Shoulder physiotherapy exercise recognition: machine learning the inertial signals from a smartwatch.

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Machine Learning in Healthcare

Background:

• Physical therapy adherence is crucial for shoulder disorder management but often poor in home settings.
• Objective adherence measurement tools for unsupervised home exercise programs are limited.
• Smartwatch technology offers potential for remote monitoring of physiotherapy compliance.

Purpose of the Study:

• To develop and evaluate a smartwatch-based system for monitoring home shoulder physiotherapy.
• To assess the feasibility of using commercial smartwatches for objective adherence measurement.
• To apply machine learning for classifying physiotherapy exercises performed at home.

Main Methods:

• Twenty healthy adults performed rotator cuff exercises while wearing a smartwatch.
• Six-axis inertial sensor data was collected for exercise classification.
• Four supervised learning algorithms (k-NN, RF, SVC, CRNN) were trained and evaluated using cross-validation.

Main Results:

• All algorithms achieved >94% accuracy in temporally stratified cross-validation.
• The Convolutional Recurrent Neural Network (CRNN) achieved 99.4% accuracy.
• Subject-stratified cross-validation showed CRNN performing best with 88.9% accuracy on unseen subjects.

Conclusions:

• This study demonstrates the technical feasibility of using smartwatches and machine learning for home physiotherapy monitoring.
• The developed system can objectively assess adherence to shoulder exercise protocols.
• This approach has the potential to improve patient outcomes through enhanced remote rehabilitation.