Motor imagery EEG signal classification based on spatiotemporal self-attention mechanism and parallel CRNN framework

Motor imagery brain-computer interface (MI-BCI) allows subjects to control external devices through imagined actions, using multi-channel electrodes to capture brain activity. However, decoding motor imagery EEG signals (MI-EEG) is challenging due to their non-stationarity and low signal-to-noise ratio. Traditional feature extraction methods, limited to single-domain analysis, often yield low classification accuracy. This paper introduces a novel approach utilizing deep learning, applying spatial and temporal self-attention mechanisms to enhance feature extraction. The proposed model adapts spatial correlations between channels and temporal dependencies between time steps, fusing these spatiotemporal features through a parallel CRNN network. Tested on the BCI Competition IV dataset, the model achieved an average classification accuracy of 77.30%, outperforming Shallow ConvNet by 5.7% and EEGNet by 3%. These results demonstrate the model’s effectiveness and superiority in MI-EEG classification.