Source-linked AI summary
Longitudinal and Graph-Augmented Prediction of Adolescent Substance Use Onset in the ABCD Study
Yixuan He, Jinni Su, Yun Kang
TL;DR
Predicting adolescent substance-use onset remains limited by reliance on baseline data and independent-participant models, despite the importance of developmental trajectories and social context. This study compares temporal and graph-based approaches in ABCD data and finds that stacked temporal XGBoost and T-GCN scores perform best across outcomes.
Problem
Existing substance-use prediction models rely mainly on baseline data and treat participants independently, limiting evidence about longitudinal trajectories and social context.
Method
Using approximately 11,860 ABCD participants across up to ten waves, the study compares baseline, longitudinal, and graph-based models across four substance-use outcomes.
Results
0.792–0.822 AUC-ROC was achieved across all four outcomes, with stacked temporal XGBoost and T-GCN scores outperforming cross-sectional baselines.
Takeaways & Limitations
Substance-use risk is best characterized by combining individual developmental trajectories with relational context, while longitudinal information remains the dominant predictive signal.
Takeaways & Limitations
The transductive graph setting allows unlabeled test-participant features to influence graph structure, although outcome labels remain masked.
Abstract
from arXiv · showhide
Early identification of adolescent substance-use risk is an important prevention challenge, yet the relative value of baseline characteristics, longitudinal trajectories, and relational context remains unclear. Using data from approximately 11,860 participants in the Adolescent Brain Cognitive Development (ABCD) Study, we compare cross-sectional, longitudinal, and graph-based approaches for predicting alcohol sipping, alcohol use, marijuana use, and alcohol/marijuana use. We evaluate tree-based models, recurrent neural networks, and Temporal Graph Convolutional Networks (T-GCNs) constructed from family, school, and feature-similarity graphs. Longitudinal models consistently outperform baseline models, with temporal XGBoost achieving the strongest standalone performance. Although T-GCNs generally do not surpass temporal XGBoost, graph-derived risk scores provide complementary information. Combining temporal XGBoost and T-GCN predictions through score-level stacking yields the best performance across all outcomes, achieving AUC-ROC values above 0.79. Feature analyses identify peer deviance, age, externalizing symptoms, parental monitoring, cultural norms, and neighborhood context as important predictors of substance use onset. These findings demonstrate the value of longitudinal modeling for substance-use prediction and suggest that graph-based representations can provide effective auxiliary risk signals.
1 Introduction
Adolescent substance-use initiation is a preventable public-health concern requiring early identification because early onset predicts later substance-use disorders and adverse developmental outcomes. This study addresses limitations of baseline, independent-observation models by evaluating longitudinal and graph-based prediction across four substance-use outcomes.
- Motivation: Early substance-use onset is associated with later substance-use disorders and adverse developmental outcomes, while delaying initiation reduces subsequent risk.These relationships motivate early identification of high-risk youth as a prevention objective.
- Motivation: ABCD repeated assessments capture evolving individual, family, peer, school, neighborhood, and cultural risk and protective factors.The paper frames substance-use vulnerability as a dynamic developmental process rather than a static baseline characteristic.
- Gap: Existing machine-learning approaches often rely on baseline or cross-sectional representations and treat participants as independent observations.The introduction identifies these practices as limitations because they underuse longitudinal trajectories and overlook relational context.
- Study objectives: The study compares baseline machine-learning, longitudinal sequence, and graph-based models for predicting alcohol sipping, alcohol use, marijuana use, and alcohol/marijuana use.The graph-based approaches include Graph Convolutional Networks (GNNs).
- Study objectives: The study tests whether graph-derived risk scores provide complementary information beyond strong temporal tabular learners.This directly evaluates whether relational representations add predictive value beyond temporal modeling.
2 Related Work
Prior ABCD research has used machine learning to identify substance-use predictors from baseline and longitudinal data. These studies highlight family, socioeconomic, sleep-related, environmental, and behavioral factors as relevant predictors.
- Machine-learning approaches: Niklason et al. (2025) used baseline XGBoost and SHAP to predict early alcohol sipping in the ABCD study.The analysis identified family norms, socioeconomic factors, and sleep-related variables as key predictors.
- Machine-learning approaches: Baseline predictors included family norms, socioeconomic factors, and sleep-related variables for early alcohol sipping.
- Machine-learning approaches: Wei et al. (2026) analyzed longitudinal ABCD data and showed that time-varying environmental and behavioral factors contribute to substance-use prediction.
