 Health professionals often prescribe patients to perform specific exercises for rehabilitation of various diseases such as stroke, Parkinson's disease, and back pain. When these exercises are performed without supervision, it becomes difficult to assess the accuracy of the patient's performance. Automated assessment of physical rehabilitation exercises aims to assign a quality score based on an RGBD video of the body movement. Recent deep learning approaches have addressed this problem by extracting CNN features from coordinate grids of skeletons obtained from videos. However, these approaches were unable to extract rich spatial temporal features from variable length inputs. In order to address this issue, we investigated graph convolutional networks, GCNs, for this task. We adapted spatio-temporal GCNs to predict continuous scores rather than discrete class labels. This allows us to process variable length inputs so that users can repeat the prescribed exercise multiple times. Additionally, our novel design incorporates self-attention of body joints which indicates their role in predicting assessment scores. Our model successfully outperformed existing exercise assessment methods on the Camor and UIPR-MD datasets. This article was authored by Swaqshar Deb, M.D. Fokr al-Islam, Shafin Rahman, and others. We are article.tv, links in the description below.