A multi-database machine learning approach to predict depression and frailty in elderly populations with chronic conditions

Abstract

Depression and frailty frequently co-occur in older adults with chronic disease, significantly impairing physical and psychological health; however, their bidirectional association remains unclear. Data from four large population-based cohorts—CHARLS, HRS, KLOSA, and SHARE—were combined, encompassing 53,544 participants aged ≥ 60 years with multimorbidity. Twelve ML algorithms were employed to predict depression (Part 1) and frailty in 16,911 cases of psychosomatic comorbidity (Part 2), with model performance evaluated by accuracy, AUC, F1-score, and SHAP values. Among the models for predicting depression, the CatBoost model demonstrated the highest AUC value (0.773), with the frailty index emerging as the strongest predictor. For frailty prediction in patients with psychosomatic comorbidity, the Gradient Boosting model performed well (AUC = 0.756). Depression severity served as a primary determinant of frailty. DCA analysis indicated consistent net clinical benefits for both models. SHAP analyses identified the frailty index as a dominant predictive feature for concurrent depression, whereas depression severity was a strong predictor of concurrent frailty in cross-sectional observations. This multi-database explainable machine learning analysis reveals robust cross-sectional predictive associations between frailty and concurrent depression among older adults with multimorbidity. The predictive models facilitate clinical risk stratification for the concurrent screening of high-risk individuals for both conditions. However, the present cross-sectional design prohibits confirmation of temporal ordering and causal reciprocal reinforcement. The findings highlight the potential of ML-based predictive models in clinical practice for the early identification and management of high-risk individuals, enabling personalized care and interventions.