Depression Severity Prediction Using Stacked Heterogenous Learning on Integrated Biomedical and Lifestyle Data

Abstract

Depression, a major global health concern, requires improved methods for timely and precise assessment. This study proposes a stacked heterogeneous ensemble framework to predict depression severity, measured by cumulative PHQ-9 scores, using multimodal data from the 2013–2014 NHANES survey. The dataset integrates six domains, such as demographics, diet, physical examination, laboratory results, medication use, and mental health responses, capturing both biomedical and lifestyle indicators. Seven base learners (SVR, Ridge, ElasticNet, KNN, Decision Tree, LightGBM, and XGBoost) were stacked in two stages. First-level outputs were fused with a Voting Regressor (Gradient Boosting Regressor + Linear Regression). The PHQ-9 scale ranges from 0 to 27, with higher values reflecting greater severity. A total of 19,560 participants were analyzed (80/20 split: 15,648 train, 3,912 test), with inverse-frequency weighting to address class imbalance. Results show the Stacking Regressor outperformed single models, achieving MSE = 1.66, RMSE = 1.29, and R2 = 0.93. These findings highlight ensemble learning’s promise for psychiatric modeling and its potential as a scalable, interpretable tool for clinical decision support.