Related Experiment Video
Updated: Aug 2, 2025
04:24
A Novel Digital Platform for a Monitored Home-based Cardiac Rehabilitation Program
Published on: April 19, 2019
11.6K
Evaluation of at-home physiotherapy
Philip Boyer1,2, David Burns3,4, Cari Whyne1,2,5
1Institute of Biomedical Engineering, University of Toronto, Toronto, Canada.
Bone & Joint Research
|April 13, 2023
Summary
Machine learning accurately tracks shoulder physiotherapy exercises using smartwatch data. Including non-exercise data improved detection, enhancing remote patient monitoring and rehabilitation outcomes.
Area of Science:
- Biomedical Engineering
- Rehabilitation Technology
- Machine Learning Applications
Background:
- Tracking adherence to at-home physiotherapy is crucial for patient engagement and rehabilitation success.
- Objective technological solutions are needed to monitor shoulder physiotherapy exercises effectively.
- Current methods face challenges in accurately capturing exercise data in diverse settings.
Purpose of the Study:
- To evaluate machine learning (ML) methodologies for detecting and classifying shoulder physiotherapy exercise data.
- To assess the performance of ML algorithms using inertial data from smartwatches in both clinic and home settings.
- To investigate the impact of data processing strategies on exercise detection and classification accuracy.
Main Methods:
- Collected inertial data from 42 patients performing shoulder physiotherapy exercises using a smartwatch.
- Employed a two-stage ML approach: out-of-distribution (OOD) detection for non-exercise data removal, followed by exercise classification.
- Evaluated strategies including exercise grouping by motion, incorporating non-exercise data in training, and patient-specific classification.
Main Results:
- A patient-specific approach with engineered features showed high in-clinic performance for exercise vs. non-exercise detection (AUROC = 0.924).
- Including non-exercise data in training significantly improved classifier performance (random forest, AUROC = 0.985).
- Highest in-clinic individual exercise classification accuracy was 0.903 using a patient-specific deep neural network approach; grouping exercises by motion type improved classification.
Conclusions:
- Machine learning, particularly with patient-specific approaches and inclusion of non-exercise data, shows strong potential for tracking shoulder physiotherapy adherence.
- Smartwatch-based inertial data collection offers a viable method for objective monitoring in both clinical and at-home settings.
- While patient-specific models show promise, their effectiveness is contingent on the quality of patient-provided data.
