Please login to view this media
- Talk
- 15/09/2021
- Canada
Machine Learning Prediction of Shoulder Patient At-Home Physiotherapy With Inertial Sensors
Description
In this presentation, Phil Boyer, a PhD candidate at the University of Toronto's Institute of Biomedical Engineering, discusses innovative machine learning applications for predicting shoulder physiotherapy adherence at home, utilizing inertial sensors from smartwatches. Boyer highlights the prevalence of rotator cuff injuries, reinforcing the impotance of physiotherapy in rehabilitation, which often takes place outside clinical settings where adherence is known to be poor.
To address this, he introduces the Smart Physiotherapy Activity Recognition System (SPARS), designed to objectively measure patients' adherence to physiotherapy exercises at home. The system starts with patients performing exercises in a controlled environment while wearing smartwatches to gather labeled data. After these initial sessions, patients continue their rehabilitation at home, allowing the system to analyze their exercise patterns using machine learning algorithms.
The crux of the talk focuses on machine learning stages, particularly out-of-distribution (OOD) prediction, which distinguishes between exercise and non-exercise activities, and the subsequent classification of specific exercises. Boyer shares insights from their research with 42 patients and 19 different exercises, outlining their validation methods and innovative approaches to training the machine learning algorithms, including a novel proxy OOD dataset.
The findings reveal significant associations between quantified adherence measured through the SPARS system and patient recovery metrics, notably pain and disability reduction over time. Boyer concludes with several key takeaways: the effectiveness of using proxy data sets for training, the benefit of grouping similar exercises to avoid penalizing patients for classification errors, and the necessity of conducting machine learning training and validation in environments representative of real-world conditions.