Cross-national predictive correlates of depressive symptoms among middle-aged and older adults with chronic conditions: an explainable machine learning analysis.

BACKGROUND: Rapid population aging has made depression among middle-aged and older adults an increasingly important public health concern. Although numerous risk factors have been identified, an integrative framework for clarifying how these risks are structured across sociocultural contexts remains lacking. Identifying cross-nationally stable vulnerability signals is therefore important for improving screening and prevention. This study aimed to identify cross-nationally stable and context-specific predictive correlates of depressive symptoms among middle-aged and older adults with at least one chronic condition using an explainable machine-learning framework. METHODS: Data were drawn from six international longitudinal cohorts-CHARLS, ELSA, HRS, KLoSA, MHAS, and SHARE-including 84,140 participants aged 50 years or older who reported at least one chronic condition. Within an explainable artificial intelligence framework, six predictive algorithms were evaluated: LR, RF, SVM, MLP, XGBoost, and EBM. SMOTE was used to address class imbalance, and SHAP analysis was applied to characterize the hierarchical structure of risk determinants. RESULTS: The models achieved moderate predictive performance across cohorts, with area under the receiver operating characteristic curve values ranging from 0.77 to 0.80. Explainable boosting machine and logistic regression performed nearly identically in most cohorts, with absolute AUC differences below 0.005 in five of six cohorts. A set of cross-nationally stable predictive correlates was identified, among which poor self-rated health, pain burden, and limitations in activities of daily living and instrumental activities of daily living were the most influential predictors. Conversely, demographic and socioeconomic factors showed greater cross-national heterogeneity, indicating that depression risk is shaped by both shared and context-specific influences. CONCLUSIONS: By integrating large-scale international cohort data with explainable machine learning, this study identified cross-national patterns of feature importance for concurrent depressive symptom status among middle-aged and older adults with chronic conditions. Poor self-rated health, somatic pain, and functional limitations may support cross-sectional screening and further assessment but should not be interpreted as early warning indicators.