Comparison of machine learning models for bradykinesia assessment using digital health technology.
Comparison of machine learning models for bradykinesia assessment using digital health technology.
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
North Chicago, US · Author affiliation
Data and Statistical Sciences, AbbVie, North Chicago, IL, United States.Location evidence
US · Author affiliation · country only
Department of Neurology, University of Rochester Medical Center, Rochester, NY, United States.Location evidence
A plain-language reading has not been prepared for this paper yet.
Original abstract
BACKGROUND: Bradykinesia, a primary symptom of Parkinson's disease, significantly impacts patients' quality of life. Traditional assessments like the MDS-UPDRS require in-person evaluation and can be subjective, resulting in patient burden and variability. METHODS: This study investigates digital health technologies and machine learning (ML) models, specifically XGBoost and Convolutional Neural Networks (CNNs), to improve objectivity and precision in bradykinesia quantification. Using data from the 12-month WATCH-PD study in early, untreated Parkinson's disease, models were evaluated against MDS-UPDRS items (3.6 pronation-supination and 3.7 toe tapping) for cross-sectional accuracy and longitudinal sensitivity. RESULTS: Results show that the class-balanced XGBoost was the best-performing model on both tasks, outperforming a stratified-random baseline and all CNN variants (cross-entropy, focal loss, and self-attention) across every imbalance-aware metric: macro-F1 0.47 and 0.40, balanced accuracy 0.47 and 0.42, quadratic weighted kappa 0.58 and 0.47, and accuracy 0.60 and 0.53 for pronation-supination and toe tapping, respectively, with the clearest gains on the under-represented higher-severity classes. A secondary longitudinal analysis found limited sensitivity to 12-month change across the ML models. DISCUSSION: Overall, machine-learning models, led by the class-balanced XGBoost, show promise in accurately predicting item-level MDS-UPDRS bradykinesia scores from wearable sensor data. Our results also highlight the importance of aligning feature selection and model architecture to intended clinical endpoints.