HADANet: hybrid attentive domain adaptation for cross-subject motor imagery EEG decoding

Motor imagery (MI) based brain-computer interfaces (BCIs) enable direct decoding of human motor intentions from neural activity. Electroencephalography (EEG) is widely used for MI decoding due to its non-invasive nature and high temporal resolution. However, large inter-subject variability in EEG signals leads to significant distribution discrepancies across subjects, which limits the generalization ability of existing decoding models. To address this issue, we propose a Hybrid Attentive Domain Adaptation Network (HADANet) for cross-subject motor imagery EEG decoding. The proposed framework employs a hierarchical convolutional feature extractor to capture complementary temporal and spectral characteristics of EEG signals. A hybrid attention mechanism further enhances discriminative spatial and channel-wise neural representations. In addition, a hybrid domain adaptation strategy combining adversarial learning and multi-kernel maximum mean discrepancy (MMD) alignment is introduced to reduce inter-subject distribution discrepancies and learn domain-invariant features. Experiments on the PhysioNet and Cho motor imagery datasets demonstrate that HADANet achieves competitive performance compared with several state-of-the-art methods, obtaining average accuracies of 82.85% and 85.87%, respectively. The results demonstrate that our framework effectively models motor imagery-related neural patterns and improves cross-subject generalization for practical BCI systems. the code in public https://github.com/curiouspeople/HADANet-.