Using Machine Learning to Derive Neurobiological Subtypes of General Psychopathology in Late Childhood

Traditional mental health diagnoses rely on symptom-based classifications. Yet this approach can oversimplify clinical presentations as diagnoses often do not adequately map onto neurobiological features. Alternatively, our study used structural imaging data and a semisupervised machine learning technique, heterogeneity through discriminative analysis, to identify neurobiological subtypes in 9- to 10-year-olds with high psychopathology endorsements (n =9,027). Our model revealed two stable neurobiological subtypes (adjusted Rand index =0.38). Subtype 1 showed smaller structural properties, elevated conduct problems and attention-deficit/hyperactivity disorder symptoms, and impaired cognitive performance compared to Subtype 2 and typically developing youth. Subtype 2 had larger structural properties, cognitive abilities comparable to typically developing youth, and elevated internalizing symptoms relative to Subtype 1 and typically developing youth. These subtypes remained stable in their neurobiological characteristics, cognitive ability, and associated psychopathology traits over time. Taken together, our data-driven approach uncovered evidence of neural heterogeneity as demonstrated by structural patterns that map onto divergent profiles of psychopathology symptoms and cognitive performance in youth.