Assessment of Rehabilitation Exercises from Depth Sensor Data

Assessing the rehabilitation exercises are essential in the recovery and treatment of various musculoskeletal conditions following surgery. According to reports, over 90% of all rehabilitative exercise sessions are conducted in a home environment. As the number of patients grows, this method becomes prohibitively expensive. Providing technology support for home-based rehabilitation is an excellent approach to address this. The patient remains at home and does the exercises in front of the camera, with the footage or data being sent to the physician for comments on the exercises. In this paper, we propose two machine learning-based models to assess the quality of exercises where the data is captured by such kinect 3D sensors. The proposed models consist of a long short-term memory(LSTM) network which uses the time series skeletal data to predict the quality of the exercises. The first model uses the predefined features proposed by the physicians. For the second model, we extract features using graph convolutional network(GCN) on the skeletal data where each node represents a body part or joint in the body and the edges represent the connection between the body parts. We conclude that LSTM is more accurate at predicting the results when GCN features are used.