From Current-Wave to Longitudinal Risk Prediction: A Leakage-Aware Stacked Ensemble Framework for Adolescent Substance Use Using the ABCD Study

Adolescent use of alcohol, nicotine, and marijuana remains a major public health concern in the United States. Early identification of youth at elevated risk is critical for prevention before use begins or escalates. We developed and evaluated a longitudinal machine learning framework to predict alcohol, nicotine, and marijuana use at the next observed assessment wave. Data came from the Adolescent Brain Cognitive Development (ABCD) Study Release 6.0. The models incorporated predictors from multiple domains, including demographics, friends, family and community context, mental health, physical health, and prior substance-related behaviors. To reduce information leakage across individuals, we implemented a leakage-aware stacked ensemble. This ensemble combined diverse base learners through out-of-fold predictions and an elastic-net meta-learner. Across all three substances, the lagged stacked ensemble outperformed the cross-sectional stack and all single base learners. Adolescents identified as highest risk showed substantially higher observed rates of substance use than would be expected under random screening. Feature-importance analyses showed that the full longitudinal models were strongly influenced by developmental timing and prior-use history. Analyses restricted to current-wave features revealed distinct substance-specific risk patterns beyond prior-use history and developmental timing. Bootstrap stability analyses identified top-ranked features showing consistent positive predictive relevance across resampled adolescents. These findings suggest that longitudinal, leakage-aware machine learning can generate substance-specific risk estimates to support targeted prevention and screening in adolescent populations.