The advancement of machine learning and deep learning systems has led to numerous innovative solutions for evaluating physical rehabilitation using sensor data. Artificial intelligence (AI)-based methods offer significant advantages over traditional approaches, including greater flexibility, enhanced patient convenience, and reduced treatment costs by minimizing the need for in-person medical visits. However, previous studies employing conventional neural networks (NNs) often failed to effectively capture the relational structure among nodes. To overcome this drawback, we suggest employing GCNs (Graph Convolutional Networks) to leverage the structural connections among nodes on the human body. The findings from our experiments show that the GCN model not only performs better but also decreases the parameter count from 5.6 million to under 160,000, all while maintaining improved accuracy. These findings underscore its potential for deployment in compact devices, mobile platforms, and commercial applications.