SAST-KAN: A Spectral Adaptive Spatio-Temporal Kolmogorov–Arnold Network for Motor Imagery EEG Decoding

Abstract

Motor imagery–based brain–computer interface (MI-BCI) provides an effective pathway for post-stroke motor rehabilitation by decoding motor intentions from EEG signals. However, altered sensorimotor rhythms, nonlinear spatio-temporal dynamics, and limited clinical data make robust and generalizable post-stroke EEG decoding challenging. To address these challenges, we propose a novel end-to-end decoding framework termed Spectral Adaptive Spatio-Temporal Kolmogorov–Arnold Network (SAST-KAN). The model jointly optimizes three functionally coupled components under a shared end-to-end classification objective: (1) an adaptive frequency-band optimization module with differentiable boundary learning to perform subject-specific, task-optimized adjustment of frequency-band boundaries; (2) a ContMix1D-enhanced temporal encoder combining local convolution and global attention to model dynamic temporal dependencies; and (3) a gated EfficientKAN classifier that improves nonlinear representation while controlling model complexity. The proposed method was evaluated on a private stroke motor-attempt dataset containing seven patients and the public BCI Competition IV-2b dataset. Experimental results show that SAST-KAN achieved an average accuracy of 84.24% and F1-score of 84.71% on the stroke dataset, outperforming the evaluated baseline methods by 2.37–7.71 percentage points in mean accuracy. In the public dataset, it reached an average accuracy of 89.72%, achieving the highest mean accuracy among the evaluated methods. These experimental results demonstrate that SAST-KAN effectively captures subject-specific spectral and temporal features and exhibits robust performance across different datasets.