A Deep Learning Framework for Efficient Online Decomposition of High-Density Surface Electromyogram into Motor Unit Spike Trains

The efficient online decomposition of high-density surface electromyogram (HD-sEMG) into motor unit spike trains (MUSTs) remains a substantial challenge. In this study, a novel deep learning (DL) framework is presented. The framework leverages a hybrid DL model to capture spatiotemporal features from HD-sEMG for fast and accurate MU spike identification. Then, the continuous MUSTs are precisely reconstructed through a post-processing approach. Besides, an asymmetric loss is integrated to improve the capability of DL model in addressing imbalanced samples. Both simulated and experimental HD-sEMG data were employed to develop and validate the framework. The offline progressive FastICA peel-off method was utilized to extract reference MUSTs to train and evaluate the proposed framework. A public HD-sEMG dataset was adopted to validate the generalizability of the framework. Results demonstrated that the proposed framework achieved accurate online estimation of MUSTs from HD-sEMG with extremely low latency (approximately 7.5 ms), yielding a matching rate of 0.98 ± 0.02, 0.92 ± 0.02 and 0.90 ± 0.02 for simulated, experimental and public dataset, respectively, significantly outperforming baseline methods (p < 0.05). This study further enriches tools for online HD-sEMG decomposition, potentially promoting their real-time applications in fields of neurorehabilitation medicine and neural motor control.