Multimodal Prediction of Distressing and Persistent Psychotic-Like Experiences in Adolescence across Environment, Genetics, and Neuroimaging

Abstract

Adolescent psychotic-like experiences, particularly those that are distressing and persistent (dpPLEs), may elevate risk for a range of psychiatric disorders, yet the relative predictive contributions of brain structure, genetics, and environment to dpPLEs remain unknown. Using four timepoints from 7,399 children in the Adolescent Brain Cognitive Development (ABCD) Study, we developed stacked ensemble machine-learning models integrating cortical morphometry and whole-brain microstructural properties, polygenic risk scores, family psychiatric history, and childhood environmental factors to predict adolescent dpPLEs. Environmental predictors alone, driven by area deprivation index, neighborhood drug-related crime, and cumulative early adversity, were the strongest predictive domain (ROC-AUC = 0.687 ± 0.003), outperforming genetic (0.619 ± 0.020) and all imaging models (0.588-0.650). Adding genetic risk to environmental features produced the best-performing model overall (ROC-AUC = 0.697 ± 0.007), whereas adding neuroimaging yielded no further incremental gain (ROC-AUC=0.683-0.696). Beyond environmental exposures, family history of psychosis, cross-disorder polygenic risk, and both globally distributed and regionally specific imaging components contributed to prediction. Childhood environmental factors thus capture a core predictive signal preceding dpPLEs, while genetic and neuroimaging features added comparatively modest discriminative prediction value, but may offer complementary biological insights.