Identifying dynamic reproducible brain states using a predictive modelling approach

Abstract

AbstractConceptually brain states reflect some combination of the internal mental process of a person, and the influence of their external environment. Importantly, for neuroimaging, brain states may impact brain-behavior modeling of a person’s traits, which should be independent of moment-to-moment changes in behavior. A common way to measure both brain states and traits is to use functional connectivity based on functional MRI data. Brain states can fluctuate in time periods shorter than a typical fMRI scan, and a family of methods called dynamic functional connectivity analyses, have been developed to capture these short time estimates of brain states. There has been a rise in the use of dynamic functional connectivity in order to find temporally specific spatial patterns of connectivity which reflect brain states, that can yield further insight into traits and behaviors. It has previously been shown that brain state can be manipulated through the use of continuous performance tasks that put the brain in a particular configuration while the task is performed. Here we focus on moment-to-moment changes in brain state and test the hypothesis that there are particular brain-states that maximize brain-trait modeling performance. We use a regression-based brain-behavior modelling framework, Connectome-based Predictive Modelling, allied to a resample aggregating approach, to identify behavior and trait related short time brain states, as represented by dynamic functional connectivity maps. We find that there is not a particular brain state that is optimal for trait-based prediction, and drawing data from across the scan is better. We also find that this not the case for in-magnet behavioral prediction where more isolated and temporally specific parts of the scan session are better for building predictive models of behavior. The resample aggregated dynamic functional connectivity models of behavior replicated within sample using unseen HCP data. The modelling framework also showed success in the estimating variance behavior in the ABCD dataset when using data from that dataset. The method detailed here may prove useful for both the study of behaviorally related brain states, and for short time predictive modelling.