An Interpretable Stacking Ensemble Model with SHAP for Geriatric Depression Prediction: Analysis of the NHANES 2005-2023 Database An interpretable stacking ensemble model for geriatric depression prediction

Abstract

OBJECTIVE: Leveraging multimodal data from the 2005-2023 National Health and Nutrition Examination Survey (NHANES) database, this study aims to develop a predictive method for the geriatric depression that combines high predictive accuracy with good interpretability, thereby providing support for in-depth exploration of the pathogenesis and risk factors of geriatric depression. METHODS: Data from 8760 participants aged 65 and older in the NHANES database from 2005-2023 are utilized to develop and validate the stacking ensemble predictive model. Depression is assessed using the Patient Health Questionnaire-9 (PHQ-9) total score meeting or exceeding 10. Before the model construction, this work employs the normalization of training data and test data, Synthetic Minority Over-sampling Technique-Random Under-Sampling (SMOTE-RUS) hybrid sampling strategy to address the class imbalance, and the recursive feature elimination method based on the random forest (RFE-RF) for feature selection. A stacking ensemble predictive framework for depression is constructed based on the primary learners (Random Forest, SVM, XGBoost, and Logistic Regression) and meta-learners (SVM and Logistic Regression). Finally, the interpretable machine learning technique SHapley Additive exPlanations (SHAP) is used to visualize the model predictive outputs. RESULTS: The XGBoost model demonstrated outstanding performance on the test set in terms of AUC (83.92%), while the Random Forest (RF) model excelled in sensitivity (71.05%). Subsequently, a specifically designed RFE-Stacking ensemble model, using RF and XGBoost as the primary learner and the SVM as the meta-learner, is developed. In comparison, this stacking ensemble model exhibits the best predictive performance with the biggest AUC (85.14%) and the highest sensitivity (78.71%). The SHAP interpretation reveals that general health condition, frequency of oral pain in the past year, marital status, history of mental health consultations in the past year, and frequency of urine leakage are the top five most influential factors in predicting the depression risk. CONCLUSION: This stacking ensemble model enhances the performance of both the primary learners and the meta-learners. This verifies the feasibility and effectiveness of the proposed model in predicting the geriatric depression. This work integrating the stacking ensemble model with SHAP offers valuable clinical references for assessing the risk of depressive symptoms, which is beneficial to develop the personalized depression interventions and preventions in the elderly.