Predicting the onset of mental health problems in adolescents.

Abstract

Objective: Mental health problems are the major cause of disability among adolescents. Personalized prevention may help to mitigate the development of mental health problems, but no tools are available to identify individuals at risk before they require mental health care.

Methods: We identified children without mental health problems at baseline but with six different clinically relevant problems at 1- or 2-year follow-up in the Adolescent Brain Cognitive Development (ABCD) study. We used machine learning analysis to predict the development of these mental health problems with the use of demographic, symptom and neuroimaging data in a discovery (N = 3236) and validation (N = 3851) sample. The discovery sample (N = 168-513 per group) consisted of participants with MRI data and were matched with healthy controls on age, sex, IQ, and parental education level. The validation sample (N = 84-231) consisted of participants without MRI data.

Results: Subclinical symptoms at 9-10 years of age could accurately predict the development of six different mental health problems before the age of 12 in the discovery and validation sample (AUCs = 0.71-0.90). The additive value of neuroimaging in the discovery sample was limited. Multiclass prediction of the six groups showed considerable misclassification, but subclinical symptoms could accurately differentiate between the development of externalizing and internalizing problems (AUC = 0.79).

Conclusions: These results suggest that machine learning models can predict conversion to mental health problems during a critical period in childhood using subclinical symptoms. These models enable the personalization of preventative interventions for children at increased risk, which may reduce the incidence of mental health problems.