By name: cfspmnet-eeg-motor-imagery-stroke description: "CFSPMNet - Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients. Use when working with MI-EEG decoding, cross-subject BCI for stroke rehabilitation, Mamba-based EEG models, or Fourier-domain token reorganization for neural decoding." license: Complete terms in LICENSE.txt metadata: arxiv_id: "2605.10111" published: "2026-05-11" authors: "Xiangkai Wang, Yun Zhao, Dongyi He, Qingling Xia, Gen Li, Xinlai Xing, Yuchi Pan, Bin Jiang" tags: [eeg, motor-imagery, mamba, stroke-rehabilitation, bci, cross-subject]
CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients
arXiv:2605.10111 | Submitted 11 May 2026 | cs.LG, cs.AI, cs.CV
Core Concept
CFSPMNet addresses the challenge of cross-patient MI-EEG decoding for stroke rehabilitation. Pathological neural reorganization makes source-learned MI representations unreliable for unseen patients. CFSPMNet models post-stroke MI-EEG as latent neural-state organization, combining a Fourier-Reorganized State Mamba Network (FRSM) with Shared-Private Prototype Matching (SPPM) for robust cross-subject adaptation.
Key Insights
Fourier-Reorganized State Mamba Network (FRSM): Represents each trial as a latent physiological token sequence, reorganizes token states in the Fourier domain, and uses Fourier-derived trial context to guide Mamba state-space propagation. This captures both band-specific spectral patterns and cross-frequency interactions.
Shared-Private Prototype Matching (SPPM): Improves target-domain pseudo-label updating by combining semantic confidence with shared-private physiological consistency, filtering confident but physiologically inconsistent target predictions.
Leave-One-Subject-Out Results: Achieves 68.23% on XW-Stroke and 73.33% on 2019-Stroke datasets, outperforming CNN, Transformer, Mamba, and adaptation-based baselines with improvements of 5.63 and 8.25 percentage points.
Neurophysiological Interpretability: Ablation, sensitivity, feature-alignment, pseudo-label selection, and neurophysiological visualization analyses confirm that Fourier-domain token-state reorganization and calibrated pseudo-label updating contribute to the performance gains.
Method Components
Fourier-Reorganized State Mamba Network
- Encodes EEG trials as latent physiological token sequences
- Reorganizes token states in the Fourier domain to capture spectral structure
- Fourier-derived trial context guides Mamba state-space propagation
- Captures both band-specific and cross-frequency interactions
Shared-Private Prototype Matching
- Maintains shared prototypes (common across subjects) and private prototypes (subject-specific)
- Pseudo-label selection filters based on both semantic confidence and physiological consistency
- Prevents propagation of confident but neurophysiologically implausible predictions
Applications
- Stroke rehabilitation BCI with MI-EEG decoding
- Cross-subject EEG decoding where training data comes from different patients
- Mamba-based neural signal processing for time-series EEG
- Domain adaptation for pathological EEG affected by neural reorganization
- Fourier-domain EEG feature extraction for spectral representation learning
Activation Keywords
- CFSPMNet
- EEG motor imagery decoding
- Mamba EEG network
- Fourier-guided EEG
- cross-subject BCI stroke
- shared-private prototype matching
- stroke rehabilitation EEG
- Fourier domain token reorganization
- Mamba state space EEG
- MI-EEG cross-patient
References
- Wang et al. (2026). CFSPMNet: Cross-subject Fourier-guided Spatial-Patch Mamba Network for EEG Motor Imagery Decoding in Stroke Patients. arXiv:2605.10111