Real-Time EMG Decomposition Across Neural Excitation Levels Using Dual Self-Attention Residual Network

Abstract

Motor unit (MU) discharge information extracted via real-time electromyogram (EMG) decomposition shows superiority in dexterous finger motion decoding. The variation of excitation levels can, however, lead to MU recruitment/de-recruitment, resulting in nonstationary EMG activities and then degraded decomposition and decoding performance; therefore, a novel online decomposition approach based on the multiple separation vector strategy was developed in this study. First, the separation vectors corresponding to different excitation levels were extracted offline via the fast independent component analysis (FastICA) algorithm and then merged to construct the separation vector pool for each MU via a previous MU action potential classification network. Under the online condition, a dual self-attention residual network was proposed to identify the excitation level, and the separation vectors were alternated correspondingly [termed alternating strategy (AS) method]. The conventional method that always used a fixed separation (FS) vector was compared. The 30-min synthetic EMG and the 15-min experiment EMG with the neural drive and the contraction strength, respectively, varying between 0% and 45% maximum voluntary contraction (MVC) were used. The experiment involved dexterous multifinger extension with isometric contractions. The results showed that the AS method obtained a higher spike consistency (87.55% $\pm ~3.75$ % versus 84.05% $\pm ~4.11$ %) with the true spike trains using the synthetic EMG and improved force prediction performance using the experiment EMG, i.e., a higher correlation ( $R^{2}$ : $0.82~\pm ~0.06$ versus $0.76~\pm ~0.07$ ) and a lower prediction error root-mean-square error (RMSE): 8.87% $\pm ~1.53$ % versus 13.61%MVC $\pm ~0.92$ %MVC) compared with the FS method. Further development of the proposed method could potentially provide a robust humanmachine interface for dexterous finger force prediction in realistic applications.