🤖 AI

Longitudinal and Graph-Augmented Prediction of Adolescent Substance Use Onset in the ABCD Study

This study demonstrates that combining longitudinal modeling with graph-augmented relational data from the ABCD Study significantly improves the prediction of adolescent substance use onset, achieving superior performance through a stacked ensemble of temporal XGBoost and Temporal Graph Convolutional Networks.

Original authors: Yixuan He, Jinni Su, Yun Kang

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yixuan He, Jinni Su, Yun Kang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every year, millions of teenagers take their first sip of alcohol or try marijuana. While many will experiment and move on, for some, this early start marks the beginning of a dangerous path toward addiction and long-term health problems. Scientists have long known that the risk of falling into substance use is not a fixed trait determined at birth; rather, it is a story that unfolds over time. A child's environment, their friendships, their family's rules, and their own changing behaviors all shift and interact as they grow from ten years old into their late teens. The challenge for researchers has been to find a way to read this unfolding story early enough to help, distinguishing the children who are simply curious from those who are heading toward trouble.

To solve this, a team of researchers turned to a massive, ongoing project called the Adolescent Brain Cognitive Development Study. This study follows nearly 12,000 children across the United States, checking in with them every year to track their lives in incredible detail. The researchers wanted to see if they could use this stream of data to predict which children would start using substances before they actually did. They tested three different ways of looking at the data: a snapshot of a child's life at a single moment, a movie of their life changing over time, and a map of how children are connected to one another through family, school, and shared experiences.

The researchers began by asking a simple question: is a single look at a child's life enough to predict the future? They took the data from the very first year of the study, which included details like age, family income, school performance, and neighborhood safety, and tried to guess who would start drinking or using marijuana years later. They used powerful computer programs to find patterns in this static information. The results were decent, but not great. The programs could identify some risk factors, such as a child's age or the rules their family had about alcohol, but they missed the dynamic changes that happen as a teenager grows up. It was like trying to predict the weather by looking at a single photograph of the sky; you can see the clouds, but you cannot see the wind shifting or the storm building.

Next, the team tried a different approach, feeding the computer the entire history of each child's life up to the point of prediction. Instead of a single snapshot, they gave the model a sequence of updates, showing how a child's behavior, friendships, and mental health changed from year to year. This method proved to be far more powerful. By watching the trajectory of a child's life, the computer could spot warning signs that a single snapshot would miss. For instance, it became clear that a sudden increase in hanging out with friends who break the rules, or a rise in behavioral problems, were strong signals that substance use might be on the horizon. The model that best captured these changing patterns was a flexible tool that could learn complex, non-linear relationships in the data, effectively saying that the story of a child's life matters more than any single chapter.

The researchers then asked if knowing who a child knows could add even more value. Adolescents do not live in a vacuum; they are influenced by their parents, their teachers, and their peers. To test this, the team built digital maps that connected children based on shared family members, shared schools, or even similar life circumstances. They used a type of artificial intelligence designed to understand these connections, hoping it would reveal hidden risks that individual data points could not show. While these maps provided useful clues, they did not outperform the method that simply tracked the child's own history over time. The connections between children added a layer of context, but the most powerful signal remained the child's own changing behavior and environment.

The breakthrough came when the researchers combined these approaches. They took the strong predictions from the timeline-based model and added the insights from the connection-based maps. This combination worked better than any single method alone. By blending the story of the individual's life with the context of their relationships, the team achieved the highest level of accuracy in predicting substance use onset. Their best model correctly identified the risk for alcohol and marijuana use with a high degree of reliability, outperforming all previous attempts that relied on static data or isolated methods.

The analysis also revealed exactly what drives these predictions. While a child's age was always a major factor, the most critical warning signs shifted as the children grew older. In the early years, family rules and neighborhood conditions were dominant. But as the children entered their teens, the influence of their peers and their own behavioral struggles became the loudest signals. A child who started associating with friends who engaged in rule-breaking behavior, or who began showing signs of externalizing problems like aggression or impulsivity, was at significantly higher risk. These findings suggest that the most effective way to protect teenagers is not just to look at their background, but to watch how their world changes around them and how they react to it.

Ultimately, this work demonstrates that predicting the future of adolescent health requires a dynamic view. A single assessment of a child's life is insufficient because the risks evolve. The most accurate picture emerges when we track the unfolding story of a child's development while also understanding the web of relationships that surrounds them. While no model can predict the future with absolute certainty, this study shows that by combining the timeline of a life with the map of its connections, we can identify the children who need help the most, offering a clearer path for prevention before the first drink or the first puff.

Technical Summary: Longitudinal and Graph-Augmented Prediction of Adolescent Substance Use Onset in the ABCD Study

Problem Statement

Early identification of adolescent substance-use risk is a critical public health challenge, as early initiation correlates with elevated risks of later substance use disorders and adverse developmental outcomes. While the Adolescent Brain Cognitive Development (ABCD) Study offers a rich, multi-domain longitudinal dataset, existing machine learning approaches face two primary limitations:

Static Representations: Most models rely on baseline or cross-sectional data, failing to fully exploit longitudinal trajectories where risk factors evolve.
Independence Assumption: Existing models typically treat participants as independent observations, ignoring the extensive evidence that peer and social environments significantly influence adolescent behavior.

