Weakly-supervised learning for Parkinson's Disease tremor detection.

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Continuous monitoring of Parkinson's Disease (PD) symptoms is crucial for effective patient care and treatment adjustment.
• Acquiring precisely labeled data for PD symptoms like tremor is difficult, necessitating weakly-supervised learning approaches.
• Existing weakly-supervised algorithms struggle with increasing data segment lengths and decreasing label precision.

Purpose of the Study:

• To evaluate standard weakly-supervised algorithms for PD tremor detection.
• To propose and assess a novel 'stratified' version of these algorithms.
• To analyze algorithm performance with increasing training segment lengths and decreasing label precision.

Main Methods:

• Evaluation of five standard weakly-supervised algorithms for PD tremor detection.
• Development and implementation of a 'stratified' approach for three algorithms, incorporating approximate tremor information.
• Analysis of algorithm performance across varying training segment lengths (30 seconds to 10 minutes).

Main Results:

• Standard weakly-supervised algorithms showed significant performance degradation with longer data segments (up to 10 minutes).
• Stratified algorithms demonstrated robust tremor detection performance, with minimal decrease as segment length increased.
• The proposed stratified approach effectively utilizes more nuanced labels for improved accuracy.

Conclusions:

• Stratified weakly-supervised algorithms offer a promising solution for continuous, automated PD symptom monitoring.
• These algorithms maintain performance despite reduced label precision associated with longer data segments.
• Improved monitoring capabilities can lead to better understanding of PD progression and optimized treatment strategies.