Comparing Task-Relevant Information Across Different Methods of Extracting Functional Connectivity

Abstract

The concept of brain states, functionally relevant large-scale activity patterns, has become popular in neuroimaging. Not all components of such patterns are equally characteristic for each brain state, but machine learning provides a possibility for extracting and comparing the structure of brain states from functional data. However, their characterization in terms of functional connectivity measures varies widely, from cross-correlation to phase coherence, and the idea that different measures provide similar or coherent information is a common assumption made in neuroimaging. Here, we compare the brain state signatures extracted from of phase coherence, pairwise covariance, correlation, regularized covariance and regularized precision for a dataset of subjects performing five different cognitive tasks. In addition, we compare the classification performance in identifying the tasks for each connectivity measure. The measures are evaluated in their ability to discriminate the five tasks with two types of cross-validation: within-subject cross-validation, which reflects the stability of the signature over time; and between-subject cross-validation, which aims at extracting signatures that generalize across subjects. Secondly, we compare the informative features (connections or links between brain regions/areas) across measures to test the assumption that similar information is obtained about brain state signatures from different connectivity measures. In our results, the different types of cross-validation give different classification performance and emphasize that functional connectivity measures on fMRI require observation windows of sufficient duration. Furthermore, we find that informative links for the classification, meaning changes between tasks that are consistent across subjects, are entirely uncorrelated between BOLD correlations and covariances. These results indicate that the corresponding FC signature can strongly differ across FC methods used and that interpretation is subject to caution in terms of subnetworks related to a task.