(10)

Where n is the life cycle of this study (219000 h), ED(t) is the load demand in time (t) (Wh).

5 Total CO2 Emissions (ECO2 T ) The (ECO2 T ) are the second objective of the (NPO). The emissions occur when there a deficit of the RESs, and (LOC) of the BS at a minimum LOC(t) = LOCmin , so will run the load with the DG. This is an undesirable condition because an increase in the CO2 emissions that lead to global warming. The emissions are related to its fuel consumption, where CO2 emission in time (t) calculated by using Eq. 11 [12]:   kg × F(t)(l/h). (11) ECO2 (t) = SECO2 l

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions

359

Where: (SE02 ) is the specific carbon dioxide emissions by liter of fuel are given as 2.7 kg/l., ECO2 (t) is the CO2 emission of the DG at a time (t) (kg). The CO2 emission of a DG throughout the lifetime of the project (25) years, it’s the sum of all CO2 emissions are given by using Eq. 12: n EC02 T = EC02 (t) (12) t=1

5.1 The Constraints The constraint in this study is (reliability) mean continues to provide the load by the electricity through the life cycle of the project and give by using Eq. 13: E T (t) ≥

ED(t) eff INV

(13)

Where ETotal (t) is the total output energy of the system at time t (Wh), and is mathematically given by using Eq. 14: E T (t) = n1 × EPV (t) + n2 × EBS (t) + n3 × EDG (t) × effINV

(14)

Where: EPV (t) is the output energy of the PV in time (t) (Wh), EBS (t) is the output energy by the BS in time (t)(Wh), and EDG (t) is the output energy by a DG in time (t)(Wh). In this study used the weighting sum method (Multi-Objective) to minimize (C T ), and (EC02 T ) by Eq. 15:





 (15) Q N i = C T N i × W1 + ECO2 T N i × W2 × COP Q(N i ) is an objective function that connected between (C T ), and(EC02 T ), (W 1 , W 2 )are the weights of (C T ), and EC02  T that used in this study to minimize the objectives together. (C T (N i )), and EC02 T N i are computed for each family (configuration) (N i ). The best families are families with the minimum C T , and EC02 T . The COP is the penalty factor for CO2 emission conversion to monetary value. In this study relied on Sweden’s emissions tax which is $150/ton, Sweden is among the countries with the highest carbon taxes in comparison to the other countries. The Flow chart of the optimization process shown in the Fig. 5, energy sources selection block shown in Fig. 6

6 Results and Discussions A new algorithm is used in this study NPO to determine the optimal size of the system, consisting of the energy supply systems (PV system, DG system, and BS system) with the number of inverters depend on the maximum electric produced by the HES, that connected with the INV and the maximum capacity of the INV to reduce the C T ,and T with continuous provide the load by the electricity for 25 years. In this study ECO2 number of the iterations was 50,LOCMax of 2400 W, LOCmin of (0.4×LOCMax ), (DOD) of (0.6 × LOCMax ).

360

A. Q. Mohammed et al.

Fig. 5. Flow chart of the optimization process

Fig. 6. Block Energy Sources Selection

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions

361

6.1 Optimal Configuration The optimal sizing of the algorithm, when used each the components (PV-BS-DG) is N = [1875 687 1], comprising of 1875 number of the PV, 687 number of the BS, and single big DG. This configuration was implemented with many inverters (46). In this configuration, DG run when (LOC) at a minimum and (EEX/DE < 0), without charge of the batteries. The behavior of the system for the first ten days of the summer season (July), shown in Fig. 7. Table 6 showed the PV-BS-DG system configuration performance for 25 years. Table 6. a PV-BS-DG configuration performance for (25) years. Configuration (N) [1875 687 1] CT ($)

1.9559 × 106

T ECO2 (kg)

3.4887 × 106

COE ($/kWh)

0.092771926

From Table 6 notes that the NPO succeed in minimizing multi-objective function include the total system cost CT , and the total CO2 emissions ET CO2.

Fig. 7. The behavior of a PV-BS- DG system configuration for the first ten days of (July)

7 Conclusions This study presented the use of a new multi-objective optimization model by using a new algorithm called Nomadic People Optimizer (NPO) to find optimal sizing of the HES comprised of the (PV systems, BS system, and the DG) to minimizing multiobjective T .With continuous function. The first objective is the CT , the second objective is EC02 provide the load demand by the electricity through the life cycle of the project 25 years as a constraint. Based on the results of this study, it is concluded as follows:

362

A. Q. Mohammed et al.

1. The optimal HES configuration was comprised of a (PV-BS-DG) system. Where T ) is (1875) number optimum sizing that gives minimum (CT ), and minimum (ECO2 of the PV, (687) number of the BS, and single big DG. 2. The NPO could accept different inputs, like the air temperature, solar irradiation, and user load demand data for developing an optimal sizing of the HES in this study. The T ) with continuous NPO also succeeds in minimize the objectives (CT , and the ECO2 provide the load by the electricity for 25 years. In the future studies require the addition of other RESs to the HES such as (wind), and also use the DG to charge the BS.

References 1. Kuang, Y., Zhang, Y., Zhou, B., Li, C., Cao, Y., Li, L., et al.: A review of renewable energy utilization in islands. Renew. Sustain. Energy Rev. 59, 504–513 (2016) 2. Sinha, S., Chandel, S.S.: Review of recent trends in optimization techniques for solar photovoltaic–wind based hybrid energy systems. Renew. Sustain Energy. Rev. 50, 755–769 (2015) 3. Iqbal, M., Azam, M., Naeem, M., Khwaja, A.S., Anpalagan, A.: Optimization classification, algorithms and tools for renewable energy: a review. Renew. Sustain. Energy Rev. 39, 640–654 (2014) 4. Evins, R.: A review of computational optimisation methods applied to sustainable building design. Renew. Sustain. Energy Rev. 22, 230–245 (2013) 5. Stevanovi´c, S.: Optimization of passive solar design strategies: a review. Renew. Sustain. Energy Rev. 25, 177–196 (2013) 6. Sharma, N., Bhat, I.K., Grover, D.: Optimization of a smooth flat plate solar air heater using stochastic iterative perturbation technique. Sol. Energy 85(9), 2331–2337 (2011) 7. Alanne, K., Saari, A.: Distributed energy generation and sustainable development. Renew. Sustain. Energy Rev. 10(6), 539–558 (2006) 8. Kannan, N., Vakeesan, N.: Solar energy for future world: a review. Renew. Sustain. Energy Rev. 62, 1092–1105 (2016) 9. Panwar, N.L., Kaushik, S.C., Kothari, S.: Role of renewable energy sources in environmental protection: A review. Renew. Sustain. Energy Rev. 15(3), 1513–1524 (2011) 10. Lewis, N.S.: Toward cost-effective solar energy use. Science 315(5813), 798–801 (2007) 11. Nozik, A.J.: Photo electrochemistry: applications to solar energy conversion. Annu. Rev. Phys. Chem. 29(1), 189–222 (1978) 12. Ogunjuyigbe, A.S.O., Ayodele, T.R., Akinola, O.A.: Optimal allocation and sizing of PV/Wind/Split-diesel/Battery hybrid energy system for minimizing life cycle cost, carbon emission and dump energy of remote residential building. Appl. Energy 171, 153–171 (2016) 13. Dufo-López, R., Bernal-Agustín, J.L.: Influence of mathematical models in design of PVDiesel systems. Energy Convers. Manage. 49(4), 820–831 (2008) 14. Sedghi, M., Hannani, S.K.: Modeling and optimizing of photovoltaic-wind-diesel hybrid systems for electriication of remote villages in Iran. Scientia Iranica B 23(4), 1719–1730 (2016) 15. Salih, S.Q., Alsewari, A.A.: A new algorithm for normal and large-scale optimization problems: nomadic people optimizer. Neural Comput. Appl. 32(14), 10359–10386 (2020) 16. Tao, H., Salih, S.Q., Saggi, M.K., Dodangeh, E., Voyant, C., Al-Ansari, N., Yaseen, Z.M., Shahid, S.: A newly developed integrative bio-inspired artificial intelligence model for wind speed prediction. IEEE Access 8, 83347–83358 (2020)

