Physiotherapy Exercise Classification with Single-Camera Pose Detection and Machine Learning.

Area of Science:

• Digital health
• Rehabilitation engineering
• Computer vision in healthcare

Background:

• Telehealth and virtual care are expanding, including physiotherapy services.
• Poor adherence to at-home physical therapy programs is a significant challenge.
• Objective tools for monitoring patient participation in remote exercise are lacking.

Purpose of the Study:

• To develop and assess an automated, unsupervised video-based system for monitoring at-home low-back and shoulder physiotherapy exercises.
• To utilize mobile phone cameras for objective assessment of exercise adherence.
• To evaluate the feasibility of using machine learning for physiotherapy monitoring.

Main Methods:

• Open-source pose detection framework to extract joint locations from videos.
• Convolutional neural network (CNN) trained on keypoint time series data to classify exercises.
• Analysis of model performance based on keypoint combinations and camera angle variations.

Main Results:

• The CNN model achieved high accuracy: 0.995 ± 0.009 for low-back and 0.963 ± 0.020 for shoulder exercises.
• Optimal performance was achieved using 12 pose estimation landmarks from the upper and lower body.
• Training the CNN with varied camera angles enhanced its robustness to filming perspective.

Conclusions:

• Smartphone cameras coupled with machine learning can effectively classify at-home physiotherapy participation.
• This technology offers a feasible, low-cost, and scalable solution for tracking adherence to physical therapy.
• The system has potential applications in various remote and supervised rehabilitation settings.