Predictive models of mental health based on resting-state functional connectivity show limited generalizability across late childhood and early adolescence
Late childhood and adolescence are marked by substantial reorganization of the brain’sfunctional connectome and the emergence of many mental health problems. Although resting-state functional connectivity (rsFC) has been widely used to predict mental health and relatedbehavioral phenotypes, it remains unclear whether such models generalize across developmentor instead reflect age-specific brain-behavior mappings. Here, we tested the developmentalgeneralizability of whole-brain rsFC-based predictive models of mental health and cognitiveability across three timepoints spanning ages 9-14 years in the Adolescent Brain CognitiveDevelopment Study. Using out-of-sample prediction and cross-timepoint model transfer, wefound that rsFC significantly predicted concurrent mental health and cognitive ability withineach age group, but models trained at one age often generalized poorly to data acquired 2-4years earlier or later, particularly for mental health. In addition to this limited transfer, wedemonstrate developmental shifts in the associations between individual functionalconnections and mental health and cognitive ability. These changes occurred alongside broaderreorganization of the functional connectome, including age-related differences in globalcovariance structure and cortical network segregation. Together, our findings indicate thatbrain-behavior mappings of mental health are highly dynamic across late childhood andadolescence and may challenge the assumption that predictive models or candidate neuralbiomarkers for mental health disorders derived at a single age remain valid across development.