Multi-objective NPO Minimizing the Total Cost and CO2 Emissions

363

17. Abdulwahab, H.A., Noraziah, A., Alsewari, A.A., Salih, S.Q.: An enhanced version of black hole algorithm via levy flight for optimization and data clustering problems IEEE. Access 7, 142085–142096 (2019) 18. Salih, S.Q., Alsewari, A.A., Al-Khateeb, B., Zolkipli, M.F.: Novel multi-swarm approach for balancing exploration and exploitation in particle swarm optimization. In: Saeed, F., Gazem, N., Mohammed, F., Busalim, A. (eds.) Recent Trends in Data Science and Soft Computing: Proceedings of the 3rd International Conference of Reliable Information and Communication Technology (IRICT 2018), pp. 196–206. Springer International Publishing, Cham (2019). https://doi.org/10.1007/978-3-319-99007-1_19 19. Al-Khazzar, A.A.A., Khaled, A.J.: A comparative study of the available measured global solar radiation in Iraq. J. Renew. Energy Environ. 4, 47–55 (2017) 20. Suresh, M., Meenakumari, R.: An improved genetic algorithm-based optimal sizing of solar photovoltaic/wind turbine generator/diesel generator/battery connected hybrid energy systems for standalone applications. Int. J. Ambient Energy, 1–8, (2019). 21. Javed, M.S., Song, A., Ma, T.: Techno-economic assessment of a stand-alone hybrid solarwind-battery system for a remote island using genetic algorithm. Energy 176, 704–717 (2019) 22. Hassan, A., El-Saadawi, M., Kandil, M., Saeed, M.: Modeling and optimization of a hybrid power system supplying RO water desalination plant considering CO2 emissions. Desalination Water Treatment 57(26), 11972–11987 (2016) 23. Sedghi, M., Kazemzadeh Hannani, S.: Modeling and optimizing of PV–wind–diesel hybrid systems for electrification of remote villages in Iran. Scientiairanica 23(4), 1719–1730 (2016)

A Real Time Flood Detection System Based on Machine Learning Algorithms Abdirahman Osman Hashi1,3(B) , Abdullahi Ahmed Abdirahman1 , Mohamed Abdirahman Elmi1 , and Siti Zaiton Mohd Hashim2 1 Faculty of Computing, SIMAD University, Mogadishu, Somalia

{aaayare,m.abdirahman}@simad.edu.so 2 Faculty of Computing„ Department of Artificail Intelligence and Big Data, Universiti

Malaysia Kelantan, 16100 Pengkalan Chepa, Kelantan, Malaysia [email protected] 3 Faculty of Informatics„ Department of Computer Science, Istanbul Teknik Üniversitesi, 34469 Masklak, ˙Istanbul, Turkey

Abstract. Flood is expressed as water overflowing onto the ground that usually is dry or an increase of water that has a significant impact on human life and it is also declared as one of the most usually natural phenomenon, causing severe financial crisis to goods and properties as well as affecting human lives. However, preventing such floods would be useful to the inhabitants in order to get a sufficient time to evacuate in the areas that might be possible floods can happen before the actual floods happen. To address the issue of floods, many scholars’ proposed different solutions such as developing prediction models and building a proper infrastructure. Nevertheless, these proposed solutions are not efficient from an economic perspective in here, Somalia. Therefore, the key objective of this research paper is to intend a new robust model which is a real-time flood detection system based on Machine-Learning-algorithms; Random-Forest, Naïve-Bayes and J48 that can detect water level and measure floods with possible humanitarian consequences before they occur. The experimental results of this proposed method will be the solution of forth mentioned problems and conduct research on how it can be easily simulate a novel way that detects water levels using hybrid model based on Arduino with GSM modems. Based on the analysis, Random-Forest-algorithm were outperformed other machine-learning-methods in-terms of accuracy over other-classification with 98.7% accuracy in-comparison with 88.4% and 84.2% for NaiveBayes and J48 respectively. The proposed method has contribution to the field of study by introducing a new way of preventing floods in the field of Artificial, data mining. Keywords: Machine learning · Naive Bayes · Random Forest · Artificial intelligence · Data mining

1 Introduction It’s well known that the ageing of natural disaster cannot be escaped however prealarming systems and proper managing can mitigate its severity to tackle this case. Most © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 364–373, 2021. https://doi.org/10.1007/978-3-030-70713-2_34

A Real Time Flood Detection System

365

of the developed countries, meteorological department has flood-monitoring cell that may not appropriately equipped with intelligent system and scalable flood alarming system or some countries may not have that department including our country, Somalia. As a result, people from areas that flood affected in the prone-areas are dealing with the results of the flood every year [1]. In Somalia, the dangerous flash floods occurred in Beledweine town of Hiran region last year had reported over 100,000 people have been displaced [2]. Consequence to that also, River-flooding has so far an impact on an estimated 620,000 people in Somalia as OCHA stated. More than around 213,800 of these people have been displaced and fled from their houses consequences of the heavy rains that happened in Ethiopia that is received across the country to be affected, especially in southern regions that is also Hiiran Region is among them, according to the UNHCR-led Protection [3]. Increased rainwater since the beginning of May last year (2019) has stated in a sharp rise in water levels in Jubba and Shabelle rivers as result and this might lead to severe flooding in central and southern regions of Somalia. According to the UNHCR-led, the flood magnitude that occurred in Baladweyn last year is reported as the highest water levels in history that occurred in that region as well as the whole regions in the country. Moreover, according to data collected by humanitarian partners resulted that the current flood levels exceed a 50 years return period as it has been affected more than 427,000 people and of these nearly 174,000 have been fled in their homes as a result of the flash and river-flooding that occurred in Hirshabelle state [2]. According to (CMHC), these floods can occur at any time of the year and are most often caused by heavy rainfall that may happen in Ethiopia that would cause then to raise the level of Shabelle and Jubba rivers. Consequences to that would be many people to evacuate and lost their houses. Hence in recent years, due to the rapid advent of communication technologies, Global Positioning System know as (GPS) that equipped with wireless devices and GSMs have been broadly deployed on various public and private positions, generating huge amount of data that could be implemented to measure water levels, locations, and so on, for fleet management [4]. In order to predict & detect flood, Machine learning applications can give valuable solutions to tackle this phenomenon case. Moreover, it is another inevitable job to resist devastation’s flood if there is possible method to inform population living around the area through appropriate and properly way in real-time [1, 8]. To date, detecting variations of water-levels in various of flood zone is widely utilized sensor technologies to share data with inhabitants [9]. The purpose of this paper is to simulate a real-time Flood Detection System based on Machine Learning that can detect water level and measure floods with possible humanitarian consequences before they occur. This paper is structured with five sections. The following section provides a background and related work of flood detection methods. The third section describes the methodology in which this framework is to be implemented. The fourth section presents the experiment design of the proposed framework. Finally, the fifth section presents a result analyze, and its conclusion.

