Longitudinal and Graph-Augmented Prediction of Adolescent Substance Use Onset in the ABCD Study
Longitudinal modeling consistently outperforms cross-sectional approaches for predicting adolescent substance use onset, with temporal XGBoost achieving the strongest standalone performance Temporal Graph Convolutional Networks (T-GCNs) built from family, school, and feature-similarity graphs do not surpass temporal XGBoost alone but provide complementary risk signals Score-level stacking of temporal XGBoost and T-GCN predictions yields the best overall performance, achieving AUC-ROC values abov
Analysis
TL;DR
- Longitudinal modeling consistently outperforms cross-sectional approaches for predicting adolescent substance use onset, with temporal XGBoost achieving the strongest standalone performance
- Temporal Graph Convolutional Networks (T-GCNs) built from family, school, and feature-similarity graphs do not surpass temporal XGBoost alone but provide complementary risk signals
- Score-level stacking of temporal XGBoost and T-GCN predictions yields the best overall performance, achieving AUC-ROC values above 0.79 across all substance use outcomes
- Key predictive features include peer deviance, age, externalizing symptoms, parental monitoring, cultural norms, and neighborhood context
- The study leverages a large-scale dataset of approximately 11,860 participants from the ABCD Study to compare cross-sectional, longitudinal, and graph-based prediction approaches
Why It Matters
This research demonstrates the practical value of combining temporal and relational modeling for public health prediction tasks, offering a blueprint for how AI practitioners can integrate multiple data modalities (longitudinal trajectories and social graphs) to improve risk prediction. The findings are directly relevant to developers building early intervention systems in healthcare and social services, where understanding both individual trajectories and peer/family influence networks can significantly enhance predictive accuracy.
Technical Details
- Dataset: Approximately 11,860 participants from the Adolescent Brain Cognitive Development (ABCD) Study, predicting four outcomes: alcohol sipping, alcohol use, marijuana use, and alcohol/marijuana co-use
- Model architectures evaluated: Tree-based models (temporal XGBoost), recurrent neural networks, and Temporal Graph Convolutional Networks (T-GCNs) constructed from three graph types: family graphs, school graphs, and feature-similarity graphs
- Comparison framework: Systematic evaluation of cross-sectional (baseline), longitudinal, and graph-based approaches to isolate the contribution of temporal dynamics versus relational context
- Ensemble method: Score-level stacking combining temporal XGBoost and T-GCN predictions, which achieved the best performance with AUC-ROC values above 0.79 across all outcomes
- Feature importance analysis: Identified peer deviance, age, externalizing symptoms, parental monitoring, cultural norms, and neighborhood context as the most influential predictors
Industry Insight
- The complementary nature of tree-based temporal models and graph neural networks suggests that hybrid ensemble strategies should be the default approach for complex social-health prediction tasks, rather than relying on a single model architecture
- For AI practitioners working in public health or social impact domains, investing in longitudinal data collection and graph construction (family, school, community networks) can yield measurable improvements in predictive performance beyond what cross-sectional data alone can provide
- The AUC-ROC above 0.79 achieved through stacking indicates that while individual models have limitations, practical deployment of early-warning systems for adolescent substance use is feasible when multiple signal sources are integrated
Disclaimer: The above content is generated by AI and is for reference only.