Predicting the risk of depression in older adults with disability using machine learning: an analysis based on CHARLS data.

Background: The advancement of artificial intelligence technologies has opened new avenues for depression prevention and management in older adults with disability (defined by basic or instrumental activities of daily living, BADL/IADL). This study systematically developed machine learning (ML) models to predict depression risk in disabled elderly individuals using longitudinal data from the China Health and Retirement Longitudinal Study (CHARLS), providing a potentially generalizable tool for early screening.

Methods: This study utilized longitudinal data from the CHARLS 2011-2015 cohort. A three-stage serial consensus approach feature selection framework (LASSO, Elastic Net, and Boruta) was employed to identify 21 robust predictors from 74 candidate variables. Ten ML algorithms were evaluated: LR, HistGBM, MLP, XGBoost, bagging, DT, LightGBM, RF, SVM, and CatBoost. Temporal external validation was performed using an independent 2018-2020 cohort to assess model generalizability. Performance was comprehensively evaluated using accuracy, AUC, F1-score, precision, and recall metrics. The SHAP framework was employed to interpret feature contribution mechanisms.

Results: Results demonstrated that the HistGBM model achieved optimal overall performance on the testing sets (AUC = 0.779, F1-score = 0.735, accuracy = 0.713), with only an 8.5% AUC difference between training and testing sets and a 10% difference between external validation and testing sets, indicating temporal stability. SHAP interpretability analysis revealed that sleep time (mean SHAP value = 0.344) in the health behavior domain and life satisfaction (0.339) and episodic memory (0.220) in the subjective perception domain contributed more significantly to prediction than traditional biomedical indicators.

Conclusion: This study developed an AI-based tool for depression risk assessment in older adults with disability through a multi-stage feature selection process and a temporal external validation framework. These findings provide a practical screening instrument and a methodological reference for implementing AI technologies in geriatric mental health applications, thereby facilitating clinical translation of predictive analytics in this field.