Related Experiment Videos
Identifying risk factors for marijuana use among male high school students using machine learning: Implications for
Han H Huynh1, Khanh T Q Le2, Minh H N Le3
1Irell and Manella Graduate School of Biological Science, City of Hope, CA, USA; Chung Shan Medical University, Institute of Medicine, Taichung, Taiwan.
Objectives:
To identify risk and protective factors associated with lifetime marijuana use among male high school students through an interpretable machine learning model, providing evidence to support early and targeted public health interventions.
Study Design:
Cross-sectional analysis of 2023 Youth Risk Behavior Surveillance System (YRBS) data for boys in grades 9-12 across the United States.
Methods:
The final analytical sample included 8285 boys after excluding missing outcomes. Thirty-six predictors spanning demographics, substance use behaviors, lifestyle, psychosocial stressors, mental health, and household environment were retained following optimization of missing data thresholds. Ensemble feature importance was determined using Logistic Regression, Linear Discriminant Analysis, and Extra Trees classifiers. Seventeen machine learning algorithms were benchmarked, with Logistic Regression selected for its combination of predictive performance and interpretability. Model performance was evaluated using accuracy, F1-score, specificity, and area under the receiver operating characteristic curve (AUC). Explainable AI methods, SHAP and LIME, provided global and individual-level feature interpretations.
Results:
Lifetime marijuana use was reported by 28.4% of participants. The optimized Logistic Regression model achieved strong performance (AUC = 0.9034; Accuracy = 0.8582; F1 = 0.8928). The most predictive factors included electronic vapor product use, cigarette smoking, alcohol consumption, frequent social media use, age, bullying and sexual violence exposure, lifetime cocaine use, academic achievement, and grade level. High academic achievement was protective. SHAP and LIME analyses confirmed the robustness and interpretability of these predictors. The optimized model was also well calibrated (Brier score = 0.104; expected-to-observed ratio = 1.01).
Conclusions:
This study applies an interpretable machine learning approach, grounded in systematic algorithm benchmarking, that provides actionable insights to identify high school boys with patterns associated with marijuana use, supporting early, focused prevention in schools and communities.
