Continuous Assessment of Function and Disability via Mobile Sensing: Real-World Data-Driven Feasibility Study.

Area of Science:

• Digital Health
• Machine Learning
• Biomedical Informatics

Background:

• Functional limitations impact clinical outcomes, mortality, and disability, particularly in older adults.
• Traditional functional assessments are time-consuming and infrequently utilized in clinical practice.
• Mobile sensing presents an opportunity for daily assessment of patient function and disability.

Purpose of the Study:

• To assess the feasibility of an interpretable machine learning pipeline.
• To predict function and disability using World Health Organization Disability Assessment Schedule (WHODAS) 2.0 outcomes.
• To utilize passively collected digital biomarkers from clinical outpatients.

Main Methods:

• Collected one-month behavioral time-series data on physical and digital activity.
• Summarized data into 64 features using statistical measures.
• Applied sequential feature selection for each WHODAS 2.0 domain and used linear regression for prediction.

Main Results:

• Machine learning models achieved an average mean absolute percentage error of 19.5% across six WHODAS 2.0 domains.
• Prediction accuracy varied by domain, with the self-care domain performing best (14.86% error) and life activities worst (27.21% error).
• Key predictive features included distance traveled, time at home, walking time, exercise time, and vehicle time.

Conclusions:

• Machine learning methods can effectively assess functional health using passively collected mobile data.
• Feature selection yields interpretable features, enhancing model explainability for clinical application.
• This approach offers a feasible alternative to traditional, time-intensive functional assessments.