Abstract
Accurate decoding of incomplete hand gestures from residual paretic surface electromyography (sEMG) remains challenging after stroke because voluntary myoelectric activity is often weak, non-stationary, and contaminated by pathological co-activation. This study proposes a Latent Synergy-Informed Purification Network (LSP-Net) for reference-guided reconstruction of residual paretic sEMG. Specifically, LSP-Net jointly models synchronous paretic- and non-paretic-side sEMG, using the non-paretic signal as both bilateral reference information and a paired waveform-level supervision target for latent residual compensation, without assuming it to represent the physiological ground truth of the paretic side. The reconstructed sEMG representations are then evaluated through a wavelet-integrated CNN-Transformer for incomplete hand gesture intention recognition. Nine participants with post-stroke hemiplegia at Brunnstrom stage III or higher performed nine bilateral hand and wrist movements. Under subject-specific calibration, LSP-Net achieved an average Pearson correlation coefficient of 0.955 and an NRMSE of 0.016, indicating close reconstruction consistency with the paired non-paretic supervision targets. Furthermore, the reconstructed representations achieved an average gesture-intention recognition accuracy of 93.62%. These results suggest that latent synergy-informed reconstruction can provide more discriminative and stable sEMG representations for subject-specific post-stroke intention decoding.