Longitudinally stable, brain-based predictive models mediate the relationships between childhood cognition and socio-demographic, psychological and genetic factors

Abstract

General intelligence, or G-Factor, of children is associated with important life outcomes, including educational attainment, employment, health, and mortality. Thus, determining the neural predictor of the G-Factor of children has become one of the main tasks in neuroscience. Here we aim to build neuroimaging-based predictive models of children's general intelligence that are longitudinally stable across two years. To achieve this goal, we used large-scale, longitudinal data with multiple neuroimaging modalities from the Adolescent Brain Cognitive Development (ABCD) study (n ~11k). We first computed modality-specific models from six MRI modalities (three task-based, resting-state, structural, and diffusion tensor imaging) of the baseline data using Elastic Net. We, then, combined the predicted values of all modality-specific models using Random Forest Opportunistic Stacking. The stacked model allowed us to predict the G-Factor of unseen, same-age (9-10-year-old) children at Pearson r=.44, better than any other modality-specific models. Based on permutation tests, this prediction was predominantly driven by activity in the parietal and frontal areas during the N-Back working-memory task. Importantly, this model was generalizable to unseen children from the follow-up data who were two years older at r=.41. Moreover, this model allowed for missingness in the data, making it possible for us to maintain around 73% of the data that would otherwise be excluded due to different artifacts occurring to any modalities. Accordingly, we developed an MRI-multimodal predictive model for the G-Factor of children that is 1) stable across years, 2) interpretable and 3) able to handle missing values.