366

A. O. Hashi et al.

2 Background and Related Work There are many natural disasters around the globe, however floods are known that they are the most critical, triggering huge damage to the human-life, infrastructure and agriculture [2, 4]. Hence there must the use of some sort of machine learning algorithm. Machine learning is one of the prominent fields in artificial intelligence that came from the improvement of self-learning algorithms to get knowledge from that data so as to create the forecasts. In these days, the data are huge, and these data can be can be converted into knowledge by using an algorithm which is the field of machine learning [5]. Machine learning gives a good effective option for taking the knowledge into data to increasingly rise the forecast models’ performance and create decisions that came from that data. Hence, the meaning of this research is if we desire to forecast the level of the river in a particular place we can use a special ML algorithm with our past data and if it is successfully recognized it, then it will do better prediction for future water levels [8]. Artificial-neural-networks, neuro-fuzzy are among the numerous ML algorithms that were stated as effective in term of short and long for flood prediction and the following subsequence explains each of these algortihms. 2.1 Artificial-Neural-Networks (ANNs) Artificial-neural-networks are systems that have numerical model with a successful proficient parallel processing. Enabling them to imitate the utilization between neural units and the biological neural network [5]. Among all ML-methods, ANNs are the most important popular learning algorithm, known to be easily changed and effective inmodeling complex flood processes and it has a tolerance with a high fault also it brings an accurate approximation [6]. If we compare convention statistical model to ANNs, ANN approach was utilized with greater accuracy for the help of predictions. Since their first time usage ANN In the 1990s, this algorithms is the most essential prevalent method for flood prediction [7]. 2.2 Adaptive-Neuro-Fuzzy Inference System (ANFIS) The fuzzy-logic of Zadeh [4] could be some soft computing technique with a qualitatively model technique using natural-language. It is also know that Fuzzy-logic is a basic mathematical-model for calculation, which works on consolidating expert knowledge into a fuzzy-inference system to able classification of different date. An FIS another play actor for human-learning through an prediction function with less complexity of computations, which gives good ability for nonlinear-modeling of extreme hydrological events [6], especially flood ones. 2.3 Decision Tree (DT) and Ensemble-Prediction-Systems (EPSs) The machine language strategy of decision tree is one of modeling predicative for suppliers with a thick application in stimulation-of-flood [8]. Decision tree uses branches from tree of decisions that is high precision to the leaves those are the target ones. In

A Real Time Flood Detection System

367

classification-trees (CT), the last factorsin a decision-tree have a separate set of values where leaves stand-for class-labels and branches on behalf of conjunctions of featureslabels [10, 15]. Meanwhile, a lot of language simulating machine alternatives were showed flood simulating model having a very tough background [9, 12]. Hence, there is a developing approach to vary from single form of prediction to an ensembleof models which is fit for not many applications, cost, dataset [13, 16].

3 Proposed Methodology As the research methodology gives a structure overview on the sequences of the follows. The overall framework of this research work will be in different phases either it is hardware or it is software development. It is known that a successful early forecasting and flood warning system will benefit to the population as it acts as a first stage of initialaction for the victims in-terms of human effect suffering and infrastructures. While SMS is an appropriate alert announcement tool that can distribute the data to the flood-victims within particular area. Hence, the first phase is to find out and select of scholarly information to acquire the adequate knowledge required to carry out this research. The main source of information and knowledge in this phase would be observing the river and data gathering from river side community. An example of the adequate knowledge is asking the river side community for the best place in the river that would be good for implementing this proposed architecture. Second phase is going to be the implementation phase. A water level sensor would be putting in the river in order to get and send dynamic real-time-data to the flood-controldevice for mining purposes about the data. This sensor device has it is own function. It will detect the water level that could be normal, above normal or a dangerous condition. After data collected and converted from analog to digital, Machine learning algorithms will be trained to decide if there is a critical condition or not. Random Forest, Naive Bayes and J48 are the machine algorithms that would do the classification base on their accoury. Last Phase will be Data sharing phase. Data that has mined from different algorithms will be transmitted to the core control unit (microcontroller PIC). PIC obtains the one has high accurate in term of their classifiers. After the high accuracy data obtained, the data can be monitored and controlled from anyplace in the region that is available GSM service. 3.1 Experimental Design The proposed framework architecture is designed to be a hybrid model based on machinelearning-algorithms with sort of hardware devices that would be able to detect water levels. Data collected from water sensor would be transmitted to the main controller which is (PIC Microcontroller).

368

A. O. Hashi et al.

Software In order to develop the forthmentioned system we decided to use a java Programming language using Audio Platform and Weka for data mining. Java is a general-purposeprogramming language that has a few implementation dependencies as possible. It is intended to let application developers write once, run anywhere. Arduino Software was used as our IDE and is a cross-platform-application that is written in the programming-language Java itself. It is used to write and upload programs to Arduino compatible boards, but also, with the help of 3rd party cores, other vendor development boards. Hardware In order to implement the system, a number of hardware devices were used. First GSMmodem and is a specialized type of modem which accepts a SIM-card, and operates over a subscription to a mobile operator, just like a mobile phone used. When a GSM-modem is connected to a computer, this allows the computer to use the GSM-modem to communicate over the mobile network. We most frequently used for sending and receiving SMS messages. Secondly, water sensor is used which is an electronic device that is designed to detect the presence of water for purposes such as to provide an alert in time to allow the prevention of water damage. PIC16F877A is an Integrated-Circuit (IC) embedded in a single chip and act as a voltage level converter. PIC16F877A is capable of converting 5V TTL Logic level to TIA/EIA-232-F level and can take up to ±30V input. It is normally used for the communication between microcontroller and Laptop/PC.

Fig. 1. Proposed framework for flood detection

Figure1 demonstrates the whole process guide for software development and hardware of the proposed-design-architecture. One water-level-sensors have been epitomized in order to offer a real-time information to the flood-control-center for processing dedications to be used lately. That sensor has special tasks. It will detect the normal level of water signals, while it transfer data to the Microcontroller and then machine learning algorithms will decide the decision of this data. Finally if there is a dangerous situation, SMS message will come from the GSM SIM900.

A Real Time Flood Detection System

369

4 Results and Discussion The anticipated model which is based on detecting the water level and training three machine-learning-algorithms to measure the accuracy of the water level those are Random-Forest-algorithm, NaiveBayes algorithm and J48 has implemented. As we mentioned, generally there are two key works done in this research: First using the Arduino and GSM devices to detect as a real time from the river and collecting these data as a dataset. Whereas second step to mining the collected data and be training these three selected algorithms in order to know the water level accuracy that is improving the accuracy performance of the flood detection methods. In the following section, the experiment results and analysis is discussed to contain all the forementioned key components: Arduino with GSM has successfully proved its essential role in generating a good data collection tool. With the proper water level parameters setting, it succeeds to achive a better-accuracy than the ordinary solution, such as Sendo-sensor. The upcoming figures are the results that obtained from the algorithms. Each table or figure is followed with additional description.

Fig. 2. Normal Water Level

As Fig. 2 demonstrates that the water level is a normal, however is it is also essential to take consideration that flood warning is increasing (Red colour) while normal water is opposide to that and dramaticily decreasing (Blue colour). Meanwhile, the three algorithms has no much different values in term of correct classified instances and incorrect classified instances while the water level is normal.

