Foundation Model Guided Dual-Branch Co-Adaptation for Source-Free EEG Decoding

Researchers from multiple institutions have introduced FUSED, a novel framework for source-free domain adaptation (SFDA) in EEG decoding. Traditional SFDA methods rely solely on the limited internal knowledge of source-pretrained models, leading to poor cross-domain generalization and unreliable pseudo-labels. FUSED addresses this by integrating a large-scale EEG foundation model (FM) pretrained on diverse data with a compact specialist model (SM) through a dual-branch co-adaptation mechanism. This allows both models to generate pseudo-labels from linear and prototype views, while a consensus filtering mechanism uses the FM's stability to identify high-quality samples. A two-stage pseudo-label refinement pipeline further suppresses error accumulation via cross-branch arbitration.

To ensure the FM adapts effectively to target data without retraining its massive parameters, FUSED first calibrates the FM's decision boundaries using mutual information maximization with the SM, then distills knowledge from the FM back to the SM. This calibrate-then-distill pipeline makes the pipeline both effective and computationally efficient. Extensive experiments across three EEG paradigms—motor imagery, emotion recognition, and SSVEP—demonstrate consistent state-of-the-art performance, validating the power of foundation-guided synergy for robust, privacy-preserving EEG decoding. This work marks the first successful integration of EEG foundation models within an SFDA framework, opening new avenues for cross-subject brain-computer interfaces.