A dynamic multi-branch EEG decoding network for motor imagery classification with preliminary clinical validation.

Motor imagery electroencephalography (MI-EEG) decoding remains challenging because of the low signal-to-noise ratio, non-stationarity, and inter-subject variability of EEG signals. This study proposes a dynamic multi-branch EEG decoding network (DMB-EDN) that jointly models temporal dynamics, learnable time-frequency patterns, and rhythm-specific spectral information. DMB-EDN combines a learnable Gabor-based time-frequency representation with physiologically guided rhythm modeling and employs trial-conditioned dynamic fusion to estimate the contribution of each branch separately for each EEG trial. This design enables adaptive coordination of complementary data-driven and physiology-guided representations. The proposed method was evaluated on the BCI Competition IV 2a dataset, the High Gamma Dataset, and a self-collected spinal cord injury (SCI) dataset. Under subject-specific evaluation, DMB-EDN achieved an average accuracy of 96.41% and a kappa of 0.952 on BCI Competition IV 2a. On the High Gamma Dataset, it achieved performance comparable to the strongest baseline under near-saturated conditions. Under leave-one-subject-out evaluation on the SCI dataset, DMB-EDN obtained an accuracy of 85.00% and a kappa of 0.700, providing preliminary evidence of improved offline cross-subject decoding. Ablation experiments confirmed the complementary contributions of the three representation branches and trial-conditioned fusion, while fusion-weight analysis revealed systematic class- and oscillation-related variations. These results demonstrate the effectiveness of DMB-EDN for EEG decoding, although validation on larger multicenter cohorts and prospective online BCI systems remains necessary.