EEG-Based Motor Imagery Decoding Combining Channel Attention and Transformer

Brain-Computer Interface (BCI) technology facilitates direct communication between the human brain and external devices, offering novel avenues for interaction and rehabilitation. Among BCIs, Electroencephalogram (EEG)-based motor imagery (MI) is a key technique, allowing cognitive signals to be transformed into executable commands. Nevertheless, challenges in accurately decoding brain signals limit the broader application of this technology. In this study, we propose the CAT-CNN-TRFM network, designed for EEG-based MI classification. The network employs a channel attention mechanism to selectively enhance relevant EEG channels. A convolutional module is subsequently used to extract spatiotemporal features, followed by a Transformer encoder to model global dependencies through multi-head attention. Finally, a fully connected layer handles the classification of signals. On the BCI IV2a dataset, CAT-CNN-TRFM achieves an average accuracy of 81.25%. Ablation studies and hyperparameter analysis confirm the essential contribution of the channel attention mechanism, Transformer module, and data augmentation strategy in optimizing the model's performance. These results highlight the promise of CAT-CNN-TRFM for addressing complex EEG signal classification challenges.