Deep Neural Network for Online Identification of Motor Unit Activities from High-Density Surface Electromyography

The deep learning-based (DL) methods have been recently reported to exhibit better robustness and real-time performance on online identifying motor unit (MU) activities from high-density surface electromyography (HD-sEMG). However, previous studies have only explored the effectiveness of temporal and spatial neural networks individually. In this study, we compare the performance of spatial and temporal neural networks to identify MU activities in real time. Both neural networks comprise the convolutional neural network (CNN) for extracting spatial information and the bidirectional gated recurrent unit (Bi-GRU) module for extracting temporal information, respectively. We further explored their performance under different training samples. Both simulated and experimental HD-sEMG recordings with an 8 × 8 electrode array from abductor pollicis brevis muscle were used to validate the methods. The MU spike trains identified offline using the automatic progressive FastICA peel-off algorithm were employed to train the neural network, which can then perform the online HD-sEMG decomposition on the unseen HD-sEMG data. Results demonstrated that the Bi-GRU method outperformed CNN method, achieving higher matching rate in both simulated and experimental data. Moreover, both methods achieved poor performance with less amount of training data. This study highlights the importance of exploiting temporal information of HD-sEMG and the amount of training data for DL-based MU identification, providing valuable insights into optimizing model design and improving decomposition performance.