STDR-Net: A Spatio-Temporal Dependency Refinement Network for Motor Imagery EEG Decoding

Abstract

Background/Objectives: Motor imagery electroencephalography (MI-EEG) contains weak, non-stationary patterns that evolve across channels and time. Effective decoding requires local sensorimotor structure, trial-wide context, and ordered temporal evolution to be represented without discarding information from earlier processing stages. Methods: We propose STDR-Net, a Spatio-Temporal Dependency Refinement Network that recalibrates convolutional tokens, models their global interactions, restores the original local representation through a cross-stage residual bridge, and refines the fused sequence with a bidirectional long short-term memory network. STDR-Net was evaluated on BCI Competition IV-2a, BCI Competition IV-2b, and the PhysioNet EEG Motor Movement/Imagery dataset against nine reproduced baselines. The PhysioNet experiment used a pooled mixed-participant trial-level split rather than an unseen-participant evaluation. Results: STDR-Net obtained accuracies of 82.69%, 86.99%, and 59.76% and Cohen’s kappa values of 0.7692, 0.7398, and 0.46, respectively. It achieved the highest observed mean accuracy on IV-2a and competitive performance on IV-2b and PhysioNet. Bridge-specific, ordering, module, and augmentation ablations supported the proposed pathway on IV-2a, whereas the differences on IV-2b were small and not significant after multiplicity correction. Conclusions: The results support information-preserving dependency refinement as a competitive design for MI-EEG decoding, while also showing that its benefit depends on the dataset and evaluation setting.