370

A. O. Hashi et al.

Fig. 3. Flood Water Level

However, as Fig. 3 shows us that water level is increased rapidly; we observed that the three algorithms has different values in term of of correct classified instances and incorrect classified instances. Based on the analyze we observed that the three machine algorithms have different variations and the upcoming table will demonstrate classifier output for their classifications as the upcoming tables will illustrate. Table 1. Detected water level in term of the three algorithms Parameters

Methods Random Forest NaiveBayes J48

Correctly classified instances 98.7%

88.4%

84.2%

Incorrect classified instance

22.8%

2.8%

2.9%

Root mean squared error

0.0904

0.1387

0.1970

Total number of instances

1000

1000

1000

Table 1, needs to illustrate that our three experiments was conducted with imbalanced data that is the actual data that obtained from real flood from the Aurdinu device. The Random Forest corrected classified Instance gained 98.7% using, whereas instance of incorrect classified is 22.8%. Secondly, NaiveBayes algorithm demonstrated betterresult than J48 algorithm. The accured classified illustrated 88.4%, while, Instances

A Real Time Flood Detection System

371

of incorrect-classified for this algorithm were 2.8%. Moreover, 84.2% of the correctclassified-Instances were achieved by applying J48 algorithm, whereas 6.9% incorrectclassified. The best result has achieved Random Forest compared other classifications with 98.7% for correct-classified-Instances, while, incorrect-classified-Instances indicated only 22.8% as mentioned before. Table 2. Detected water level in term of accuracy by class Method

Accuracy by class True positive

True negative

Recall

Random Forest

0.989

0.004

0.976

NativeBayes

0.886

0.014

0.888

J48

0.842

0.021

0.842

Table 2 demonstrates the precision by class of-classifications. In order to increase information benefit from the data collected from water levels, we trained a number of machine learning methods to ascertain the appropriate techniques that could able to produce great performance and also accuracy. Random Forest algorithm was achieved high performance then other methods. Because of that Random Forest has gotten the highest True Positive which is 0.989% whereas J48 got the lowest one which is 0.842%. Meanwhile, in term of True negative; Ramdom forest is achived the lowest negative which 0.004% whereas J48 achieved the highest value for 0.021%. In order to avoid over-fitting issues and generating easy to set constraints, Random Forest can deal with supervised learning algorithms and utilize a huge number of decision-tree-models. Using this model will provide help and supports to those are living around the river areas that always face many circumstances that are coming from the rivers such as flooding.

5 Conclusions Flood-detection systems have been developed for immediate response to the authority people before it happens. It will inform you the state of the current water-level by using Aurdino-sensor-network, which will then provide SMS notification if there is a dangerous situation through GSM modem. Three machine-learning-algorithms were tested to classify data. It is an neglectable that Random-Forest-algorithm were outperformed other machine-learning-methods in-terms of accuracy over other-classification with 98.7% accuracy comparerd with 88.4% using NaiveBayes algorithm that plays an essential role. Furthermore, J48 achieved 84.2% accuracy close to the NaiveBayes, however it is slightly lower than that algorithm.. However, this proposed method can be further-improved or enhanced to achieve-todo more advanced technology and well applications that is capable for data mining in the next phase of research. For future enhancement, this proposed architecture can be advanced by adding by Video surveillance and GPS-module to track-the-equipment that

372

A. O. Hashi et al.

installed in different-areas. Finally, clustering algorithms can be applied for machinelearning-algorithm in order to improve the results of proposed model. Acknowledgments. The authors would like to express their cordial thanks to SIMAD University, for the Research University Grant no. 15. The authors would also like to acknowledge grateful to SIMAD Research Center for their support and making this research a success.

References 1. Khalaf, M., Hussain, A.J., Al-Jumeily, D., Fergus, P., Idowu, I.O.: Advance flood detection and notification system based on sensor technology and machine learning algorithm. In: 2015 International Conference on Systems, Signals and Image Processing (IWSSIP), pp. 105–108. IEEE (2015) 2. Halbeeg: 100,000 People Displaced by Floods in Beledweyne, 27 April 2017. https://en. halbeeg.com: https://en.halbeeg.com/2018/04/27/100000-people-displaced-by-floods-in-bel edweyneofficial/ 3. OCHA: Floods: Response plan. Humanitarian Country Team and partners. K. Elissa, “Title of paper if known” (2018, unpublished). 4. Baydargil, H.B., Serdaroglu, S., Park, J.S., Park, K.H., Shin, H.S.: Flood detection and control using deep convolutional encoder-decoder architecture. In: 2018 International Conference on Information and Communication Technology Robotics (ICT-ROBOT), pp. 1–3. IEEE (September 2018) 5. Mosavi, A., Rabczuk, T., Varkonyi-Koczy, A.R.: Reviewing the novel machine learning tools for materials design. In: Luca, D., Sirghi, L., Costin, C. (eds.) Recent Advances in Technology Research and Education, pp. 50–58. Springer International Publishing, Cham (2018) 6. Li, L., Xu, H., Chen, X., Simonovic, S.: Streamflow forecast and reservoir operation performance assessment under climate change. Water Resour. Manag. 24, 83 (2010) 7. Wu, C., Chau, K.-W.: Data-driven models for monthly streamflow time series prediction. Eng. Appl. Artif. Intell. 23, 1350–1367 (2010) 8. Zadeh, L.A.: Soft computing and fuzzy logic. In: Fuzzy Sets, Fuzzy Logic, and Fuzzy Systems: Selected Papers by Lotfi a Zadeh, pp. 796–804. World Scientific, Singapore (1996) 9. Choubin, B., Khalighi-Sigaroodi, S., Malekian, A., Ahmad, S., Attarod, P.: Drought forecasting in a semi-arid watershed using climate signals: a neuro-fuzzy modeling approach. J. Mt. Sci. 11, 1593–1605 (2014) 10. Choubin, B., Khalighi-Sigaroodi, S., Malekian, A., Ki¸si, Ö.: Multiple linear regression, multi-layer perceptron network and adaptive neuro-fuzzy inference system for forecasting precipitation based on large-scale climate signals. Hydrol. Sci. J. 61, 1001–1009 (2016) 11. Dineva, A., Várkonyi-Kóczy, A.R., Tar, J.K.: Fuzzy expert system for automatic wavelet shrinkage procedure selection for noise suppression. In: Proceedings of the 2014 IEEE 18th International Conference on Intelligent Engineering Systems (INES), Tihany, Hungary, 3–5 July 2014, pp. 163–168 (2014) 12. Hashi, A.O., Hashim, S.Z.M., Anwar, T., Ahmed, A.: A robust hybrid model based on KalmanSVM for bus arrival time prediction. In: Saeed, F., Mohammed, F., Gazem, N. (eds.) Emerging Trends in Intelligent Computing and Informatics: Data Science, Intelligent Information Systems and Smart Computing, pp. 511–519. Springer International Publishing, Cham (2020) 13. Tiwari, M.K., Chatterjee, C.: Development of an accurate and reliable hourly flood forecasting model using wavelet–bootstrap–ANN (WBANN) hybrid approach. J. Hydrol. 394, 458–470 (2010)

A Real Time Flood Detection System

373

14. Amir Mosavi, K.-W.C.: Review flood prediction using machine learning models. Water 2018, 1–41 (2018) 15. Hameed, S.S., et al.: Filter-wrapper combination and embedded feature selection for gene expression data. Int. J. Adv. Soft Comput. Appl. 10(1), 90–105 (2018) 16. Sajedi-Hosseini, F., Malekian, A., Choubin, B., Rahmati, O., Cipullo, S., Coulon, F., Pradhan, B.: A novel machine learning-based approach for the risk assessment of nitrate groundwater contamination. Sci. Total Environ. 644, 954–962 (2018)

Extracting Semantic Concepts and Relations from Scientific Publications by Using Deep Learning Fatima N. AL-Aswadi1,2 , Huah Yong Chan1(B) , and Keng Hoon Gan1 1 School of Computer Sciences, Universiti Sains Malaysia, 11800

Gelugor, Pulau Pinang, Malaysia [email protected], {hychan,khgan}@usm.my 2 Faculty of Computer Sciences and Engineering, Hodeidah University, Hodeidah, Yemen

Abstract. With the large volume of unstructured data that increases constantly on the web, the motivation of representing the knowledge in this data in the machineunderstandable form is increased. Ontology is one of the major cornerstones of representing the information in a more meaningful way on the semantic Web. The current ontology repositories are quite limited either for their scope or for currentness. In addition, the current ontology extraction systems have many shortcomings and drawbacks, such as using a small dataset, depending on a large amount predefined patterns to extract semantic relations, and extracting a very few types of relations. The aim of this paper is to introduce a proposal of automatically extracting semantic concepts and relations from scientific publications. This paper introduces a novel relevance measurement for concepts, and it suggests new types of semantic relations. Also, it points out of using deep learning (DL) models for semantic relation extraction. Keywords: Concept extraction · Deep learning · Ontology construction · Relevance measurements · Semantic relation discovery

