Study Overview and Methodology
A study published on arXiv (identifier arXiv:2608.14578) leverages longitudinal tracking data from approximately 11,860 participants in the Adolescent Brain Cognitive Development (ABCD) Study to construct an early risk prediction system for the initiation of alcohol and cannabis use.
The research team evaluated and benchmarked cross-sectional, longitudinal, and graph-based approaches. Evaluated architectures included decision tree models, Recurrent Neural Networks (RNNs), and Temporal Graph Convolutional Networks (T-GCN). The T-GCN relational graphs were constructed based on family ties, school environments, and feature-level similarities across individuals.
Predictive Performance and Hybrid Architecture
Experimental results revealed that models utilizing longitudinal data consistently outperformed baseline cross-sectional models. In standalone evaluations, temporal XGBoost achieved the highest individual performance. Although the standalone T-GCN architecture did not outperform temporal XGBoost on its own, graph-extracted risk scores provided valuable complementary signals.
By integrating predictions from temporal XGBoost and T-GCN via score-level stacking, the hybrid pipeline delivered the highest overall predictive performance across all target behaviors—including alcohol sipping, full alcohol use, cannabis use, and co-use—achieving an Area Under the ROC Curve (AUC-ROC) exceeding 0.79.
Key Risk Predictors
Feature importance analysis highlighted the primary predictive factors for adolescent substance use initiation, including:
* Peer deviance and social influence * Chronological age * Externalizing behavioral symptoms * Level of parental monitoring * Cultural norms and neighborhood socioeconomic environment