Stacked LoRA Boosts EEG Foundation Models for Motor Imagery Decoding

Brain-computer interfaces (BCIs) using electroencephalography (EEG) face a significant hurdle: the wide variation in brain signals between individuals. This variability often forces the use of subject-specific models, requiring extensive recalibration for each user. Recent advancements in EEG foundation models offer a promising path, but they still need further adaptation to be truly effective for specific tasks. This paper introduces Stacked LoRA, a novel adaptation framework designed to address this challenge for motor imagery classification. It structurally separates universal knowledge from individual neural patterns by using both a global adapter, trained across all subjects, and subject-specific adapters for individual variability. The study compared three adaptation strategies: purely subject-specific LoRA, global LoRA, and the proposed Stacked LoRA. Across multiple benchmark datasets, Stacked LoRA consistently delivered the highest accuracy, particularly demonstrating its strength in clinical scenarios where signal variability is high. The findings suggest that while a shared adapter suffices for large, diverse groups, subject-specific adaptation is crucial for complex clinical data.