Uncovering Dynamic Neural Information Flow with Continuous-Time Weighted Dynamic Bayesian Networks

Abstract

Understanding how information dynamically flows within neural systems is a crucial problem in neuroscience. Traditional approaches often assume stationary or quasi-stationary functional networks, which fail to capture the time-varying dynamics of interactions among neural variables. To address this limitation, we introduce Continuous-Time weighted Dynamic Bayesian Networks (CTwDBN), a non-stationary graphical modeling framework for uncovering smoothly time-varying conditional dependencies. Validation on synthetic datasets demonstrated that CTwDBN reliably recovers the structure and dynamics of ground-truth information flow. Application to electrophysiological recordings during a guided saccade task revealed temporal fluctuations in conditional dependencies in the cortical network that persisted an order of magnitude longer than the receptive field dynamics. In the resting-state cortex, CTwDBN revealed persistent fluctuations within a low-dimensional dependency space reflecting canonical anatomical motifs. These results highlight CTwDBN as a versatile analytical framework for capturing dynamic information flow in neural systems with broad applicability to complex biological and artificial systems.