Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Directed connectivity mapping through a self-attention mechanism in the human brain foundation model

Myeonggyo Jeong¹; ¹Sung Kyun Kwan University

Presenter: Myeonggyo Jeong

Brain foundation models pretrained on large-scale human fMRI data recently demonstrate an unprecedented performance in predicting neural signals across diverse brain states, yet what these models learn about functional dynamics remains largely unknown. Here, we leveraged a ‘self-attention (SA)’ mechanism, a computational core of these models to yield a spatiotemporally weighted sum of past brain signals to explain that of a targeted brain area, for mapping of their connectome directionality (effective connectivity; EC). We have thoroughly validated accuracy of our framework in both simulation and empirical settings. In simulation with known ground-truth directed networks, SA-derived EC outperformed classical algorithms (VAR, GC, rDCM) across all evaluation metrics. Next, applying our method to empirical fMRI (HCP) revealed hierarchically characteristic EC patterns in the human brain, suggesting that the SA of brain foundation models could be reformulated as an efficient in-vivo marker for a non-linear mapping of directed brain interactions.

Topic Area: Methods, Tools, Theory & Neural Coding