This study addresses these gaps by evaluating whether integrating longitudinal temporal modeling and graph-based relational context can improve the prediction of substance-use onset (alcohol sipping, alcohol use, marijuana use, and combined use) compared to standard baseline approaches.

Methodology

Data and Preprocessing

Dataset: Approximately 11,860 participants from the ABCD Study, spanning up to 10 temporal waves (T1–T10, ages 10–17).
Features: 53 dimensions including sociodemographic, individual, family, peer, school, neighborhood, cultural, and structural variables. Missing values (~55% of feature-wave cells) were handled via Full-Information Maximum Likelihood with EM (FIML-EM).
Tasks:
- Task 1 (Baseline): Predict final-wave onset using only Wave 1 features.
- Task 2 (Temporal): Predict final-wave onset using all features from waves prior to the outcome.

Graph Construction

To model relational context, the authors constructed 17 graph variants representing distinct developmental mechanisms:

ID-based Graphs: Family, School, Site, and unions thereof (e.g., School+Family).
Feature-Similarity Graphs: k-Nearest Neighbors (kNN) with varying sparsity ( $k=5, 10, 20$ ) and Cosine similarity graphs (thresholds 0.5%–10%).
Structure: Graphs were constructed per wave using missingness-aware distance calculations. Self-loops were included, and adjacency matrices were binarized and row-normalized.

Models Evaluated

Baseline (Task 1): Elastic Net, Random Forest, XGBoost, MLP, and Graph Convolutional Networks (GCN).
Temporal (Task 2): LSTM, GRU, Temporal XGBoost (flattened $T \times F$ $T \times F$ input), and Temporal Graph Convolutional Networks (T-GCN).
- T-GCN Architecture: 2-layer GCN spatial encoder per wave followed by a GRU temporal update.
Ensemble/Stacking:
- Score-Level Stacking: A logistic regression or XGBoost model combining risk scores from Temporal XGBoost and selected T-GCNs (Family, kNNsym5, Cos0.5%).
- Augmented XGBoost: Appending T-GCN risk scores directly to the flattened feature set.

Key Contributions

Systematic Comparison: A rigorous evaluation of baseline, longitudinal, and graph-based frameworks for predicting four distinct substance-use outcomes.
Diverse Graph Strategies: Implementation of multiple graph-construction strategies capturing family, school, and latent feature-similarity relationships to model social and contextual influences.
Graph-Augmented Stacking Framework: Demonstration that a score-level stacking approach combining temporal XGBoost and T-GCN predictions yields the highest predictive performance, achieving AUC-ROC values above 0.79 across all outcomes.

Results

Predictive Performance

Longitudinal Superiority: Temporal models significantly outperformed baseline models. Temporal XGBoost achieved the strongest standalone performance, with AUC-ROC gains over baseline XGBoost ranging from 0.086 to 0.171.
Graph Performance:
- In Task 1, standard tree-based models (XGBoost, Random Forest) outperformed baseline GCNs.
- In Task 2, T-GCNs generally did not surpass Temporal XGBoost alone. Family graphs and sparse feature-similarity graphs (e.g., kNNsym5) performed best among graph variants.
Ensemble Success: The Score-Level Stacking approach (combining Temporal XGBoost and T-GCN scores) achieved the best results across all outcomes:
- AUC-ROC: 0.796 (Alcohol Sips) to 0.822 (Alcohol).
- AUC-PR: Consistent improvements over standalone models, confirming gains are not solely due to class imbalance handling.
- Note: Directly appending graph scores to raw features (Augmented XGBoost) underperformed score-level stacking, suggesting graph scores act better as calibrated ensemble predictors than raw features in high-dimensional spaces.

Feature and Domain Importance

Task 1 (Baseline): Dominated by Demographic (Age), Structural (Immigration bias), Neighborhood, and Cultural factors (Religious rules on alcohol).
Task 2 (Temporal): The predictive signal shifted toward Individual and Peer domains. Key predictors included Peer Deviance (at later waves, e.g., t8), Externalizing Problems, and Age.
Graph Contribution: Graph models emphasized similar features (Age, Peer Deviance, Externalizing) but provided complementary relational risk summaries. In the stacked model, the Temporal XGBoost score was the largest contributor, with graph scores providing consistent, non-zero auxiliary signals.

Significance and Claims

The paper claims that longitudinal modeling is the primary driver of predictive performance for adolescent substance-use onset, as developmental trajectories contain risk information irrecoverable from a single baseline snapshot.

While graph-based models did not independently surpass strong temporal tabular learners (like Temporal XGBoost), the study demonstrates that graph-derived risk scores provide complementary information. By integrating these relational signals through score-level stacking, the proposed framework achieves state-of-the-art performance (AUC-ROC > 0.79) and offers a promising direction for incorporating social and contextual structure into risk prediction.

The authors conclude that future substance-use risk is best characterized by a combination of individual developmental trajectories and relational context, with the proposed stacking framework effectively balancing these signals. Practical deployment considerations, such as threshold selection based on intervention costs, are noted as future work.