Abstract
This study aims to identify depressive risks in elderly individuals with subjective cognitive decline (SCD) and develop a predictive model using machine learning algorithms to enable timely interventions.Data from the 2015 and 2018 waves of the China Health and Retirement Longitudinal Study (CHARLS) were used, including 1,921 elderly individuals. Depression was assessed with the CESD-10 scale. Three machine learning models-Gradient Boosting, Random Forest, and Boosted XGBoost-were used to predict depression risk over three years, incorporating 10 demographic, 5 health, 13 chronic disease, 3 lifestyle, and 2 physical function factors. Lasso feature selection identified 10 key variables for model training. Model performance was evaluated using ROC curves, AUC, sensitivity, specificity, accuracy, calibration, and decision curve analysis. Among all evaluated models, Boosted XGBoost demonstrated the highest predictive accuracy in the test set (AUC = 0.893), outperforming both Gradient Boosting (AUC = 0.887) and Random Forest (AUC = 0.861). However, Random Forest (RF) achieved superior sensitivity. Consequently, we performed feature importance analysis using both Boosted XGBoost and RF models. The results identified five significant predictors of depression in older adults with subjective cognitive decline (SCD): educational attainment, digestive health status, arthritis diagnosis, sleep duration, and residential location.The machine learning model developed in our study demonstrates strong predictive performance for depression risk among older adults with subjective cognitive decline (SCD), enabling early identification of high-risk individuals. These findings provide a scientific foundation for understanding depression progression mechanisms and developing personalized intervention strategies.