Table 1_Predicting multiple mental health outcomes in adolescents using explainable machine learning models.docx

SHapley Additive exPlanations (SHAP) were used to examine feature contributions. To minimize target leakage, outcome-specific feature sets were used, with PSQI and RSES excluded from the depression model because of direct or substantial conceptual overlap with PHQ-9 content, and PSQI excluded from the well-being model because of overlap with WHO-5 content.</p>Results<p>XGBoost showed the strongest out-of-sample predictive performance across all four outcomes, explaining 49% of the variance in depression (R² = 0.49, 95% CI: 0.44–0.53; RMSE = 3.52; MAE = 2.81), 55% in anxiety (R² = 0.55, 95% CI: 0.50–0.59; RMSE = 3.04; MAE = 2.47), 60% in perceived stress (R² = 0.60, 95% CI: 0.56–0.64; RMSE = 3.71; MAE = 2.87), and 50% in psychological well-being (R² = 0.50, 95% CI: 0.45–0.54; RMSE = 3.38; MAE = 2.68).</p>Discussion<p>Emotional dysregulation and loneliness were consistently among the most influential features, while academic stress, family support, social media addiction, self-esteem, and sleep quality showed outcome-specific contributions.

SHAP rankings for the depression model were stable across five-fold cross-validation, with emotional dysregulation and loneliness consistently occupying the highest ranks.</p>