Synergizing Gradient Boosting and Random Forest: An Interpretable Dual-Model Framework to Unveil Key Stressors and Mechanisms in Adolescent Educational Mental Health

Abstract

Adolescent educational stress, shaped by intricate interactions among psychological, physiological, and environmental factors, poses a substantial challenge to educational mental health. Traditional assessment methods, however, struggle to capture these dynamic relationships, thereby limiting the effectiveness of interventions. To address this gap, the present study introduces an interpretable dual-model framework integrating Gradient Boosting Machine (GBM) and Random Forest (RF). Leveraging data from 1,000 adolescents, this framework identifies key stressors and their underlying mechanisms through hyperparameter optimization and multi-modal validation (Spearman correlations, SHapley Additive exPlanations [SHAP] analysis, and feature importance rankings). The framework achieved high predictive accuracy (R² > 0.80, MAE < 0.15). Key findings include that self-esteem emerges as the dominant stress predictor (ΔR² ≈ 0.13), followed by academic performance (ΔR² ≈ 0.11). SHAP visualizations further revealed nonlinear threshold effects (e.g., those related to academic performance) and anxiety-mediated pathways. Additionally, model comparisons indicated that RF exhibited superior noise robustness (MAE = 0.135 versus GBM’s 0.146), whereas GBM better captured linear relationships in physiological variables. By leveraging feature importance rankings, the framework enables targeted stress interventions, thus optimizing resource allocation in educational mental health.