Construction of risk prediction model for depression among Chinese older adults: An empirical study based on charls database

Objective To identify predictors of depressive symptoms among Chinese older adults and develop an interpretable machine learning prediction model. Methods Data from 14,469 participants in CHARLS 2018–2020 were analyzed. Depressive symptoms in 2020 were defined using the CES-D-10 scale. After feature engineering and Random Forest-based feature selection, 25 predictors were used to train six machine learning models. Model performance was evaluated using Accuracy, Precision, Recall, F1-score, AUROC, AUPRC, and Brier score. Recall-oriented threshold tuning and SHAP analysis were also conducted. Results After AUPRC-based hyperparameter tuning, the Stacking Ensemble model showed the best overall performance under the default threshold, with an Accuracy of 0.731, AUROC of 0.758, AUPRC of 0.603, and Brier score of 0.182. Its Recall was 0.412 under the default threshold but improved to 0.690 after threshold tuning. SHAP analysis identified pain distress, education level, sleep duration, self-rated health, gender, basic ADL difficulties, and 10-year survival expectation as key contributors. Conclusion The Stacking Ensemble model showed moderate predictive ability for depressive symptoms among Chinese older adults. It may support preliminary community-based identification of high-risk individuals, but further validation is needed before clinical application.