Probing EEG Foundation Model Attention Maps for Lightweight Interpretable Emotion Recognition

Foundation Models (FMs) trained on EEG signals promise subject- and dataset-independent pipelines for brain-computer interface (BCI) applications, but have yet to consistently outperform task-specific architectures, and their large representations are costly to compute and store. As a means of advancing the understanding of the Foundation Models' inner mechanisms, we propose a lightweight alternative that exploits the self-attention mechanism of REVE, a pretrained EEG FM, to derive features similar to weighted Functional Connectivity (FC) without any fine-tuning. Inspired by kernel-based FC matrices, we aggregate the raw first-layer attention weights of a single head over the full duration of dataset trials into weighted channel-to-channel adjacency matrices. These matrices replace standard encoder representation vectors as classification features, substantially reducing computational cost. We further introduce a label-free, symmetry-based criterion, motivated by the emergence of near-symmetric attention patterns in trained encoder-decoder transformers, ranking attention heads by how much discriminative information they carry. Classical machine learning classifiers are then trained on the single most informative head's attention-derived FC matrix and achieve balanced accuracy and F1 scores on the SEED-IV emotion-recognition dataset that are competitive with several fully fine-tuned state-of-the-art EEG Foundation Models, requiring at least two orders of magnitude fewer calculations than the standard use of the encoder. Subsequent cortical region analysis revealed a significant over-representation of frontal region connections by ranking the most discriminative FC matrix connectivity edges, validating our methods as consistent with prior affective-computing literature. These results suggest that interpretable, computationally inexpensive linear probing of attention maps is a viable route to deploying EEG Foundation Models in resource-constrained settings.