Dynamic functional connectome predicts individual working memory performance across diagnostic categories.

Working memory impairment is a common feature of psychiatric disorders. Although its neural mechanisms have been extensively examined in healthy subjects or individuals with a certain clinical condition, studies investigating neural predictors of working memory in a transdiagnostic sample are scarce. The objective of this study was to create a transdiagnostic predictive working memory model from whole-brain functional connectivity using connectome-based predictive modeling (CPM), a recently developed machine learning approach. Resting-state functional MRI data from 242 subjects across 4 diagnostic categories (healthy controls and individuals with schizophrenia, bipolar disorder, and attention deficit/hyperactivity) were used to construct dynamic and static functional connectomes. Spatial working memory was assessed by the spatial capacity task. CPM was conducted to predict individual working memory from dynamic and static functional connectivity patterns. Results showed that dynamic connectivity-based CPM models successfully predicted overall working memory capacity and accuracy as well as mean reaction time, yet their static counterparts fell short in the prediction. At the neural level, we found that dynamic connectivity of the frontoparietal and somato-motor networks were negatively correlated with working memory capacity and accuracy, and those of the default mode and visual networks were positively associated with mean reaction time. Moreover, different feature selection thresholds, parcellation strategies and model validation methods as well as diagnostic categories did not significantly influence the prediction results. Our findings not only are coherent with prior reports that dynamic functional connectivity encodes more behavioral information than static connectivity, but also help advance the translation of cognitive "connectome fingerprinting" into real-world application.