1 Introduction The substantial growth of unstructured data makes manually ontology construction a hard and laborious task as well as it is time-consuming. This unstructured data contains much useful knowledge, but unfortunately, this knowledge is not in the machineunderstandable form, it is just in a human-understandable form [1, 2]. Therefore, constructing the ontologies is considered an important task to make this data in the machineunderstandable form as well as human-understandable form. The ontology is a data model to represent a set of concepts and the relationships among those concepts within a domain [1]. Many applications, such as Automated Fraud Detection, Semantic Searching, Decision-Support and Question-Answering (QA) systems are built based on ontologies [3–6]. Most of the recent research direct their efforts towards using ontologies because © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 374–383, 2021. https://doi.org/10.1007/978-3-030-70713-2_35

Extracting Semantic Concepts and Relations from Scientific Publications

375

these systems rely on results of knowledge modelling [7]. Thus, by using ontologies, the query or scenario construction, as well as the inferencing, are enriched [5, 7]. There are many existing ontologies repositories or tools that are constructed or seek to construct ontologies either manually, cooperatively or automatically, for example, WordNet, which is considered one of the oldest and most popular ontology repositories. It is a high accuracy resource that was manually constructed by linguists. However, the progress of WordNet is quite slow comparing with streaming data across the Web, as well as it lacks many modern terms, such as Covid-19, cloud computing, deep learning or even netbook [8]. Another example of ontology repositories is YAGO (Yet Another Great Ontology) [9], it is an ontology that built on top of both WordNet and Wikipedia. YAGO uses the Wikipedia category pages rather than using information extraction methods to leverage the knowledge of Wikipedia. However, Wikipedia categories are often quite fuzzy and irregular [8] (it does not follow the expected pattern and it is open edit for anyone), that is considered one of the disadvantages of YAGO repository. Also, YAGO uses structured data for building the ontology which might result to waste the space if not all arguments of n-array facts are known. In addition, YAGO is relatively little help if WordNet neither contains some of the related concepts [8]. There are many other ontology repositories that their ontologies were extracted from structured contents of Wikipedia pages such as Freebase, BabelNet, and DBpedia. On the other hand, many ontology extraction tools try to extract and construct the ontology either cooperatively or automatically, such as Text-to-Onto [10], SYNDIKATE (SYnthesis of DIstributed Knowledge Acquired from TExts) [11, 12], CRCTOL (Concept-Relation-Concept Tuple based Ontology Learning) [13], and ProMine [14]. Some of them using structured or semi-structured data as input to extract and construct the ontologies such as ProMine and Text-to-Onto, while others using unstructured data to extract the ontologies such as SYNDIKATE and CRCTOL. However, most of the existing ontology extraction tools have many shortcomings and drawbacks. For example, some of them depend on human intervention in the whole of their tasks such as Text-to-Onto. In addition, most of them, such as Text-to-Onto and CRCTOL, depend on predefined templates for relation extraction that lead to very low recall results [2, 15, 16]. Moreover, some of these tools use small dataset such as Text-to-Onto which used only 21 web articles as the input dataset. Nowadays, many researches that try to extract ontologies from scientific publications have begun to emerge, such as in [17, 18]. The [17] study is an association rules-based approach for enriching the domain ontology rather than extracting new domain ontology. This study depends partially on lexical similarity measures, but in many cases, there is no correlation between the lexical similarity of concept names and the semantic concept similarity because of the high complexity of language or the uncoordinated ontology development. An example of this shortcoming, the concepts pair (table, stable) has lexical similarity while there is not semantically matching. In the [18] study, the authors defined NTNU system that aims to extract the keyphrases and relations from scientific publications using multiple conditional random fields

376

F. N. AL-Aswadi et al.

(CRFs), this study has many limitations and shortcomings as the author stated themselves. One main limitation of these limitations is that this study extract only two types of relations they are synonym and hyponym. In addition, the authors stated that their multiple CRF models with the help of rules have improved the performance on the development set, but the performance was worse on the testing set. Besides all of the above, with continuous scientific development, new fields and terms constantly appear such as Covid-19 in the medical domain or deep learning in the IT domain. So there is a serious need to develop a new technique that can automatically extract and construct ontologies that represent the knowledge. This paper gives a proposal of automatically extracting the semantic concepts and relations from scientific publications by using DL. The rest of this paper is organized as the following: Section 2 gives a look at the ontology construction challenges. Section 3 explores the Deep Belief Network (DBN), while Sect. 4 presents the proposed work. Finally, we concluded in Sect. 5.

2 Ontology Construction Challenges Ontology construction process may conduct by one of the three ways: manual construction (fully performed by experts), cooperative construction (most or all ontology construction tasks are supervised by experts), and automatic construction (automatically performed with limited intervention by users or experts). The main two tasks of the ontology constructing process are extracting the concepts, as well as extracting and mapping the relationship between these concepts. Getting a high degree of precision and recall for these extracted relationships means getting a high degree of precision and reliability of the constructed ontology. The four main drawbacks and shortcomings in the most existing ontology construction research, which is considered the main challenges for automatic ontology construction, are: 1. Most of them not use the efficient relevant measurement to avoid noisy data such as [10, 13] 2. Most of them depend on large amount predefined patterns such as in [8, 13, 17]. 3. Many of them use a small dataset for constructing the ontologies such as in [10, 13] 4. Most of them extract very limited relations almost do not exceed synonym, hyponym, hypernym, meronyms, and/or holonyms relations such as in [8, 13, 14, 18]. The validity of these challenges is discussed and evidenced on our previous work [19] that presents and discusses in details the approaches, prominent systems of ontology construction and their challenges.

3 Deep Belief Network DL is a branch of neural networks (NNs), the difference between traditional NN and DL is in their architectures. NN have shallow architectures (one hidden layer); while DL has deep architectures (more than one hidden layer) and every hidden layer learns

Extracting Semantic Concepts and Relations from Scientific Publications

377

a new extracted features (concepts or relations) from the previous layer. The shallow architectures can effectively solve many simple, well-constrained or defined problems, but their modelling and representational power are limited [20]. Hence, for more complicated real-world applications such as human speech and natural language understanding, where we do not have enough predefined patterns or where we do not have a clear perception of problems, the deep architectures have more abilities when dealing with these complicated problems rather than shallow architectures [20]. As well as DL can handle a large amount of data in an effective and efficient way. Deep Belief Network (DBN) (with its respective variations) is one of the milestones models on the DL [21–23]. It is a multi-layer, unsupervised or supervised, and feedforward architectures. DBN is a generative graphical model that consists of a stack of Restricted Boltzmann machines (RBMs) [20, 24–26]. RBM is a symmetrical graph (each visible node is connected with each hidden node) that consists of two layers: a layer of visible nodes and a layer of hidden nodes with no connections in the same layer [20, 24, 27]. Figure 1 shows an example of DBN. Each layer in the DBNs has a double role, it serves as the hidden layer to the nodes that come before and as the visible layer to the nodes that come after. The training of DBN can be a discriminating training for inference problem, classification problem; or a generating training to generate training data [21, 28].

