High-Resolution Motor State Detection in Parkinson's Disease Using Convolutional Neural Networks

Area of Science:

• Neurology
• Biomedical Engineering
• Artificial Intelligence

Background:

• Advanced Parkinson's disease (PD) patients experience motor fluctuations.
• Objective monitoring of these motor states (OFF, ON, DYSKINETIC) is crucial but challenging.
• Current monitoring methods are often subjective or limited to clinical settings.

Purpose of the Study:

• To develop and validate a deep learning model for classifying motor states in PD patients.
• To assess the feasibility of using a single wrist-worn IMU sensor for continuous, real-world monitoring.
• To correlate model predictions with expert clinical assessments.

Main Methods:

• Collected 8,661 minutes of IMU data from 30 PD patients in unscripted environments.
• Used a convolutional neural network (CNN) with 1-minute data windows.
• Expert clinicians annotated motor states (OFF, ON, DYSKINETIC) every minute using MDS-UPDRS and AIMS scales.
• Validated the model on unseen patient data.

Main Results:

• Achieved a 3-class balanced accuracy of 0.654 for classifying OFF, ON, and DYSKINETIC states.
• Model demonstrated high correlation with per-subject annotations (r=0.83/0.84) and 1-minute windows (r=0.64/0.70).
• Sensitivity/specificity varied by state, with OFF/DYSKINETIC states showing higher specificity (0.89).

Conclusions:

• Deep learning models can effectively classify motor states in Parkinson's disease using single IMU data.
• This approach shows feasibility for objective, long-term motor state detection in free-living conditions.
• Potential for improved management and understanding of PD motor fluctuations.