Triglyceride-glucose-Chinese visceral adiposity index (TyG-CVAI) outperforms other TyG-derived indices in predicting 5-year advanced cardiovascular-kidney-metabolic syndrome: a prospective community-based cohort and machine learning study in China

The Cardiovascular-Kidney-Metabolic (CKM) syndrome is a recently proposed integrated disease concept. Its advanced stages (Stages 3–4) are characterized by subclinical or clinical cardiovascular disease, severe chronic kidney disease, and significantly elevated mortality risk. Triglyceride-glucose(TyG) and TyG-derived indices can comprehensively reflect insulin resistance (IR) and obesity; however, the predictive performance of these indices for advanced CKM syndrome has not been systematically compared. This study aimed to compare the predictive value of 10 TyG and TyG-derived indices for 5-year advanced CKM syndrome and to identify the optimal index. This prospective community-based cohort study (Yulin, Guangxi, China) included 2072 participants free of advanced CKM at baseline (median follow-up: 54.7 months). Ten TyG and TyG-derived indices were calculated (TyG, TyG-WC, TyG-WHtR, TyG-BMI, TyG-BRI, TyG-WWI, TyG-LAP, TyG-ABSI, TyG-VAI, TyG-CVAI). Restricted cubic spline (RCS) regression assessed dose–response relationships, and Cox proportional hazards regression estimated effect sizes. Incremental predictive value was evaluated using a clinical reference model (Model Clinical) and a simplified model (Model 2) as a sensitivity analysis. IDI and time-dependent AUC were the primary metrics, with NRI as secondary (1000 bootstrap resamples). Three feature selection methods (LASSO, Boruta, and RF-RFE) were jointly applied within the training set to identify the optimal TyG index, and their stability was assessed using 1000 bootstrap resamples. Six machine learning algorithms (RF, XGBoost, LightGBM, MLP, AdaBoost, and Bagging) were used to construct prediction models. The optimal model was selected based on test-set discrimination, calibration (Hosmer–Lemeshow P ≥ 0.05), and clinical utility, with the Youden index used for threshold determination. Model performance was assessed using AUC, calibration plots, Brier score, and decision curve analysis (DCA). A risk stratification cutoff was derived from the training set using the Youden index and applied to the test set. SHAP was used to interpret the final optimal model. Internal validation used 10-fold cross-validation and bootstrap optimism correction (1000 resamples); sensitivity analyses included inverse probability of censoring weighting (IPCW), exclusion of participants with hypertension, diabetes, or medication use, additional adjustment for baseline CKM stage, and separate analyses for Stage 3 and Stage 4 outcomes, and redefinition of advanced CKM excluding Stage 3 events defined solely by eGFR < 30 mL/min/1.73 m 2. Multiple comparison corrections (FDR and Bonferroni) were applied. Subgroup analyses were performed to assess consistency across age, gender, smoking, drinking, and exercise strata. After a 5-year follow-up, 247 of the 2,072 participants developed incident advanced CKM events (incidence density: 2.79 per 100 person-years). RCS analysis revealed that, except for TyG-ABSI, which exhibited a U-shaped non-linear association ( P for nonlinearity = 0.037, crossing points: 0.651 and 0.712), the remaining nine indices showed significant linear dose–response relationships with advanced CKM risk (all P for overall < 0.05, all P for nonlinearity > 0.05). In the fully adjusted model, TyG-CVAI demonstrated the largest effect size (per 1-SD increase: HR = 1.290, 95% CI 1.143–1.456, P < 0.001). In the incremental predictive analysis using Model Clinical, TyG-CVAI showed the highest continuous NRI (0.363, 95% CI 0.198–0.493, P < 0.001) and a modest but statistically significant IDI (0.009, 95% CI 0.001–0.022, P = 0.016). At 60 months, adding TyG-CVAI yielded a small yet statistically significant increase in AUC (from 0.789 to 0.799; ΔAUC = 0.010, P < 0.001). This was the largest improvement among all indices. All three feature selection methods consistently identified TyG-CVAI as the optimal index. Based on comprehensive evaluation of test set AUC, calibration curves, and Brier score, the XGBoost model (AUC = 0.839, 95% CI 0.796–0.883, Brier score = 0.084, Hosmer–Lemeshow P = 0.200) achieved the best balance between discrimination and calibration and was selected as the final optimal model. Internal validation showed a 10-fold cross-validated AUC of 0.855 ± 0.038 and a bootstrap optimism-corrected AUC of 0.870, indicating low overfitting risk. DCA demonstrated positive net benefit of the XGBoost model compared to the "treat all" strategy within the threshold probability range of 0.10–0.40. Calibration curves showed good agreement between predicted probabilities and observed outcomes. SHAP analysis of the final six-feature XGBoost model illustrated the contributions of age, gender, hypertension, diabetes, LDL-C, and TyG-CVAI. Using a cutoff of 0.14, the test set was stratified into low- and high-risk groups, with 5-year cumulative incidences of advanced CKM of 5.9% vs. 38.1% (HR = 8.356, 95% CI 5.105–13.676, P < 0.001). Subgroup analyses showed consistent associations across strata (all P for interaction > 0.05). Sensitivity analyses, including exclusion of participants with hypertension, diabetes, or medication use, additional adjustment for baseline CKM stage, and separate analyses for Stage 3 and Stage 4, yielded consistent results. These findings remained robust after FDR and Bonferroni correction, and TyG-CVAI was selected in 100% of bootstrap resamples. Among the 10 TyG and TyG-derived indices evaluated, TyG-CVAI demonstrated superior predictive performance for 5-year advanced CKM syndrome, with the largest effect size and highest continuous NRI, although its incremental value beyond established clinical predictors was modest. These findings were internally consistent across multiple machine learning approaches. TyG-CVAI may serve as a low-cost auxiliary tool for identifying individuals at high risk of advanced CKM in community settings, pending external validation.