Fig. 1. An example of DBN that was stacked of 3 RBMs

4 Methodology Our proposed work aims to handle the above shortcomings by suggesting an enhancement for concepts relevance measures for handling the first shortcoming and by suggesting six more relation types for handling the fourth shortcoming as well as by using DL techniques for handling second and third shortcomings. That is because DL can handle a large amount of data in an efficient and effective way as well as because using predefined patterns can give a reasonable precision, but a very low recall because that any relation

378

F. N. AL-Aswadi et al.

is not within the predefined patterns cannot be detected. While DL is based on the deep learning fundament [28]. We proposed these suggestions based on our knowledge and literature that we presented and discussed in [19]. 4.1 Concept Relevance Measurements From the existing studies on ontology construction, the concepts relevance measurements that are used are Term Frequency-Inverse Document Frequency (TF-IDF) that is shown in Eqs. (1), (2) and (3); Domain Relevance measures (DR) that is shown in Eq. (6); Domain Consensus measures (DC) that is shown in Eqs. (4) and (5); and Domain Relevance value (DRM) that is shown in Eqs. (7) and (8) . tf = idf = log2

count of concept c in d total number of concepts in d

(1)

the size of d count of documents where concept c appears tf − idf = tf × idf

DC =

 dj ∈D











p c, dj × log2



1   p c, dj

(3)  (4)

  freq c ∈ dj

  p c, dj = n freq c ∈ dj j = 1 dj ∈ D     freq c ∈ Dj   DR c, Dj = n freq c ∈ Dj j=1 DRM =

λ(c) =

(2)

|log λ(c)| − min|log λ| tf (c) df (c) × × max(tf ) max|log λ| − min|log λ| min(df )

max pk1 (1 − p)n1−k1 pk2 (1 − p)n2−k2 p

max p1k1 (1 − p1 )n1−k1 p2k2 (1 − p2 )n2−k2

, λ ∈ [0, 1]

(5)

(6)

(7)

(8)

p1 ,p2

Where c refers to the concept, d refers to the corpus, D refers to domain corpora, λ refers to the likelihood ratio, n is the number of document collections, k1 and k2 refer to the frequencies of concept (c) in the domain and contrasting domain, n1 and n2 refer to the total number of concepts in the domain and contrasting domain respectively, p refers to conditional probability for the concept (c), p1 and p2 refer to the probability of concept (c) in the domain and contrasting domain respectively, and df refers to the document frequency of a concept (c) in the document. These measurements are good in some aspects and have weakness in others, to name a few, TF-IDF is used in some studies such as [10, 29–31] to select the relevant concepts

Extracting Semantic Concepts and Relations from Scientific Publications

379

or objects for the domain. However, it is not suitable for identifying the significant concepts of a corpus [13]; it is a trustworthy measure for identifying important keywords in individual documents but not for all the corpus. On the other hand, the DRM measure that is suggested and used in [13], is a good measure for identifying the significant concepts of a small corpus. However, according to our experiments, it is not efficient for big corpus due to it estimated the likelihood ratio for each term, which causes extremely high computing cost. Also, DC and DR measures are used together in [32] study to identify the significant concept for the selected domain. However, DR value merely depends on the concept’s frequency in the target domain corpus and the contrasting corpora so if the size of one of this corpus is adjusted; the result of DR would be greatly different [13]. Besides all the above, all the previous measurements donot take into account the time factor to identify the significant concepts for the domain. For explaining the importance of the time factor to measure the relevant concept, suppose that we have corpora in the IT domain, the programming language is a concept in IT domain and at the same time it is a sub-domain of it that has concepts related to it. Now, suppose that after extracting the concepts we get the following instances: Basic language, Pascal language, C language, c# language, Java language, Python language and XXX language (suppose it is a new language that just appears). Then without using time factor in relevance measures, the new appeared language (XXX language) might have a low relevant value and not identifying as a significant, while Basic language is identified as a significant. So, using the time factor is essential to identifying the relevant concepts. By using it, the new appeared language (XXX language) would have a better relevant value, and it would identify as a significant. In contrast, Basic language would have a low relevant value and then it is identified as “old” (old programming language) and be less significant. Based on all the above, we suggest a new concept relevance measurement that aims to solve the previous problems. It is Domain Time Relevance (DTR) measure that estimates the rate of repetition of the concept at all times of domain corpora so that the greater frequency value of the concept with the progress of time, means the greater degree of relevance for this concept. The domain time concept of a concept c in domain D at time t is given by Eq. (9): 1       DC c, Dti+1 − DC c, Dti DRT =

(9)

I =n−1

Where c refers to the concept, t refers to the time, and D refers to domain corpus. 4.2 Semantic Relations Most of the existing studies extracted very limited relations almost do not exceed synonym, hyponym, hypernym, meronyms, and/or holonyms relations such as in [8, 13, 14, 18]. Based on our knowledge and the initial analysis for around 1000 scientific articles that are collected from SCOPUS engine, we can define the types of semantic relations for concepts on scientific domains as it is shown in Table 1. We have suggested six more relation types (homonyms, usage, result, comparison, model, and dependence) for handling the semantic relations shortcoming.

380

F. N. AL-Aswadi et al. Table 1. Semantic relation types

Relation type

Example

Linguistic relation

Equal

Data, Information

Synonyms

Is_A

Bubble Sort, Sorting Algorithm

Hyponyms Hypernyms

Has_A

Algorithm, Performance

Holonyms

Different_of

Plant, Plant or CNNa , CNNb

Homonyms

Part_of

Red Blood Cells, Blood

Meronyms

Used_to Used_by

Technology, Waste Food Technology, Human

Usage

Result_of

Reliable Ontology, Precision Relation

Result

Compared_to

Bubble Sort, Merge Sort

Comparison

Use_A -

Image Classification, Machine Learning

Model

Depend_On

Performance, Data Size

Dependence

a Name of a TV news channel. b Abbreviation of Convolutional Neural Network.

4.3 Extracting Semantic Concepts and Relations Figure 2 shows the workflow of the proposed work to extract the semantic concepts and relations from scientific publications. In our previous work [19], we discussed and presented in details the literature review about ontology construction, OL, DL, and the DL for OL. In this paper, we introduce a proposal of extracting semantic concepts and relations by using DBN. In this work, DBN is used to classify the extracted concepts and to extract the relations between concepts. Also, it is used to classify the extracted semantic relations under the main relations types that were defined in Table 1. After pre-processing the text and extracting the concepts from the text by using ngram and other concept extraction methods and after applying the proposed relevance measure, the terms and tags bags are built in binary representation. Then the system assigning the part of speech (POS) and syntactic tags to each individual term in binary representation. This combining aims to build the feature vectors (training file for DBN). After building appropriate DBN and training it, the trained DBN can be used to classify the concepts and to extract the relations. The concept classification by using DBN will be through two processes. First one is detection process, in this process DBN has only two target outputs: “yes” and “no” where “yes” means that c1 ⊆ c2 while “no” means c1 c2 for each input vector (ci refers to concept i). The second process is the classification process, in this process the second and the third up the level of the k most relevant concepts are used as the DBN target outputs. The k most relevant concepts are identified by using proposed measures. The same processes are done for semantic relations detection and classification except that the target outputs of the classification process are the relations identified in Table 1.

Extracting Semantic Concepts and Relations from Scientific Publications

381

Fig. 2. Workflow to extract semantic concepts and relations from scientific publications.

5 Conclusion In this paper, we introduced new relevance measurement for concepts as well as we introduced six new types of semantic relations, they are homonyms, usage, result, comparison, model, and dependence. Furthermore, we presented a proposal for extracting semantic concepts and relations automatically from scientific publications by using DBN. This proposal aims to address the four main shortcomings and drawbacks in current ontology

382

F. N. AL-Aswadi et al.

extraction systems as it pointed out above. In the future work, we will illustrate in details with experimental results how this proposed work has be performing and enhancing the ontology construction process.

