FA-STTM: A hybrid Transformer-Mamba network for motor imagery EEG classification

The brain–computer interface (BCI) offers transformative solutions for individuals with neural impairments, with motor imagery electroencephalogram (MI-EEG) widely applied in various BCI applications. However, the limited performance and generalizability of MI-EEG decoding remain significant challenges. In this paper, we propose a frequency-based attention spatial-temporal Transformer-Mamba network, named FA-STTM. We first design a wavelet-based spatial attention and a group-wise channel attention to emphasize task-relevant oscillation band features. Next, we develop a Transformer-Mamba module to extract global dependencies and accelerate training. Finally, we replace traditional fully connected layers with Kolmogorov–Arnold Networks to enhance representational power. The experimental results demonstrate that, in subject-dependent evaluations, our method achieves state-of-the-art performance both with data augmentation (BCICIV-2a: 82.9 %, BCICIV-2b: 87.81 %) and without data augmentation (BCICIV-2a: 77.89 %, BCICIV-2b: 86 %). Furthermore, in the challenging subject-independent evaluations, FA-STTM also demonstrates outstanding decoding performance (BCICIV-2a: 61.13 %, BCICIV-2b: 77.04 %). To the best of our knowledge, this is the first attempt to combine the Mamba and Transformer architectures in EEG decoding. The comparative studies demonstrate that the Transformer-Mamba design outperforms traditional Transformer architecture in terms of decoding accuracy, stability and training speed. Extensive qualitative and visualization experiments further highlight the superior performance and interpretability of our model.