MSHANet: A Multiscale Hybrid Attention Network for Motor Imagery EEG Decoding

Brain-computer interface (BCI) technology has significant applications in neuro rehabilitation and motor function restoration, especially for patients with stroke or spinal cord injury. Motor imagery electroencephalog-raphy (MI-EEG) is widely used in BCIs, but its nonlinear dynamics and inter-subject variability limit decoding accuracy. In this paper, a multiscale hybrid attention network (MSHANet) for MI-EEG decoding, which consists of spatiotemporal feature extraction (STFE), talking head self-attention (THSA), dynamic squeeze-and-excitation attention (DSEA), and a temporal convolutional network (TCN), is proposed. MSHANet was evaluated via within-subject experiments using BCI Competition IV Datasets 2a and 2b, as well as EEGMMID, achieving decoding accuracies of 83.56%, 89.75%, and 75.66%, respectively. In cross-subject experiments on the three datasets, the mode lattained accuracies of 69.93% on BCI-2a, 81.85% on BCI-2b, and 79.67% on EEGMMID. In addition, we propose an electrode spatial structure-aware encoder. This technique encodes the spatial positions of electrodes in the original data, enabling the model to obtain richer spatial electrode information at the input stage. In within-subject and cross-subject tasks on BCI-2a, this encoding improved the decoding performance by 2.83% and 2.91%, respectively. Visualization was also employed to elucidate feature distributions and the effec tiveness of its attention mechanisms. Experimental results demonstrate that MSHANet performs exceptionally well in MI-EEG decoding tasks and has high potential for clinical applications, particularly in neurorehabilitation and motor function reconstruction.