References 1. Mishra, S., Jain, S.: A study of various approaches and tools on ontology. In: 2015 IEEE International Conference on Computational Intelligence & Communication Technology (CICT), pp. 57–61 (2015) 2. Maimon, O., Browarnik, A.: Ontology learning from text: why the ontology learning layer cake is not viable. Int. J. Signs Semiot. Syst. 4(2), 1–4 (2015) 3. Zou, X.: A survey on application of knowledge graph. J. Phys. Conf. Ser. 1487, 012016 (2020) 4. Ergeta, M.: Introduction to Knowledge Graphs and their Applications (2019). https://med ium.com/analytics-vidhya/introduction-to-knowledge-graphs-and-their-applications-fb5b12 da2a8b. Accessed 30 Sept 2020 5. da Silva, J.W.F., Venceslau, A.D.P., Sales, J.E., Maia, J.G.R., Pinheiro, V.C.M., Vidal, V.M.P.: A short survey on end-to-end simple question answering systems. Artif. Intell. Rev. 53(7), 5429–5453 (2020) 6. Al-Ghuribi, S.M., Noah, S.A.M.: Multi-criteria review-based recommender system-the state of the art. IEEE Access. 7, 169446–169468 (2019) 7. Franco, W., Avila, C.V.S., Oliveira, A., Maia, G., Brayner, A., Vidal, V.M.P., et al.: Ontologybased question answering systems over knowledge bases: a survey. In: ICEIS, vol. 1, pp. 532– 539(2020) 8. Arnold, P., Rahm, E.: Extracting semantic concept relations from Wikipedia. In: Proceedings of the 4th International Conference on Web Intelligence, Mining and Semantics, WIMS14, Thessaloniki, Greece, pp. 1–11. ACM (2014) 9. Suchanek, F.M., Kasneci, G., Weikum, G.: YAGO: a core of semantic knowledge. In: Proceedings of the 16th International Conference on World Wide Web, Banff, Alberta, Canada, pp. 697–706. ACM (2007) 10. Maedche, A., Volz, R.: The text-to-onto ontology extraction and maintenance environment. In: Proceedings of the ICDM-Workshop on Integrating Data Mining and Knowledge Management, San Jose, California, USA (2001) 11. Hahn, U., Romacker, M.: The SYNDIKATE text knowledge base generator. In: Proceedings of the 1st International Conference on Human Language Technology Research, San Diego, pp. 1–6. Association for Computational Linguistics (2001) 12. Hahn, U., Marko, K.G.: Ontology and lexicon evolution by text understanding. In: Proceedings of the ECAI 2002 Workshop on Machine Learning and Natural Language Processing for Ontology Engineering, OLT 2002, Lyon, France (2002) 13. Jiang, X., Tan, A.H.: CRCTOL: a semantic-based domain ontology learning system. J. Am. Soc. Inform. Sci. Technol. 61(1), 150–168 (2010) 14. Gillani Andleeb, S.: From text mining to knowledge mining: an integrated framework of concept extraction and categorization for domain ontology, p. 146. Department of Information Systems, Budapesti Corvinus Egyetem, Budapest (2015) 15. Al-Ghuribi, S.M., Alshomrani, S.: Bi-languages mining algorithm for extraction useful web contents (BiLEx). Arab. J. Sci. Eng. 40(2), 501–518 (2015) 16. Wong, W., Liu, W., Bennamoun, M.: Ontology learning from text: A look back and into the future. ACM Comput. Surv. (CSUR) 44(4), 20 (2012) 17. Paiva, L., Costa, R., Figueiras, P., Lima, C.: Discovering semantic relations from unstructured data for ontology enrichment: association rules based approach. In: 2014 9th Iberian Conference on Information Systems and Technologies (CISTI), pp. 1–6 (2014)

Extracting Semantic Concepts and Relations from Scientific Publications

383

18. Lee, L.-H., Lee, K.-C., Tseng, Y.-H.: The NTNU system at SemEval-2017 task 10: extracting keyphrases and relations from scientific publications using multiple conditional random fields. In: Proceedings of the 11th International Workshop on Semantic Evaluation, Vancouver, Canada, pp. 951–955. Association for Computational Linguistics (2017) 19. Al-Aswadi, F.N., Chan, H.Y., Gan, K.H.: Automatic ontology construction from text: a review from shallow to deep learning trend. Artif. Intell. Rev. 53(6), 3901–3928 (2020) 20. Deng, L.: Three classes of deep learning architectures and their applications: a tutorial survey. APSIPA Trans. Sig. Inf. Process. (2012) 21. Mo, D.: A survey on deep learning: one small step toward AI. Department of Computer Science, University of New Mexico, USA (2012) 22. Arel, I., Rose, D.C., Karnowski, T.P.: Deep machine learning - a new frontier in artificial intelligence research [research frontier]. IEEE Comput. Intell. Mag. 5(4), 13–18 (2010) 23. Chen, X.-W., Lin, X.: Big data deep learning: challenges and perspectives. IEEE Access. 2, 514–525 (2014) 24. Deng, L., Yu, D.: Deep learning. Sig. Process. 7, 3–4 (2014) 25. Hinton, G.E.: Deep belief networks. Scholarpedia 4(5), 5947 (2009) 26. Hinton, G.E., Osindero, S., Teh, Y.-W.: A fast learning algorithm for deep belief nets. Neural Comput. 18(7):1527–1554 (2006) 27. Fischer, A., Igel, C.: An introduction to restricted Boltzmann machines. In: Iberoamerican Congress on Pattern Recognition, pp. 14–36. Springer, Heidelberg (2012) 28. Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M.P., et al.: A survey on deep learning: algorithms, techniques, and applications. ACM Comput. Surv. 51(5) (2018). Article 92 29. Dahab, M.Y., Hassan, H.A., Rafea, A.: TextOntoEx: automatic ontology construction from natural English text. Exp. Syst. Appl. 34(2), 1474–1480 (2008) 30. Cimiano, P., Völker, J.: Text2Onto. In: Montoyo, A., Mu´noz, R., Métais, E. (eds.) Natural Language Processing and Information Systems: Proceedings of the 10th International Conference on Applications of Natural Language to Information Systems, NLDB 2005, Alicante, Spain, 15–17 June 2005, pp. 227–238. Springer, Heidelberg (2005) 31. Kolbe, N., Vandenbussche, P.-Y., Kubler, S., Traon, Y.L.: LOVBench: ontology ranking benchmark. In: Proceedings of the Web Conference 2020, Taipei, Taiwan, pp. 1750–1760. Association for Computing Machinery (2020). 32. Missikoff, M., Velardi, P., Fabriani, P.: Text mining techniques to automatically enrich a domain ontology. Appl. Intell. 18(3), 323–340 (2003)

Effectiveness of Convolutional Neural Network Models in Classifying Agricultural Threats Sayem Rahman(B) , Murtoza Monzur, and Nor Bahiah Ahmad Faculty of Engineering, School of Computing, Universiti Teknologi Malaysia UTM, 81310 Skudai, Johor, Malaysia {rsayem2,muamurtoza2}@graduate.utm.my

Abstract. Smart farming has recently been gaining traction for more productive and effective farming. However, pests like monkeys and birds are always a potential threat for agricultural goods, primarily due to their nature of destroying and feeding on the crops. Traditional ways of deterring these threats are no longer useful. The use of highly effective deep learning models can pave a new way for the growth of smart farming. This study aims to investigate the manner in which deep learning convolutional neural network (CNN) models can be applied to classify birds and monkeys in agricultural environments. The performance of CNN models in this case is also investigated. In this regard, four CNN variants, namely, VGG16, VGG19, InceptionV3 and ResNet50, have been used. Experiments were conducted on two datasets. The experimental results demonstrate that all the models have the capability to perform classification in different situations. Data quality, parameters of the models, used hardware during experiments also influence the performance of the considered models. It was also found that the convolutional layers of the models play a vital role on classification performance. The experimental results achieved will assist smart farming in opening new possibilities that may help a country’s agriculture industry, where efficient classification and detection of threats are of potential importance. Keywords: Smart farming · Convolutional neural network (CNN) · Deep learning · Computer vision · Image processing

