Uncovering functional connectivity patterns predictive of cognition in youth using interpretable predictive modeling

Functional MRI studies have identified functional connectivity (FC) patterns associated with behavioral traits using whole-brain or region-wise predictive models. However, whole-brain approaches often suffer from limited generalizability and interpretability due to the high-dimensionality of FC data. Conversely, region-wise models inherently isolate predictions, ineffective for characterizing contributions of the whole brain FC patterns in predicting a target trait. In this study, we propose an interpretable predictive model that learns fine-grained FC patterns predictive of behavioral traits, jointly at the regional and participant levels, to characterize the overall association of FC patterns with a target trait. Our model learns both a relevance score and a dedicated prediction model for each brain region, then integrates the regional predictions to generate a participant-level prediction, capturing the collective association of FC patterns with the trait. We validated our method using FC data from 6798 participants in the Adolescent Brain and Cognitive Development (ABCD) study for predicting cognition. Our interpretable predictive model identified the cingulo-parietal, retrosplenial-temporal, dorsal attention, salience, and cingulo-opercular networks as collectively predictive of cognitive traits. The interpretable model significantly improved prediction accuracy and facilitated the characterization of fine-grained differences in FC patterns across cognitive domains. Furthermore, the learned relevance scores enhanced region-wise predictions of longitudinal cognitive measures in the ABCD cohort and cognitive traits in an external Human Connectome Project Development (HCP-D) cohort. These findings suggest that our method effectively characterizes generalizable and fine-grained FC patterns linked to cognition in youth.