CFSPMNet: EEG Decoding for Stroke MI

• CFSPMNet is a cross-patient adaptation framework for EEG motor imagery decoding in stroke rehabilitation that calibrates latent neural states.
• It combines a Fourier-Reorganized State Mamba Network (FRSM) for spectral token reorganization with a Shared-Private Prototype Matching (SPPM) module for robust pseudo-label filtering.
• Empirical results on stroke EEG datasets demonstrate significant improvements over current methods, underscoring its potential in clinical brain-computer interfaces.

Searching arXiv for CFSPMNet and closely related names to verify nomenclature and related papers. CFSPMNet, short for Cross-subject Fourier-guided Spatial-Patch Mamba Network, is a cross-patient adaptation framework for EEG motor imagery (MI) decoding in stroke patients. It was proposed for the clinically important setting in which a model is trained on EEG from several stroke patients and must decode MI from an unseen stroke patient with no target labels. The method is motivated by the claim that post-stroke transfer failure is not merely a generic feature-distribution shift, but a disturbance in latent neural-state organization involving task-related EEG dynamics, aperiodic activity, local excitability, cross-regional coordination, and trial-level context. To address this, CFSPMNet combines a Fourier-Reorganized State Mamba Network (FRSM) with Shared-Private Prototype Matching (SPPM), coupling trial-wise latent state modeling with pseudo-label filtering based on both semantic confidence and physiological consistency (Wang et al., 11 May 2026).

1. Clinical setting and problem formulation

CFSPMNet is situated in rehabilitation-oriented brain-computer interfaces (BCIs) that use MI-EEG as a non-invasive signal source. In this setting, stroke patients may still imagine movement even when voluntary movement is weak or absent, and decoded left/right hand imagery can be used to drive virtual reality, robotic assistance, neurostimulation, and closed-loop rehabilitation (Wang et al., 11 May 2026).

The central problem is cross-patient generalization. Stroke alters the aperiodic spectral background of EEG, the excitability of local neural populations, sensorimotor coordination across regions, and the trial-level state trajectory during MI. As a result, the same MI label may correspond to markedly different latent EEG organization across patients. CFSPMNet frames this as a problem of latent neural-state calibration rather than only statistical alignment.

A plausible implication is that the method targets a failure mode specific to stroke rehabilitation BCIs: a decoder may appear class-discriminative on source subjects while remaining unreliable on an unseen patient because the patient’s task-related neural states are organized differently. The model therefore seeks to preserve both class semantics and physiologically meaningful structure during adaptation.

2. Architectural organization and latent tokenization

CFSPMNet has two principal components: FRSM, which acts as the feature encoder, and SPPM, which governs target-domain pseudo-label updating (Wang et al., 11 May 2026). FRSM converts each EEG trial into a sequence of spatial-patch physiological tokens, reorganizes token states in the Fourier domain, and uses the resulting trial-specific Fourier context to guide a Mamba state-space model. SPPM then determines whether target predictions should be accepted as pseudo-labels by requiring both high classifier confidence and consistency with source-derived physiological prototypes.

The initial trial representation is defined as follows:

$$\begin{aligned} F_{\mathrm{temp}} &= \operatorname{Concat}(\mathcal{T}_1(X),\dots,\mathcal{T}_M(X)),\ F_{\mathrm{emb}} &= \mathcal{F}(\mathcal{S}(F_{\mathrm{temp}})),\ Z^{(0)} &= \operatorname{PE}\!\left(\sqrt{D}\,\Pi(F_{\mathrm{emb}})\right)\in\mathbb{R}^{L\times D}. \end{aligned}$$

Here, $\mathcal{T}_m(\cdot)$ denotes temporal branches, $\mathcal{S}(\cdot)$ spatial mapping, $\mathcal{F}(\cdot)$ temporal fusion, $\Pi(\cdot)$ patch projection, and $\operatorname{PE}$ positional embedding. The paper interprets each token as a compact observation of local neural population activity, while the token sequence captures how local states evolve into a global trial-level MI state.

Component	Stated role
FRSM	Feature encoder with Fourier-domain token-state reorganization and Mamba propagation
SPPM	Pseudo-label updating using semantic confidence plus physiological consistency

This organization is not presented as conventional spectral feature engineering. Instead, the Fourier domain is used to reorganize latent token states, and the adaptation mechanism explicitly rejects target samples that are confident yet physiologically mismatched.

