Task-specific AI modeling methods for parkinson disease diagnosis

다국어 초록 (Multilingual Abstract)

Recent studies in Parkinson's Disease (PD) diagnosis increasingly integrate AI technology using wearable sensors and deep learning. However, conventional works show limitations such as observation focusing controlled movements, isolated analysis for each symptom, and inefficiency from threshold-based single classifiers. To address these challenges, researches adopting enhanced machine learning methodologies including ensemble learning algorithms.
This study suggests practical diagnosis methods and ensemble AI model to improve accuracy and robustness. Hence, two experiments, Gait Analysis and Tremor Analysis during Walking, are integrated using the PD-BioStampRC21 dataset and a single accelerometer. The HMM and VA-AdaBoost framework for gait analysis achieved high performance (93% Accuracy, 1.0 Sensitivity). For tremor analysis, we propose a practical method using a single sensor in daily life environment but achieved 87% accuracy and outstanding 0.96 AUC. Further, newly proposed novel features, APIT and APF, improve diagnostic performance by capturing the subtle movement characteristics of tremor.

다국어 초록 (Multilingual Abstract)

Recent studies in Parkinson's Disease (PD) diagnosis increasingly integrate AI technology using wearable sensors and deep learning. However, conventional works show limitations such as observation focusing controlled movements, isolated analysis for each symptom, and inefficiency from threshold-based single classifiers. To address these challenges, researches adopting enhanced machine learning methodologies including ensemble learning algorithms.
This study suggests practical diagnosis methods and ensemble AI model to improve accuracy and robustness. Hence, two experiments, Gait Analysis and Tremor Analysis during Walking, are integrated using the PD-BioStampRC21 dataset and a single accelerometer. The HMM and VA-AdaBoost framework for gait analysis achieved high performance (93% Accuracy, 1.0 Sensitivity). For tremor analysis, we propose a practical method using a single sensor in daily life environment but achieved 87% accuracy and outstanding 0.96 AUC. Further, newly proposed novel features, APIT and APF, improve diagnostic performance by capturing the subtle movement characteristics of tremor.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1. Diagnosis of Parkinson's Disease 1
• 1) PD Symptoms and Identification Methods 1
• 2) Diagnosis of PD using AI algorithms 2
• 2. Direction and Objectives of Research 4
• 3. Future Paper Structure 8
• Chapter 2. Related Works 9
• 1. Conventional Diagnosis Methods and Limitation 9
• 1) Early Stage of PD Detection Methods 9
• 2) Wearable Sensors in Movement Disorder Assessment 14
• 2. Machine Learning Application in PD Detection 16
• 1) Increasing challenge with Machine learning techniques 16
• 2) Hidden Markov Models in Biological Signal Processing 18
• 3) Ensemble Learning Approaches for Medical Diagnosis 20
• Chapter 3. Research Design and Experimental Frameworks 26
• 1. Framework I; Ensemble Learning for Gait Pattern Analysis 26
• 1) Overview of Method 26
• 2) Hidden Markov Model (HMM) Segmentation Detection 28
• 3) Feature Extraction 28
• 4) Optimized AdaBoost Classification 29
• 2. Framework II; Analysis of Tremor Dynamics during Walking 29
• 1) Overview of Method 29
• 2) Data Collection and Preprocessing 31
• 3) Comprehensive Feature Extraction 31
• 4) VA-AdaBoost Classification Model 32
• 3. Dataset and Preprocessing 32
• 4. Evaluation of the proposed Methods 36
• Chapter 4. Gait Analysis based on HMM driven Model 40
• 1. Walking Segmentation based on HMM model 40
• 1) Feature Extraction for HMM-Based Segmentation 41
• 2) HMM Configuration and Parameter Optimization 43
• 3) State Sequence Prediction and Walking Segmentation Identification 45
• 4) Feature Extraction for PD Classification 46
• 5) Optimized AdaBoost Algorithm with Variance-Based Weight Initialization 47
• 2. Experimental Setup 49
• 3. Result and Performance Analysis 52
• 1) HMM Transition Matrix and Emission Parameters 52
• 2) Superiority of HMM Segmentation 54
• 3) Comparison with Other Ensemble Learning Model 57
• 4. Discussion of Findings and Summary 62
• Chapter 5. Analysis of Hand Tremor Activity During Walking 64
• 1. Methodology Development 64
• 1) Feature Extraction for Tremor Analysis 64
• 2) Theoretical Background of Peak Dynamic Features 66
• 3) Moving Variance-Enhanced Scanning Mechanism 69
• 4) Ensemble Learning Framework 71
• 2. Experimental Setup 72
• 3. Result and Performance Analysis 75
• 1) Performance of the Ensemble Learning Model 75
• 2) Impact of Novel Features; APIT and APF 80
• 3) Comparison with Previous Work 85
• 4. Integrated Analysis combined Tremor and segmented Gait signal 87
• 5. Discussion of Findings and Summary 92
• Chapter 6. Conclusion and Future Works 94
• 1. Summary of Key Finding and Contribution 94
• 1) Principal Findings 94
• 2) Major Theoretical and Methodological Contributions 97
• 3) Limitation of Current Research 99
• 4) Clinical Application 100
• 2. Future Research Directions 102
• 1) Multimodal Integration Possibilities 102
• 2) Longitudinal Studies for Disease Progression Monitoring 103
• 3) Advanced Algorithmic Developments 1104
• 4) Technology Integration and Scalability 105
• Reference 107
• Appendix 124
• Abstract (국문 초록/요약) 130