Beyond out-of-sample: robust and generalizable multivariate neuroanatomical patterns of psychiatric problems in youth

Abstract

Mapping differential brain structures for psychiatric problems has been challenging due to a lack of regional convergence and poor replicability in previous brain-behavior association studies. By leveraging two independent large cohorts of neurodevelopment, the ABCD and Generation R Studies (total N = 11271), we implemented an unsupervised machine learning technique with a highly stringent generalizability test to identify reliable brain-behavior associations across diverse domains of child psychiatric problems. Across all psychiatric symptoms measured, one multivariate brain-behavior association was found, reflecting a widespread reduction of cortical surface area correlated with higher child attention problems. Crucially, this association showed marked generalizability across different populations and study protocols, demonstrating potential clinical utility. Moreover, the derived brain dimension score predicted child cognitive and academic functioning three years later and was also associated with polygenic scores for ADHD. Our results indicated that attention problems could be a phenotype for establishing promising multivariate neurobiological prediction models for children across populations. Future studies could extend this investigation into different development periods and examine the predictive values for assessment of functioning, diagnosis, and disease trajectory in clinical samples.