Development and validation of an interpretable machine learning model for predicting the risk of 8-year all-cause mortality in Cardiovascular-Kidney-Metabolic Syndrome among older adults: A multicenter and cohort study.

Background: Cardiovascular-kidney-metabolic syndrome (CKM) among elder adults due to age-related physiology is a high burden, but long-term mortality risk prediction is understudied. This study aims to develop and validate an explainable machine learning model to predict 8-year all-cause mortality. Methods: This study used HRS (2012-2020) and CHARLS (2011-2020) database, performed data cleaning and multiple imputation, plotted Kaplan-Meier curves with log-rank tests, conducted competing risk analyses, and pooled estimates using Rubin's rules. Cox regression was used to adjust for confounders. In CHARLS, variables were selected using LASSO (retained if selected ≥ 70%), after sensitivity and collinearity checks. We compared six survival models with 10-fold cross-validation and evaluated performance with AUC, DCA, calibration curves, and the Brier score, with external validation in HRS. SHAP was used to explain feature importance. Results: 8,473 participants (CHARLS 4,460; HRS 4,013) was included. Pooled 8-year mortality by CKM stages 0-4 was: HRS 9.2%, 4.94%, 9.14%, 12.57% and 18.97%; CHARLS 15.69%, 9.92%, 16.76%, 25.51% and 26.19%. Each one-stage increase in CKM was associated with a 40% and 15% higher mortality risk (both P < 0.001). The Cox model performed best: internal AUC 76.7% (95%CI: 76.5%-76.8%); external AUC 76.6% (95%CI: 76.2-77.0). Top three SHAP features: age, smoking and cystatin C. Conclusions: The Cox model showed good discrimination, calibration, and interpretability for predicting 8-year mortality risk in older adults with CKM syndrome across Chinese and American populations. Research findings indicate that interventions targeting the predictors cystatin C and gait speed may help reduce the long-term risk of mortality.