Motor imagery (MI)-elicited electroencephalogram (EEG) signals are widely used in brain-computer interface (BCI) systems. However, decoding MI-EEG remains challenging due to its low signal-to-noise ratio, nonstationarity, and complex temporal dynamics. To address these issues, this paper proposes a novel MI-EEG classification model, EEG-ConvMamba, which integrates multi-branch convolutional neural networks (CNNs) with a Mamba module. The CNN is used to extract local spatiotemporal features, suppress noise, and reduce sequence redundancy, while the Mamba module, based on the state space model (SSM), efficiently captures long-range temporal dependencies, followed by a fully connected layer for classification. The proposed method was evaluated on three public MI-EEG datasets, BCI-IV-2a, High-Gamma, and OpenBMI, under a subject-dependent evaluation protocol. Experimental results show that EEG-ConvMamba achieved average accuracies of 80.06%, 97.09%, and 72.26% on the three datasets, respectively. Compared with current state-of-the-art (SOTA) methods, the proposed model achieved the best performance and demonstrated statistically significant improvements in the vast majority of comparisons. In addition, Smooth Grad-CAM, a BCI signal visualization method, was employed to interpret the learned features and further enhance the interpretability of the proposed model. The source code is available at https://github.com/Jianxi-Huang/EEG-ConvMamba .
Huang et al. (Tue,) studied this question.