Transcultural prediction model for late-life depression based on multi-cohort machine learning and explainable AI

Late-life depression is a global health concern with heterogeneous risk factors across populations. This study aimed to develop and validate a machine learning model for depression prediction in older adults using harmonized data from the United States and China.

We harmonized data from the Health and Retirement Study (HRS, n = 6865) and China Health and Retirement Longitudinal Study (CHARLS, n = 4476) for adults aged ≥60 years. Depression was assessed using validated scales in both cohorts. The Boruta algorithm was used for feature selection. 17 machine learning algorithms were evaluated, with HRS data split into training (70 %) and internal validation (30 %), and CHARLS data used for external validation. Model performance was assessed using AUC, decision curve analysis, calibration plots, and SHapley Additive Explanations (SHAP).

The Gradient Boosting Machine (GBM) model achieved the best performance, with AUCs of 0.752 (95 % CI: 0.735-0.768) in HRS training, 0.763 (95 % CI: 0.737-0.788) in HRS validation, and 0.717 (95 % CI: 0.702-0.732) in CHARLS validation. The model showed good calibration and positive net benefit across relevant clinical thresholds. SHAP analysis identified self-rated health, functional dependency, self-rated memory, arthritis, and ADL score as top predictors with consistent effects across populations.

We developed a robust and interpretable machine learning model for predicting late-life depression that generalizes across culturally distinct populations. The results highlight both common predictive factors and the need for population-specific considerations in clinical application.