Post-stroke Fine Hand Motion Intention Recognition Based on sEMG Decomposition and Residual Spiking Neural Networks

Fine motor dysfunction of the hand severely impacts activities of daily living in stroke survivors. Accurate decoding of motion intentions from surface electromyography (sEMG) is critical for enabling survivors to participate actively in robot-assisted rehabilitation. Motion intention recognition methods using motor unit spike trains (MUSTs) derived from sEMG decomposition have demonstrated superior performance compared to conventional sEMG-based methods. However, these methods inadequately leverage the inherent spatiotemporal sparse coding efficiency of MUSTs and the full potential of sEMG decomposition remains underutilized in post-stroke populations. This study proposes a hand motion intention recognition framework integrating sEMG decomposition with a residual spiking neural network (Res-SNN). sEMG signals were recorded from 14 neurotypical individuals and 7 stroke survivors performing 35 fine hand and wrist movements. The performance of Res-SNN was evaluated separately in neurotypical and post-stroke cohorts, and compared with a traditional sEMG-based deep residual network (ResNet) and a MUST-based convolutional SNN (CSNN). Results indicate that Res-SNN achieved classification accuracies above 0.95 for both cohorts, significantly surpassing those of ResNet (neurotypical: 0.84±0.08; post-stroke: 0.90±0.04). While Res-SNN showed comparable accuracy to CSNN in neurotypical subjects (0.99±0.01 vs. 0.96±0.08, P =0.48), it substantially outperformed CSNN in stroke survivors (0.95±0.03 vs. 0.71±0.16, P <0.001). Moreover, Res-SNN exhibited low inference power consumption (5.41 mJ·s). By integrating sEMG decomposition with Res-SNN, this study provides a high-accuracy and energy-efficient solution for post-stroke intention recognition, advancing the application of neural decoding technologies and neuromorphic computing in human-machine interfaces.