Robust decoding for MI-EEG: a hybrid transformer network using multi-perspective collaborative attention and dynamic...

Deep learning methods have been extensively applied in brain-computer interface (BCI) systems based on motor imagery (MI) for decoding electroencephalogram (EEG) signals. However, most existing hybrid architectures often struggle to effectively eliminate redundant noise in multi-channel signals and lack adaptability to the inherent non-stationarity and distribution drift of EEG signals. This work proposes a novel end-to-end hybrid attention Transformer network (HATNet) for EEG classification. HATNet first employs a convolutional neural network to extract local spatio-temporal features. To overcome the limitations of existing models, it fuses a Collaborative Attention Mechanism for Lightweight Channels, which dynamically recalibrates feature channels through multidimensional pooling strategies, including entropy pooling, to achieve precise spatial noise suppression. Addressing the non-stationary nature of EEG signals, an innovative Dynamic Hyperbolic Tangent module drives the Transformer encoding layer, adapting in real-time to data distribution drifts and significantly enhancing the model's ability to capture individual variations. Furthermore, cross-layer residual fusion pathways deeply integrate global contextual features with raw local spatio-temporal features. To ensure clear scope definition, experiments explicitly distinguish between primary MI tasks and auxiliary motor execution (ME) tasks. HATNet's performance was evaluated on three primary MI benchmark datasets, namely BCIC-IV-2a, BCIC-IV-2b, and the large-scale OpenBMI, as well as one auxiliary ME dataset, HGD. Experimental results demonstrate that HATNet achieves state-of-the-art performance across all analyses. In subject-dependent evaluations, average accuracy rates reached 81.25%, 86.65%, and 69.57% on the three primary MI datasets respectively, and 96.20% on the auxiliary ME dataset. Furthermore, in subject-independent evaluations, it achieved 60.88%, 80.79%, and 76.28% on the MI datasets respectively, alongside 73.95% on the ME dataset. Through multidimensional feature selection and dynamic adaptive modeling, HATNet exhibits superiority and robustness in enhancing both MI and ME decoding performance.