TremorFusion: AI-driven feature extraction for multi-class Parkinson's tremor classification using CSVM and DeepK-CNN

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Parkinson's disease (PD) significantly impacts patients' quality of life.
• Accurate tremor detection and classification are crucial for effective PD management.
• Current methods may lack the precision for real-time, remote patient assessment.

Purpose of the Study:

• To evaluate a wrist-worn accelerometer for detecting and classifying five distinct Parkinson's disease tremor types.
• To develop and compare novel machine learning models for tremor analysis.
• To assess the feasibility of a wearable sensor for remote PD patient monitoring.

Main Methods:

• Utilized a wrist-worn accelerometer to collect tremor data from 90 PD patients.
• Developed a Cubic Support Vector Machine (CSVM) model with novel feature extraction and Poincaré-based analysis.
• Implemented a hybrid DeepK-CNN model integrating Non-negative Matrix Factorization (NMF), k-Nearest Neighbor (KNN), and Convolutional Neural Network (CNN).
• Employed leave-one-patient-out fivefold cross-validation for performance evaluation.

Main Results:

• The CSVM model achieved high performance: 95.28% mean sensitivity, 95.51% mean specificity, and a low false alarm rate (FAR) of 0.15/24h.
• The DeepK-CNN model demonstrated good performance with 87.51% mean sensitivity and 91.65% mean specificity.
• Statistical analysis confirmed the significance of both models (p<0.001 for CSVM, p<0.05 for DeepK-CNN).

Conclusions:

• A wrist-worn accelerometer combined with the CSVM model offers a highly accurate and feasible solution for real-time Parkinson's disease tremor classification.
• The proposed methods support personalized treatment strategies and remote patient assessment, facilitating clinical adoption.
• Wearable sensor technology holds significant potential for improving the management of neurological disorders like Parkinson's disease.