1 Introduction With the rise of civilization and modernization, mankind has greatly benefitted from many aspects of science and technology, but the shortcomings have also been increasing. When it comes to managing food for the world, which certainly has the largest population than ever before, the available resources have not been on par to support it. Consequently, there exist problems regarding hunger issues, and the threat toward agricultural goods also cannot be denied. Among the threats to agricultural farms is the pest problem. Birds and monkeys are a prominent threat to this particular sector, which certainly affects the productivity of any farm. Recent studies showed that millions of dollars are being wasted due to animals and birds destroying crops. The traditional ways to prevent this are failing © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 F. Saeed et al. (Eds.): IRICT 2020, LNDECT 72, pp. 384–395, 2021. https://doi.org/10.1007/978-3-030-70713-2_36

Effectiveness of Convolutional Neural Network Models

385

[1], as they are not sufficiently effective, hence it is time to find a sustainable solution to minimize the damage to crops as much as possible. Even animals are increasingly becoming smarter nowadays, and this requires finding more technical and up-to-date solutions. Using deep learning and image processing to classify and detect them away could lead to being a possible solution for this particular problem. Reduction in production efficiency due to birds and animal attacks in agriculture fields, orchards and ponds, are both long-standing and high-cost problems [2]. There are millions of cases worldwide and billions of dollars of wastage on a yearly basis due to this less-considered yet highly affecting problem. Traditional techniques are not effective enough to control bird and animal attacks on farms. At the same time, farming is the most important source of the gross domestic product (GDP), as it contributes around 22% of a country’s overall GDP [1]. Traditional techniques such as scarecrows or other violent ways of scaring away the pests are not long term solutions. This new modern era needs a long term and effective solution to scare away birds and animals. Previous research works mostly focused on bird detection and producing specific sounds using the Viola Jones algorithm, image pre-processing, features extraction and template matching [1–3]. However, the long term effects of previous works remain questionable because of the smart nature of some animals such as monkeys. Therefore, this study highlights both birds and animals, specifically monkey classification, using the Wild Bird Image [4] dataset and 10 Monkey Species [5]. When it comes to classifying any intruder, the use of images is the best option, as it leads to fast classification and recognizing the target object. The application of convolutional neural networks (CNNs) of deep learning in this sector is becoming a new norm. There are numerous CNN models available that are being used successfully for object classification and detection. Using CNN models to classify the intruders with the help of image classification allows farmers to be notified in detecting possible threats for their farms. If it can be implemented successfully, it will certainly help to grow a country’s economy. This study applied several existing CNN variants with two datasets to investigate the performance of these models. The model variants considered to analyze overall performance are VGG16, VGG19, InceptionV3 and ResNet50. The champion of the ILSVRC 2014 contest was the GoogleNet, also known as Inception model. VGG net was the runner up of that contest. Other two models also have very high accuracy in terms of classification from images. The remainder of this paper is organized as follows. Section 2 provides an overview of previous works to the particular domain. Section 3 explains the architecture of the convolutional neural network (CNN) model applied. Section 4 presents the proposed methodology for this study. The results and performance of the CNN models are discussed in Sect. 5. Finally, Sect. 6 concludes this paper and discusses potential future work.

2 Related Work A survey by USDA’s Natural Agricultural Statistics Service (NASS) in 1994 shows that nearly half of the field crop producers face losses that are estimated to be around $316 million due to wildlife invasion [2]. Another survey by USDA NASS (1999) in

386

S. Rahman et al.

seven states and two crops, reported a loss of tens of millions of dollars and a reduction in product quality each year, directly by birds [6]. The use of different types of deterring techniques such as different scarecrows, gas cannons, sensor triggered devices, laser beams, and flashlights, has a short duration of effectiveness in terms of deterring the threats [2, 6, 7]. Classifying and detecting the original threat is always challenging, hence, deep learning-based computer vision solutions always achieve better performance, as these are successful in classifying the objects correctly. Many applications such as image classification, speech recognition, and language processing learning techniques use deep learning [8]. Deep learning-based convolutional neural network (CNN) models are currently being used to achieve high precision and effectiveness when dealing with large amounts of image data. CNNs work with typical feedforward neural networks. Backpropagation CNNs comprise many hierarchical layers like feature map layers and classifications layers [9]. A proposed project [3] named “Smart Scarecrow” used image processing with feature extraction and template matching to detect birds, and then used specific sounds to deter them. The study achieved an accuracy of around 90.47%, which was satisfying. Another research team used Context-SVM for object classification and detection using VOC 2007 and 2010 datasets. They achieved the highest mean A.P. of 73.0, and won a competition using the VOC datasets [10]. Image processing and template matching were used to successfully detect and analyze the bird’s motion using the Wild Birds in a Wind Farm dataset [1]. Deep learning models were used by [11] to locate the monkey features in their natural habitat. A total of 6000 high-resolution image dataset of monkeys from the Primate Research Institute was used to conduct this research in 2018 [11]. Bird species from their natural habitats were successfully identified by extracting features from the bird’s colour, wings, beak and so on by another research group. Using CNN, the study achieved an 80% success rate in predicting the bird’s class from images. The CaltechUCSD Birds 200 (CUB-200-2011) dataset was used for the study [12]. The work by [13] involved a complete system model in classifying and detecting birds and other objects. The study pre-trained CNN models like Inception v3, ResNet v2, NASNet mobile, and MobileNet. This particular study achieved 100% accuracy in some cases, indicating the possibility and high success rate of trained CNN models [13].

3 The Architecture of Convolutional Neural Network (CNN) CNN is capable of extracting certain features from the input data. The existing neurons inside the CNN models are capable of extracting high-level abstraction features of the extracted features from the previous layers. The architecture of deep learning CNN models has four main types of building blocks, namely, Convolution Layers, Non-Linearity (Rectified Linear Unit), Max-Pooling and Fully Connected or Classification Layers [14]. Figure 1 illustrates the basic CNN architecture and shows the different layers involved. 3.1 Convnet Layers First Layer: Convolutional Layer. A CNN model has at least one convolutional layers, and one or more fully connected layers. Identifying features with the help of a feature

Effectiveness of Convolutional Neural Network Models

387

Fig. 1. Basic convolutional neural network architecture showing different layers [14].

map in the dataset is known as convolutional operation, and convolutional layers handle these types of operations. Pooling Layer. Max pooling layer reduces noise in the filters. After reducing the noise, the maximum value is extracted and then put in the pooled feature map by the maxpooling layer. This maximum value functions as the input for the next layer. Pooling layers handle the problem of overfitting. Other common pooling layer operations are average pooling, stochastic pooling, spectral pooling, spatial pyramid pooling and multiscale orderless pooling. Fully Connected Layer. Identifying the classification output is the responsibility of the fully connected layers, which are positioned after the convolutional layers, giving the output of the classifier. High-level reasoning decisions are made in the fully connected layers. Loss Layer. Other than the four main layers, there is also a layer referred to as the loss layer, where the fully connected layers serve as the lost layers that compute the loss or error. The error is the penalty for the imbalance between the actual and the desired output. Softmax loss is used to predict a single class out of some mutually exclusive classes, and is widely used as a loss function.

3.2 Activation Functions Another important part of the neural network is activation functions. These activation functions take single numbers, and upon some mathematical computations, determine the output of a neural network. Each neuron in the network has the function attached to it. Activation or deactivation of the neuron depends on the activation function based on the model’s prediction. Some commonly used functions are Sigmoid, Tanh, ReLU (Rectified Linear Unit) and Leaky ReLU.

388

S. Rahman et al.

Sigmoid. Real-Valued functions and output values are taken as an input by the Sigmoid activation function. The input is taken in the range of 0 and 1, as shown in Eq. (1) [14]. f (x) = max(0, x)

(1)

Tanh. Tanh is considered a scaled-up version of sigmoid, as shown in Eq. (2). It generates output values ranging from −1 to 1. Tanh is preferred over Sigmoid because it has fewer drawbacks than the Sigmoid function. The reason is Tanh is zero centred, so gradients do not oscillate between positive and negative values [14]. f (x) =