A MACHINE LEARNING APPROACH TO PREDICTING HIGH-RISK IRRITABILITY TRAJECTORIES ACROSS THE TRANSITION TO ADOLESCENCE

Abstract

Adolescence is a sensitive developmental period that presents a crucial opportunity for early intervention to mitigate risk for future psychopathology. Transdiagnostic symptoms are important indicators of risk. Irritability, characterized by proneness to anger, frustration, and temper outbursts, is a transdiagnostic symptom associated with negative mental health outcomes in youth and adults. Prior work has characterized youth with varying levels of irritability across different developmental periods and identified irritability trajectories that signify high risk. Data were from the Adolescent Brain Cognitive Development (ABCD) Study, which is a 10-year longitudinal study that tracks the brain development, cognitive skills, physical health, and psychosocial functioning of a large, national sample starting from preadolescence. The baseline sample consisted of 11,862 9-10-year-old preadolescent youth. Irritability was parent-rated at baseline, 1-year, 2-year, 3-year, and 4-year follow-ups on the Child Behavior Checklist (CBCL) irritability index. Latent class growth analysis (LCGA) was used to determine developmental trajectories of irritability. Four machine learning approaches were applied to develop predictive models for youth irritability trajectories. The baseline (preadolescent) variables covered a wide range of domains (youth psychopathology, youth physical health, neurocognitive abilities, psychosocial environment, demographics, parent psychopathology, and structural neurobiology). LCGA identified four distinct irritability trajectories: persistent low irritability (n = 8692, 73.28%), moderate irritability and decreasing (n = 1083, 9.13%), low to moderate irritability and increasing (n = 1449, 12.22%), and chronic high irritability (n = 638, 5.38%). The machine learning models demonstrated strong performance in detecting probability of being in the chronic high irritability trajectory, but poorer performance in classifying the two irritability trajectories that were characterized by change in irritability. In addition, the machine learning models showed strong performance in differentiating the persistent low irritability trajectory from the other trajectories. The top predictors indicated that the youth psychopathology domain produced the most relevant predictors. The machine learning models also highlighted several novel predictors for irritability trajectory which merit further research. Furthermore, behavioral and clinical variables outperformed the structural neurobiology variables in predicting irritability trajectories. The present study provides a foundation for future development of models to predict irritability trajectories during a critical developmental period for early intervention.