UnEBOLT: A Unified Model for EEG-to-BOLD Translation and Functional Connectivity Reconstruction

Abstract

Functional magnetic resonance imaging (fMRI) provides high-resolution, whole-brain dynamic information, but is costly and immobile, limiting its utility in low-resource settings. EEG-to-fMRI translation via deep learning offers a promising alternative, enabling access to deep brain activity from scalp EEG signals in naturalistic settings. However, current state-of-the-art methods for EEG-to-fMRI translation require training separate models for each brain region, limiting efficiency and scalability. Here, we introduce UnEBOLT, a Unified model for EEG-to-BOLD Translation. UnEBOLT is an end-to-end framework that predicts whole-brain fMRI time series from EEG by adaptive multi-region decoding within a single model. This approach enables efficient and comprehensive inference while also reconstructing subject-specific functional connectivity matrices, a representation that provides insight into neuronal interactions and which has been successfully utilized for clinical biomarkers. Our results show that UnEBOLT achieves comparable performance to dedicated ROI-specific models while scaling to multi-region prediction. Additionally, the reconstructed fMRI time series enable functional connectivity estimation, which may have broad applications in neuroscience.