3 Data Description, Preparation, and Prediction Tasks
The study uses approximately 11,860 ABCD participants followed across up to 10 waves to compare baseline-only and temporal prediction of final-wave substance-use outcomes. It also evaluates 17 graph variants representing family, school, site, and feature-similarity relationships.
- Cohort and features: Approximately 11,860 ABCD participants were followed across up to T=10 annual or midyear waves, with ages ranging from 10 at T1 to 17 at T10.The modeling table contained 48 standardized continuous or ordinal Z features spanning sociodemographic and individual characteristics.
- Prediction tasks: Baseline-only prediction uses wave-1 features, whereas temporal prediction uses all waves before the final outcome wave to predict the final-wave outcome.For every outcome and random seed, participants were split into stratified 70% training, 15% validation, and 15% test sets using seeds 0, 10, 20, 30, and 40.
- Graph construction: The study constructs 17 graph variants spanning family, school, site, and feature-similarity relationships to represent distinct developmental mechanisms.Family graphs capture shared genetic and household environmental influences; school graphs approximate peer exposure and shared institutional environments; site graphs capture regional and recruitment-related context; and feature-similarity graphs represent latent homophily.
- Graph construction: Seven ID graphs connect participants sharing school, site, family, or unions of these identifiers: school, site, family, school+site, school+family, site+family, and school+site+family.Family graphs exclude implausibly large family groups to maintain sparsity.
- Graph construction: Feature-similarity graph construction uses a feature dimension of F+2=55 after expanding one-hot sex and race encodings before removing dependent dimensions during modeling.The dependent dimensions are removed during model training and evaluation to avoid multicollinearity.
4 Experiments
Experiments across four substance-use outcomes show that longitudinal models outperform baseline approaches, while graph-derived risk scores add complementary information through score-level stacking. Feature analyses consistently highlight age, peer deviance, externalizing behaviors, cultural norms, and family-related factors.
- Experimental design: Task 1 compared Elastic Net, Random Forest, XGBoost, MLP, and graph neural networks across four outcomes.Task 1 GCNs used two layers, 64-dimensional hidden representations, and sparse adjacency matrices.
- Task 1 results: 0.706 AUC-ROC: XGBoost achieved this for Alcohol Sips Ever, while GCN models reached approximately 0.62–0.69.Tree-based models, especially XGBoost and Random Forest, consistently outperformed baseline GNNs and MLPs in cross-sectional experiments.
- Task 2 results: 0.086 to 0.171 AUC-ROC gains: temporal XGBoost improved over baseline XGBoost across outcomes, and the score-logit XGBoost+TGCN stacker ranked highest for all four Task 2 outcomes.GRU, LSTM, and temporal XGBoost achieved AUC-ROC values between 0.78 and 0.82 across outcomes.
- Task 2 results: 0.634 to 0.646 AUC-PR: for marijuana use, temporal XGBoost improved to the score-logit stacker, which also achieved the highest AUC-PR across all Task 2 outcomes.AUC-PR rankings were largely consistent with AUC-ROC results; for the composite alcohol/marijuana outcome, the reported increase begins at 0.856.
- Graph analyses: Family and school graphs generally outperformed site-based graphs, while family and feature-similarity T-GCN scores contributed nonzero complementary signals in the stacker.The temporal XGBoost score remained the largest stacker contribution, and direct feature appending often underperformed score-level stacking.
- Feature importance: Age, peer deviance, externalizing behaviors, cultural substance-use norms, and family-related factors were consistently identified as important predictors across outcomes.Temporal models emphasized later-wave deviant peer exposure, age, and externalizing problems, whereas cross-sectional models repeatedly selected cultural, structural, and neighborhood features.
5 Conclusion and Future Work
Longitudinal trajectories provide the dominant predictive improvement in ABCD substance-use prediction, while graph neural networks contribute informative but generally weaker signals. Graph-specific T-GCN risk scores can modestly improve temporal XGBoost while preserving interpretable graph contributions and domain signal.
- Longitudinal trajectories are the dominant source of predictive improvement, with temporal XGBoost substantially outperforming baseline XGBoost.
- Graph neural networks are informative but usually weaker than temporal XGBoost, particularly when graph aggregation dilutes individual trajectories.
- Graph-specific T-GCN risk scores can modestly improve temporal XGBoost while preserving interpretable graph contributions and domain signal.