SCATNet: A Sinc-CNN-Attn-Transformer Framework for Motor Imagery EEG Classification

This paper proposes SCATNet, an end-to-end framework for decoding motor imagery (MI) EEG signals. SCATNet integrates a parameterized Sinc-convolution front end, bandgrouped spatial convolutions, the Convolutional Block Attention Module (CBAM), and a standard Transformer encoder to achieve synergistic spectral interpretability and long-range dependency modeling. The model was evaluated on the BCI-IV-2a and BCII-V-2b datasets, achieving average accuracies of 80.48% and 87.08%, respectively, with Cohen’s kappa values of 0.7397 and 0.7415, outperforming several representative baseline methods. Experimental results demonstrate that the proposed approach effectively enhances the discriminative performance and robustness of MI-EEG signals while maintaining a lightweight architecture.