AUTHOR=Lin Chuang , Chen Chen , Cui Ziwei , Zhu Xiujuan TITLE=A Bi-GRU-attention neural network to identify motor units from high-density surface electromyographic signals in real time JOURNAL=Frontiers in Neuroscience VOLUME=18 YEAR=2024 URL=https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2024.1306054 DOI=10.3389/fnins.2024.1306054 ISSN=1662-453X ABSTRACT=

To utilize surface electromyographics (sEMG) for control purposes, it is necessary to perform real-time estimation of the neural drive to the muscles, which involves real-time decomposition of the EMG signals. In this paper, we propose a Bidirectional Gate Recurrent Unit (Bi-GRU) network with attention to perform online decomposition of high-density sEMG signals. The model can give different levels of attention to different parts of the sEMG signal according to their importance using the attention mechanism. The output of gradient convolutional kernel compensation (gCKC) algorithm was used as the training label, and simulated and experimental sEMG data were divided into windows with 120 sample points for model training, the sampling rate of sEMG signal is 2048 Hz. We test different attention mechanisms and find out the ones that could bring the highest F1-score of the model. The simulated sEMG signal is synthesized from Fuglevand method (J. Neurophysiol., 1993). For the decomposition of 10 Motor Units (MUs), the network trained on simulated data achieved an average F1-score of 0.974 (range from 0.96 to 0.98), and the network trained on experimental data achieved an average F1-score of 0.876 (range from 0.82 to 0.97). The average decomposition time for each window was 28 ms (range from 25.6  ms to 30.5 ms), which falls within the lower bound of the human electromechanical delay. The experimental results show the feasibility of using Bi-GRU-Attention network for the real-time decomposition of Motor Units. Compared to the gCKC algorithm, which is considered the gold standard in the medical field, this model sacrifices a small amount of accuracy but significantly improves computational speed by eliminating the need for calculating the cross-correlation matrix and performing iterative computations.