3. Fourier-Reorganized State Mamba Network

Within each encoder block, FRSM first normalizes the token states, transforms them along the token dimension with real FFT, applies a learnable complex Fourier mixer with spectral filtering and sparsification, and then returns to the original domain (Wang et al., 11 May 2026):

$$\begin{aligned} \widetilde{Z}^{(\ell)} &= \operatorname{LN}(Z^{(\ell)}),\qquad \Omega^{(\ell)}=\operatorname{rFFT}((\widetilde{Z}^{(\ell)})^{\top}),\ \widehat{\Omega}^{(\ell)} &= \operatorname{SoftShrink}_{\tau_f}(\mathcal{M}_{f}(\Omega^{(\ell)}))\odot W_f+\Omega^{(\ell)},\ Z_{\mathrm{enh}}^{(\ell)} &= \operatorname{iFFT}(\widehat{\Omega}^{(\ell)})+(\widetilde{Z}^{(\ell)})^{\top}. \end{aligned}$$

The stated effect of this operation is twofold: it reorganizes token-state arrangement in the spectral domain and suppresses weak or noisy perturbations through spectral sparsification. The paper’s interpretation is that Fourier space provides a global complex-basis perspective on a trial’s latent state organization, including both periodic and aperiodic structure.

The reorganized spectrum is then partitioned by masks $\{M_b\}_{b=1}^{B}$ to form complementary spectral states and a Fourier-derived context:

$$\begin{aligned} Z_b^{(\ell)}&=\operatorname{iFFT}(\widehat{\Omega}^{(\ell)}\odot M_b)+(\widetilde{Z}^{(\ell)})^{\top},\ C^{(\ell)}&=\operatorname{LN}\!\left(\mathcal{H}([Z_1^{(\ell)},\dots,Z_B^{(\ell)}])+Z_{\mathrm{enh}}^{(\ell)}\right),\ S^{(\ell)}&=\sigma(W_s C^{(\ell)}),\qquad B^{(\ell)}=W_b C^{(\ell)}. \end{aligned}$$

This context conditions the Mamba branch:

$$\begin{aligned} [U^{(\ell)},R^{(\ell)}]&=\operatorname{Split}(W_{\mathrm{in}}\widetilde{Z}^{(\ell)}),\ \bar{U}^{(\ell)}&=U^{(\ell)}\odot(1+S^{(\ell)}),\qquad \bar{R}^{(\ell)}=R^{(\ell)}+B^{(\ell)},\ h_t^{(\ell)}&=\bar{A}_t^{(\ell)}h_{t-1}^{(\ell)}+\bar{B}_t^{(\ell)}u_t^{(\ell)},\ o_t^{(\ell)}&=\bar{C}_t^{(\ell)}h_t^{(\ell)}+D^{(\ell)}\odot u_t^{(\ell)},\ z_t^{(\ell,\ast)}&=W_{\mathrm{out}}^{(\ell)}\!\left(o_t^{(\ell)}\odot\operatorname{SiLU}(\bar{r}_t^{(\ell)})\right),\ Z^{(\ell+1)}&=Z^{(\ell)}+\operatorname{Drop}(Z^{(\ell,\ast)}). \end{aligned}$$

The paper’s key claim is that Mamba does not operate on an ordinary token stream; rather, its input and residual pathways are modulated by trial-specific Fourier context. This suggests that selective state-space propagation is being tailored to the current trial’s reorganized physiological state, which the authors present as the core mechanism of FRSM.

4. Shared-Private Prototype Matching and adaptation protocol

SPPM addresses a specific adaptation failure mode: in stroke EEG, a target prediction may be confident but wrong because it does not match the physiological structure of the predicted class (Wang et al., 11 May 2026). To counter this, the method constructs source-domain class prototypes in a private-signature space and accepts target pseudo-labels only when they satisfy both a semantic and a physiological criterion.

Source signatures are defined by

$\mathcal{T}_m(\cdot)$0

For each class $\mathcal{T}_m(\cdot)$1, the prototype and tolerance are computed as

$\mathcal{T}_m(\cdot)$2

Target pseudo-label calibration is then written as

$\mathcal{T}_m(\cdot)$3

A target sample is therefore accepted only if the classifier is sufficiently confident and its physiological signature lies within the predicted class’s acceptance envelope. The paper explicitly argues that this is important because stroke-induced reorganization can produce high-confidence predictions that remain physiologically inconsistent.

