Multivariate Time-Lagged Multidimensional Pattern Connectivity (mvTL-MDPC) for EEG/MEG Functional Connectivity Analysis

Abstract

Multidimensional connectivity methods are critical to reveal the full pattern of complex interactions between brain regions over time. However, to date only bivariate multidimensional methods are available for time-resolved EEG/MEG data, which may overestimate connectivity due to the confounding effects of spurious and indirect dependencies. Here, we introduce a novel functional connectivity method which is both multivariate and multidimensional, Multivariate Time-lagged Multidimensional Pattern Connectivity (mvTL-MDPC), to address this issue in time-resolved EEG/MEG applications. This novel method extends its bivariate counterpart TL-MDPC to estimate how well patterns in an ROI 1 at time point t1 can be linearly predicted from patterns of an ROI 2 at time point t2 while partialling out the multivariate contributions from other brain regions. We compared the performance of mvTL-MDPC and TL-MDPC on simulated data designed to test their ability to identify true direct connections, using the Euclidean distance to the ground truth to measure goodness-of-fit. These simulations demonstrate that mvTL-MDPC produces more reliable and accurate results than the bivariate method. We therefore applied this method to an existing EEG/MEG dataset contrasting words presented in more or less demanding semantic tasks, to identify the dynamic brain network underlying controlled semantic cognition. As expected, mvTL-MDPC was more selective than TL-MDPC, identifying fewer connections, likely due to a reduction in the detection of spurious or indirect connections. Dynamic connections were identified between bilateral anterior temporal lobes, posterior temporal cortex and inferior frontal gyrus, in line with recent neuroscientific models of semantic cognition.