CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients
Researchers have introduced CFSPMNet, a novel cross-patient adaptation framework designed to improve motor imagery electroencephalography (MI-EEG) decoding for post-stroke rehabilitation. Addressing the challenge of pathological neural reorganization that makes source-learned representations unreliable for unseen patients, the model combines a Fourier-Reorganized State Mamba Network (FRSM) with Shared-Private Prototype Matching (SPPM). FRSM reorganizes latent physiological token states in the Fourier domain to guide state-space propagation, while SPPM enhances pseudo-label updating by filtering physiologically inconsistent predictions. Experimental results on two stroke MI-EEG datasets demonstrate that CFSPMNet significantly outperforms existing CNN, Transformer, and Mamba-based baselines. The framework achieved average accuracies of 68.23% on the XW-Stroke dataset and 73.33% on the 2019-Stroke dataset, representing substantial improvements over previous methods. These findings suggest that modeling latent neural-state organization can effectively enhance cross-patient Brain-Computer Interface (BCI) decoding, offering a promising non-invasive route for personalized stroke rehabilitation technologies. The source code for the framework has been made publicly available.
Wire timeline
CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients
Researchers have introduced CFSPMNet, a novel cross-patient adaptation framework designed to improve motor imagery electroencephalography (MI-EEG) decoding for post-stroke rehabilitation. Addressing the challenge of pathological neural reorganization that makes source-learned representations unreliable for unseen patients, the model combines a Fourier-Reorganized State Mamba Network (FRSM) with Shared-Private Prototype Matching (SPPM). FRSM reorganizes latent physiological token states in the Fourier domain to guide state-space propagation, while SPPM enhances pseudo-label updating by filtering physiologically inconsistent predictions. Experimental results on two stroke MI-EEG datasets demonstrate that CFSPMNet significantly outperforms existing CNN, Transformer, and Mamba-based baselines. The framework achieved average accuracies of 68.23% on the XW-Stroke dataset and 73.33% on the 2019-Stroke dataset, representing substantial improvements over previous methods. These findings suggest that modeling latent neural-state organization can effectively enhance cross-patient Brain-Computer Interface (BCI) decoding, offering a promising non-invasive route for personalized stroke rehabilitation technologies. The source code for the framework has been made publicly available.
cs.AI updates on arXiv.org