A Deep Learning Based Assessment Method for Rehabilitation Exercises

The most common symptom of stroke is limb dysfunction, resulting in a reduced quality of life for the patient. Patients can be rehabilitated through specific exercises movements. However, when the exercises are performed without the presence of a healthcare provider, patients and families cannot be informed about the correctness of the exercises movements. Recent graph convolutional neural network (GCN) approaches address this problem by extracting features from a coordinate grid of skeletal data obtained from videos. The traditional GCN network cannot accurately reflect the spatial dependencies between nodes due to the fixed graph structure. In fact, the interactions between nodes may change with time and environment, so the fixed neighbor relationship cannot adapt to the dynamic change of graph structure. To address this problem, a hierarchical decomposed graph structure is proposed in this work for predicting continuous scores instead of discrete labels. In addition, this work introduces an attention-guided hierarchy aggregation module to better extract the features of the main hierarchical edge sets. The results of the simulation demonstrate that the proposed HDGCN-LSTM significantly outperforms the existing algorithms interms of MAD, RMS and MAPE.