Directed Functional Connectivity by Variational Cross-mapping of Psychophysiological Variables

Abstract

AbstractClassic methods of exploratory time-lag methods of directed connectivity hinge on wide-sense stationarity, which restricts their application to resting-state or time-windowed connectivity. Thus, nonlinear temporal effects in task-evoked neural activity observed in fMRI-BOLD data may be underestimated. Instead of a combinatorial search through dynamic causal models to account for such effects, we propose a modification of variational cross-mapping methods. We take inspiration from psychophysiological interactions to determine a priori covariance functions of nonparametric generative Gaussian process models to compare evidence of directional influence between paired regions of interest. To validate the method, we determine Bayes factors for correct inferred directionality in simulated bidirectional Lorenz-Rossler systems or neurovascular signals evoked by neural networks. We demonstrate that the method maintains strong evidence of correct event-coupling direction with high sensitivity and minimal specificity sacrifice. We further validate the method on a well-known object response dataset, by assessing whether it uncovers previously documented effective connectivity findings on face processing from BOLD responses mapped onto cortical surface parcellations of ventrolateral cortical areas. Our results agree with current effective connectivity findings about face processing, suggesting vertices of the inferior occipital gyrus send information to the superior temporal sulcus and the fusiform cortex, without direct connectivity between the fusiform cortex and the superior temporal sulcus.