Predicting mental health risk in adolescents using interpretable machine learning: Evidence from SHAP analysis.

BACKGROUND/OBJECTIVE: Adolescent mental health is a critical public health concern; however, the combined contribution of emotional, demographic, and behavioral factors to poor mental health remains incompletely understood. This study applied an interpretable machine learning framework to identify key predictors of elevated mental health risk among adolescents and characterize their relative and non-linear contributions. METHODS: Data were drawn from the 2021 Youth Risk Behavior Survey (YRBS), comprising 16,946 U.S. high school students. Elevated mental health risk was defined using the YRBS current mental health item (Q85): reporting that mental health was not good most of the time or always during the past 30 days. The distinct 12-month sadness/hopelessness item (Q25) was retained as a predictor in the primary XGBoost model. The operating threshold was selected by maximizing the Youden index using five-fold out-of-fold predictions in the training set and was then locked before evaluation in the held-out test set. SHapley Additive exPlanations (SHAP) were used to assess feature importance and directional and non-linear patterns. RESULTS: Overall, 28.0% of adolescents were classified as having elevated mental health risk, with higher prevalence among females (39.9% vs. 16.9%) and peak prevalence at ages 16-17. The primary XGBoost model achieved an AUC of 0.806 and PR-AUC of 0.603. At the training-derived threshold of 0.285, accuracy was 0.749, sensitivity 0.754, specificity 0.747, and balanced accuracy 0.750. SHAP analysis identified sadness/hopelessness as the dominant predictor, followed by sex, sleep duration, screen time, and breakfast frequency. In a sensitivity analysis excluding sadness/hopelessness, discrimination decreased (AUC 0.736; PR-AUC 0.510). CONCLUSION: The revised primary model achieved balanced sensitivity and specificity above 0.700 using a threshold selected entirely within the training data. Sadness/hopelessness was the strongest predictor, while demographic and behavioral factors provided additional predictive information. Because sadness/hopelessness is conceptually close to the current mental health outcome and the data are cross-sectional, the model should be interpreted as an auxiliary risk-stratification tool rather than a standalone clinical screening or diagnostic instrument.