Abstract
Decoding motor imagery (MI) electroencephalogram (EEG) signals, a keynon-invasive brain-computer interface (BCI) paradigm for controlling externalsystems, has been significantly advanced by deep learning. However, MI-EEGdecoding remains challenging due to substantial inter-subject variability andlimited labeled target data, which necessitate costly calibration for newusers. Many existing multi-source domain adaptation (MSDA) methodsindiscriminately incorporate all available source domains, disregarding thelarge inter-subject differences in EEG signals, which leads to negativetransfer and excessive computational costs. Moreover, while many approachesfocus on feature distribution alignment, they often neglect the explicitdependence between features and decision-level outputs, limiting their abilityto preserve discriminative structures. To address these gaps, we propose anovel MSDA framework that leverages a pretrained large Brain Foundation Model(BFM) for dynamic and informed source subject selection, ensuring only relevantsources contribute to adaptation. Furthermore, we employ Cauchy-Schwarz (CS)and Conditional CS (CCS) divergences to jointly perform feature-level anddecision-level alignment, enhancing domain invariance while maintaining classdiscriminability. Extensive evaluations on two benchmark MI-EEG datasetsdemonstrate that our framework outperforms a broad range of state-of-the-artbaselines. Additional experiments with a large source pool validate thescalability and efficiency of BFM-guided selection, which significantly reducestraining time without sacrificing performance.