Beyond BMI: Chinese Visceral Adiposity Index (CVAI) improves machine learning prediction of 9-year cardiometabolic multimorbidity in the CHARLS cohort

Background Driven by rapid population aging and evolving lifestyle patterns across China, middle-aged and older individuals are at substantially elevated risk of cardiovascular disease (CVD) and diabetes mellitus (DM). The two disorders often co-occur and share metabolic risk determinants, collectively termed cardiometabolic multimorbidity (CMM). However, systematic predictive research targeting the composite endpoint of CMM remains relatively limited. Methods Based on data from the China Health and Retirement Longitudinal Study (CHARLS) spanning 2011 to 2020, three separate cohorts were established for DM (n = 7,903), CVD (n = 9,997), and CMM (n = 9,490). Four machine learning algorithms, namely logistic regression, random forest, XGBoost, and LightGBM, were applied to develop 9-year risk prediction models. Model performance was evaluated using AUC, calibration curves, and Brier score, and decision curve analysis, with SHAP analysis applied for model interpretability. Results The 9-year cumulative incidence rates of DM, CVD and CMM were 11.8%, 22.7% and 28.9%, respectively. DM models achieved the best performance (AUC 0.703–0.715), with XGBoost reaching the highest AUC (0.715). CVD prediction proved the most challenging (AUC 0.604–0.630), with Logistic Regression performing best (0.630). CMM models showed intermediate performance (AUC 0.621–0.629), with XGBoost achieving the highest AUC (0.629). SHAP analysis revealed that glycemic and obesity-related indices were key predictors for DM; age and blood pressure parameters dominated CVD prediction; and the Chinese Visceral Adiposity Index (CVAI) together with depressive symptoms (CESD) emerged as the top predictors for CMM. Simplified scorecards retained most predictive performance for DM (AUC 0.681) and CMM (AUC 0.625) but failed for CVD (AUC 0.436). Conclusions DM was predicted with the highest accuracy using routine clinical data, while CVD prediction remained suboptimal, suggesting that novel biomarkers or imaging are needed. The composite CMM endpoint showed intermediate predictability, with CVAI and depressive symptoms as key predictors, underscoring the roles of visceral adiposity and mental health. Simplified scorecards offer practical tools for DM and CMM screening in primary care, but their failure for CVD highlights the inherent complexity of cardiovascular risk stratification.