EDANet: Efficient domain-adaptive attention neural network for EEG classification of motor imagery

Brain-Computer Interface (BCI) is a cutting-edge technology enabling communication between the brain and external devices. However, due to the non-stationarity of electroencephalography (EEG) signals, which leads to domain discrepancies between source and target domains, and the difficulty in extracting robust features from low signal-to-noise ratio (SNR) EEG signals, the EEG-based BCIs face significant challenges. In this study, an efficient domain-adaptive attention neural network (EDANet) is proposed for motor imagery decoding. In this model, a domain-adaptive spatial filter and a bidirectional attention temporal convolutional module (Bi-ATCN) are proposed to extract more useful features. The domain-adaptive spatial filter reduces domain discrepancies by aligning covariance matrices of EEG signals across different sessions and enhances the overall SNR by emphasizing the importance of distinct electrode channels. Compared to conventional unidirectional temporal models, the proposed Bi-ATCN captures both forward and backward temporal dependencies, leading to richer temporal context modeling. Moreover, Bi-ATCN integrates an efficient bi-layer attention mechanism (EBAM) to further improve temporal feature representation. To evaluate the proposed approach, extensive experiments were conducted on two publicly available EEG datasets BCIC IV-2a and BCIC IV-2b, achieving competitive average classification accuracies of 84.11% and 86.03%, respectively. Compared to state-of-the-art models, EDANet demonstrates superior classification performance, highlighting its potential for enhancing the practical application of BCIs.