An Unsupervised Feature Learning Approach for Elucidating Hidden Dynamics in rs-fMRI Functional Network Connectivity.

Abstract

Dynamic functional network connectivity (dFNC) data extracted from resting state functional magnetic resonance imaging (rs-fMRI) recordings has played a significant role in characterizing the role that brain network interactions play in a variety of brain disorders and cognitive functions. dFNC analyses frequently use clustering methods to identify states of network activity. However, it is possible that these states are dominated by a few highly influential networks or nodes, which could obscure condition-related insights that might be gained from networks or nodes less influential to the clustering. In this study, we propose a novel feature learning-based approach that could contribute to the identification of condition-related activity in formerly less influential networks or nodes. We demonstrate the viability of our approach within the context of schizophrenia (SZ), applying our approach to a dataset consisting of 151 participants with SZ and 160 controls (HCs). We find that the removal of some connectivity pairs significantly affects the underlying states and magnifies the differences between participants with SZ and HCs in each state. Given our findings, we hope that our approach will contribute to the characterization and improved diagnosis of a variety of neurological conditions and functions. Clinical Relevance- Our approach could contribute to the characterization and diagnosis of many neurological conditions.