Spatially and Temporally Guided Bayesian Optimization for Brain Effective Connectivity Learning from fMRI and EEG Data

ICLR 2026 Conference Submission14977 Authors

Keywords: Brain Effective Connectivity, Bayesian Optimization, multimodal, Causal Learning, fMRI, EEG

TL;DR: STBO-EC

Abstract: Brain effective connectivity (EC) characterizes the causal and directional interactions among brain regions and plays a central role in understanding cognition and neurological disorders. Constructing EC networks from multimodal neuroimaging such as functional Magnetic Resonance Imaging (fMRI) and electroencephalography (EEG) is challenging, since most existing methods rely on feature concatenation or linear mapping, neglecting structural consistency and nonlinear cross-modal dynamics. In this work, we propose STBO-EC, a spatially and temporally guided framework for multimodal EC learning. First, we develop an anatomy-informed spatial alignment strategy that leverages known brain region coordinates to establish structurally consistent correspondences between EEG electrodes and fMRI regions. Second, we design a time-slice-based alignment and fusion mechanism to effectively bridge the temporal resolution gap between fast EEG activity and slow fMRI signals. Finally, to tackle the high dimensionality and nonlinear dependencies of fused multimodal data, we employ a low-rank parameterized Bayesian optimization method (DrBO), which enables efficient exploration of the exponential EC search space while providing uncertainty-aware inference. Experiments on two real EEG–fMRI datasets demonstrate that STBO-EC consistently outperforms state-of-the-art baselines across multiple evaluation metrics.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 14977