Machine Learning in the Parkinson's disease smartwatch (PADS) dataset

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Smart devices offer novel tools for movement disorder research.
• A lack of comprehensive, annotated movement datasets hinders ML model development for home-based disease detection.
• Existing research needs robust data for reliable diagnosis and treatment monitoring.

Purpose of the Study:

• To create a comprehensive dataset of movement data and clinical annotations for Parkinson's disease (PD) and its differential diagnoses (DD).
• To develop and validate machine learning (ML) models using smart device data for disease detection and monitoring.
• To facilitate home-based assessment and management of movement disorders.

Main Methods:

• A three-year cross-sectional study involving 504 participants (PD, DD, healthy controls).
• Utilized a multi-modal smartphone app with electronic questionnaires and smartwatch measures during neurologist-designed assessments.
• Applied an integrative ML approach combining signal processing and deep learning, followed by cross-validation.

Main Results:

• Achieved 91.16% balanced accuracy in classifying Parkinson's disease (PD) vs. healthy controls (HC).
• Attained 72.42% balanced accuracy in differentiating PD from differential diagnoses (DD).
• Highlighted challenges in distinguishing between similar movement disorders.

Conclusions:

• The developed dataset and ML models show significant potential for smart device-based movement disorder detection.
• Distinguishing between PD and its differential diagnoses remains a key area for future research.
• The comprehensive annotations support further investigation into phenotypical biomarkers for movement disorders.