Simultaneous Estimation of Wrist Joint Angle and Torque during Isokinetic Contraction based on HD-sEMG

The establishment of a natural and smooth human-computer interface is crucial for myoelectric control, which requires an effective decoding method for movement intention. Based on high-density surface electromyography (HD-sEMG), this study explored a method to simultaneously estimate wrist joint angle and torque during isokinetic contraction. Ten able-bodied individuals were instructed to complete wrist isokinetic flexion and extension tasks with different movement patterns, and the HD-sEMG signals were collected. To decode these signals, a convolutional neural network (CNN) incorporating the global attention mechanism was established, named global attention convolutional neural network (GACNN). Six other decoding models were also used to continuously estimate the wrist joint angle and torque, including support vector machine (SVM), residual network (ResNet), long short-term memory (LSTM), transformer-based model (TBM), muscle synergy-based graph attention networks (MSGAT-LSTM), and spatio-temporal feature extraction network (STFEN). Evaluation metrics including normalized root mean square error (NRMSE) and Pearson’s correlation coefficient (PCC) were applied to evaluate the estimation performance of the seven models. The GACNN showed significantly better estimation performance than SVM, LSTM, ResNet, STFEN and it also demonstrated superior performance over TBM and MSGAT-LSTM in some estimation cases. On average, for all subjects, NRMSE and PCC of the GACNN were 0.080 ± 0.013 and 0.955 ± 0.016. The result shows the superiority of the neural network incorporating global attention mechanism, which is of great significance for the application of human-computer interaction.