Longitudinally stable, brain-based predictive models explain the relationships of childhood intelligence with socio-demographic, psychological and genetic factors

Abstract

AbstractChildhood intelligence is strikingly predictive of major life outcomes. Understanding the brain underpinnings of early-life intelligence is of prime clinical and public health importance, but has so far remained elusive. Here, we demonstrate that it is possible to arrive at models of the brain that are both predictive and explanatory of childhood intelligence. For this we leverage the unique power of the large-scale, longitudinal, multimodal MRI data from the Adolescent Brain Cognitive Development (ABCD) study (n ∼11k) and create a novel predictive modeling framework that integrates machine-learning and structural-equation-based methodologies. Our predictive models combine six MRI modalities (task-fMRI from three tasks, resting-state fMRI, structural MRI, DTI) using machine-learning. In terms of prediction, our models achieve an unprecedented longitudinal association (r=.41) with childhood intelligence across two years in unseen data. We found fronto-parietal networks during a working-memory task to drive childhood-intelligence prediction. In terms of explanation, our models significantly explain variance in childhood intelligence due to (1) key socio-demographic and psychological factors (proportion mediated=18.65% [17.29%-20.12%]) and (2) genetic liability, as reflected by the polygenic score of cognitive ability (proportion mediated=15.6% [11%-20.7%]). In summary, our work shows that novel models of human multimodal neuroimaging data are powerful in helping us predict and explain variation in childhood intelligence.