Individual-Specific fMRI-Subspaces Improve Functional Connectivity Prediction of Behavior

Abstract

There is significant interest in using resting-state functional connectivity (RSFC) to predict human behavior. Good behavioral prediction should in theory require RSFC to be sufficiently distinct across participants; if RSFC were the same across participants, then behavioral prediction would obviously be poor. Therefore, we hypothesize that removing common resting-state functional magnetic resonance imaging (rs-fMRI) signals that are shared across participants would improve behavioral prediction. Here, we considered 803 participants from the human connectome project (HCP) with four rs-fMRI runs. We applied the common and orthogonal basis extraction (COBE) technique to decompose each HCP run into two subspaces: a common (group-level) subspace shared across all participants and a subject-specific subspace. We found that the first common COBE component of the first HCP run was localized to the visual cortex and was unique to the run. On the other hand, the second common COBE component of the first HCP run and the first common COBE component of the remaining HCP runs were highly similar and localized to regions within the default network, including the posterior cingulate cortex and precuneus. Overall, this suggests the presence of run-specific (state-specific) effects that were shared across participants. By removing the first and second common COBE components from the first HCP run, and the first common COBE component from the remaining HCP runs, the resulting RSFC improves behavioral prediction by an average of 11.7% across 58 behavioral measures spanning cognition, emotion and personality.\n\nHighlightsO_LIWe decomposed rs-fMRI signals into common subspace & individual-specific subspace\nC_LIO_LICommon subspace is shared across all Human Connectome Project (HCP) participants\nC_LIO_LICommon subspaces are different across runs, suggesting state-specific effects\nC_LIO_LIIndividual-specific subspaces are unique to individuals\nC_LIO_LIRemoval of common subspace signals improve behavioral prediction by 11.7%\nC_LI