The adaptation protocol is leave-one-subject-out (LOSO). One stroke patient is held out as target, all remaining patients are source, training uses labeled source data and unlabeled target data, and the process is repeated so each subject becomes target once. Optimization is two-stage. In Stage I, the model is trained on labeled source data and prototypes are built. In Stage II, target probabilities, private signatures, and accepted pseudo-labels are recomputed and the following objective is optimized:

$\mathcal{T}_m(\cdot)$4

This establishes a calibration procedure in which the source domain stabilizes class structure and the target domain is incorporated only through calibrated pseudo-labels.

5. Empirical results, ablations, and analytical evidence

CFSPMNet was evaluated on two stroke MI datasets under LOSO evaluation (Wang et al., 11 May 2026). XW-Stroke comprises 24 clinically screened acute ischemic stroke patients used in the study subset, with 30 channels and 2-class MI. 2019-Stroke contains 15 stroke patients, 63 channels, and 2-class MI.

Dataset	CFSPMNet result	Strongest competitor
XW-Stroke	68.23% ± 5.13	SSTDA: 62.60% ± 5.57
2019-Stroke	73.33% ± 18.71	SSAS: 65.08% ± 20.12

The reported improvements are +5.63 percentage points on XW-Stroke and +8.25 percentage points on 2019-Stroke. The method also achieved the highest F1-score and kappa on both datasets: 66.78% F1 and 0.365 kappa on XW-Stroke, and 75.11% F1 and 0.467 kappa on 2019-Stroke. On XW-Stroke, the baseline comparisons were reported as statistically significant with $\mathcal{T}_m(\cdot)$5.

The ablation studies assign distinct roles to FRSM and SPPM. Removing SPPM reduced accuracy to 63.96% on XW-Stroke and 71.00% on 2019-Stroke. Removing dynamic pseudo-label updating reduced it further to 63.65% and 64.75%, respectively. Removing Fourier context guidance yielded 63.75% on XW-Stroke and 69.25% on 2019-Stroke. The paper also reports that removing either the high-frequency or low-frequency branch degrades performance, with the low-frequency branch being especially important on both datasets.

The analytical results extend beyond scalar accuracy. The t-SNE plots show a progression from scattered raw representations with weak class structure, through intermediate adaptation with better separation but ambiguous target clusters, to final CFSPMNet features that are compact, well separated, and aligned with source structure. The paper states that this suggests FRSM improves state organization and SPPM suppresses physiologically mismatched pseudo-labels. Block-level visualizations indicate that Fourier-derived context is more coherent than raw branch modulation, and pseudo-label selection visualizations show that accepted target samples lie where both the semantic confidence margin and the physiological consistency margin are positive. Topographic sensitivity maps further show non-uniform channel patterns, with dataset-specific differences consistent with patient-specific stroke neurophysiology.

6. Limitations, implications, and nomenclature

The paper explicitly notes several limitations of the current evidence base (Wang et al., 11 May 2026). Public stroke EEG datasets remain small; the XW-Stroke experiments use a clinically screened subcohort, not the full cohort; and standard deviations are relatively large, reflecting strong inter-patient variability. The datasets also lack full clinical metadata, so the model cannot explicitly incorporate lesion location, lesion volume, hemisphere, corticospinal tract involvement, stage of recovery, medication, rehabilitation history, or motor scores. The authors therefore identify multi-center and online validation as necessary future steps, and argue that reducing false feedback is clinically important because incorrect pseudo-labels in a closed-loop system may reinforce wrong intent decoding.

A plausible implication is that CFSPMNet’s main contribution is conceptual as much as architectural: it reframes cross-patient stroke MI decoding as neural-state calibration. In that view, successful transfer depends not only on aligning class-discriminative features, but also on preserving the physiological organization that makes those features meaningful for rehabilitation-oriented BCIs.

The name CFSPMNet should also be distinguished from several unrelated architectures with superficially similar acronyms. CFPN, the Cross-layer Feature Pyramid Network, is a salient object detection method based on direct cross-layer communication in feature pyramids (Li et al., 2020). CFPNet, the Channel-wise Feature Pyramid network, is a real-time semantic segmentation model built around dilated convolution channels (Lou et al., 2021). CSFPN, the Cascaded Sparse Feature Propagation Network, addresses point-based interactive segmentation through sparse graph propagation (Zhang et al., 2022). These methods operate in computer vision and address different tasks, whereas CFSPMNet is an EEG adaptation framework for stroke MI decoding.