Design and Optimisation of an Integrated Multi-Source Ensemble Learning Framework for Predicting Teacher Depression Risk

Existing approaches to predicting teacher depression rely predominantly on single-source self-reported data and conventional statistical models, limiting their ability to capture the complex interactions among psychological, occupational, financial, and contextual risk factors and reducing their generalizability across teaching populations. This study designed and optimised an integrated multi-source ensemble learning framework for predicting teacher depression risk. Structured survey data comprising PHQ-9 depression severity scores together with demographic, occupational, financial, and institutional-support indicators were integrated into a unified feature space and refined using a consensus feature-selection strategy based on Random Forest, XGBoost, and LightGBM feature importance rankings. Four heterogeneous machine learning algorithms (Random Forest, CatBoost, LightGBM, and XGBoost) were trained following Synthetic Minority Oversampling Technique (SMOTE)-based class balancing and combined using a stacked ensemble architecture optimised through hyperparameter tuning, probability calibration, and stratified cross-validation. Model interpretability was examined using SHapley Additive exPlanations (SHAP), while external generalizability was evaluated on a held-out dataset from Kenyan public schools. The optimised stacked ensemble outperformed all individual base learners, achieving an accuracy of 78.0%, a recall of 80.0%, and an F1-score of 82.8%, with institutional-support, occupational, and financial-strain variables emerging as the most influential predictors of teacher depression risk. Although the ensemble demonstrated superior discriminatory performance, its probability calibration was inferior to that of individually calibrated models, and the framework's robustness to noisy or incomplete data was not empirically assessed. Grounded in Ensemble Theory and the Job Demands–Resources framework, the findings demonstrate that integrating heterogeneous data sources through ensemble learning substantially improves the prediction of teacher depression risk while highlighting that enhanced predictive accuracy alone does not guarantee deployment-ready clinical decision support. Consequently, calibrated individual learners may be more appropriate for referral and intervention decisions, whereas the prominence of institutional-support factors underscores the need for policy interventions that strengthen school-based mental health systems. Further validation across diverse educational contexts and under varying data-quality conditions is recommended to enhance the framework's robustness and practical applicability.