An Interpretable Deep Learning Optimized Wearable Daily Detection System for Parkinson's Disease.

Area of Science:

• Biomedical Engineering
• Neurology
• Artificial Intelligence

Background:

• Parkinson's disease (PD) monitoring requires objective gait analysis for disease progression tracking.
• Current methods for assessing PD symptoms during daily activities are often subjective and infrequent.

Purpose of the Study:

• To develop an accurate, objective, and passive deep learning algorithm for detecting Parkinson's disease (PD) during daily walking.
• To identify representative spatiotemporal motor features indicative of PD.
• To optimize the algorithm using interpretable AI for enhanced efficiency.

Main Methods:

• Collected motion data from 100 subjects (PD patients and healthy controls) using five inertial measurement units (IMUs) on the wrist, ankle, and waist.
• Applied continuous wavelet transform to sensor data and trained a 6-channel convolutional neural network (CNN) for classification.
• Utilized gradient-weighted class activation mapping and 3D CNN visualization for model interpretability and optimization.

Main Results:

• The waist-mounted sensor achieved the highest classification accuracy (98.01% ± 0.85%) and AUC (0.9981 ± 0.0017).
• Key PD-related gait features were identified in the lower frequency band (0.5-3Hz).
• Optimizing the model based on visual interpretation reduced data processing by 50% while maintaining high performance (AUC=0.9929 ± 0.0019).

Conclusions:

• Interpretable deep learning models, particularly using waist-worn sensors, offer a highly accurate and efficient method for passive Parkinson's disease detection.
• Visual interpretation of AI models can guide feature selection, reducing computational costs without compromising diagnostic performance.
• This approach represents a novel advancement in the intelligent diagnosis and monitoring of Parkinson's disease.