Subject-Independent Deep Learning Framework for Motor Imagery Electroencephalogram Decoding in Neurorehabilitation

Abstract

Motor imagery (MI) has emerged as a pivotal paradigm in non-invasive brain-computer interfaces (BCIs) for neurorehabilitation, enabling motor function restoration through mental rehearsal of movements. However, traditional MI electroencephalogram (EEG) classification models face significant challenges due to high inter-subject variability and the expensive requirement of annotated EEG data for each new subject. To tackle these limitations, we introduce a deep learning framework, the Dual-branch Subject-aligned Generalization Network (DSGNet). DSGNet simultaneously extracts temporal and spectral EEG features through dual complementary convolutional branches and incorporates a novel class alignment loss to enforce domain-invariant representation across subjects, enabling generalization to unseen individuals without requiring subject-specific labeled data. We evaluate DSGNet on four public MI-EEG datasets—OpenBMI, BCI Competition IV 2a, SHU Version 5, and BCI Competition IV 2b—under a rigorous leave-one-subject-out cross-validation protocol. Experimental results show that DSGNet achieves the highest accuracy on the three-class and four-class datasets, with improvements of 0.22% and 2.15% over the strongest baselines, respectively, while maintaining comparable performance on the binary-class dataset. These findings highlight the effectiveness of class-structure alignment in developing reliable subject-independent BCI systems for neurorehabilitation.