We gathered participants from the NHANES and CHARLS cohort studies, focusing on US and Chinese residents aged 45 years and above. The survey design and extensive data collection methods of the NHANES have been previously described (26). This study focused on individuals who reported having CVD and had complete baseline data on eGDR, TyG, TG/HDL-C, and METS-IR. These data were used to examine the association between IR indices and the current incidence of CVD. Accordingly, 22,922 individuals aged 45 years and above were obtained from the NHANES surveys between 2005 and 2018. Participants with missing information on CVD (n=189), eGDR (n=3,135), TyG, TG/HDL-C ratio, METS-IR (n=10,179), or other covariates (n=2,199) were excluded from the analysis ( Figure 1A ). A total of 44,462 participants from NHANES were included. We included participants with or without prevalent CVD to evaluate the relationship between CVD and IR indices in the NHANES.

Regarding the CHARLS, a detailed survey design and enrollment criteria were documented (27). A baseline survey was conducted from June 2011 to March 2012, selecting a nationally representative sample of 17,705 participants from 10,229 households. Participants were regularly followed up every two years through face-to-face interviews performed by the interviewer using a computer to administer and record responses to survey questions. The follow-up survey waves were performed in 2013, 2015, 2018, and 2020. We included participants who were not diagnosed with CVD at baseline and had complete follow-up results for IR indices. The incidence of CVD was assessed during follow-up investigations using, the latest data collected in Wave 4 of 2020. The CHARLS data were used to evaluate the association between these IR indices and the future incidence of CVD. Consequently, 17,314 individuals aged 45 years and above from the CHARLS were included. Participants with missing information on eGDR (n=7,564), TyG, TG/HDL-C, and METS-IR (n=267); those without follow-up information in the 2020 wave (n=1,388); those diagnosed with CVD at baseline (n=2), and those missing other covariates (n=316) were excluded from the analysis ( Figure 1B ). For the analysis of future CVD incidence among the CHARLS participants, we included only those who were free of prevalent CVD at baseline.

During the NHANES examination, certified examiners measured blood pressure (BP) multiple times using standardized protocols. Four readings were averaged to determine the BP. Additionally, body mass, height, and WC data were obtained. The BP of the participants were expressed as the average of several measurements in the NHANES and CHARLS. Blood samples were obtained from participants in both NHANES and CHARLS, and further measurements were conducted according to standard procedures. The biochemical parameters measured included blood urea nitrogen (BUN), uric acid (UA), hemoglobin, HbA1c, total cholesterol (TC), HDL-C, TG, and low-density lipoprotein cholesterol (LDL-C). Other covariates included age, sex, education, marital status, alcohol consumption, and smoking status (current smoker, former smoker [NHANES only], never smoker [NHANES only], or non-smoker [CHARLS only]).

Hypertension was diagnosed in participants who either self-reported a diagnosis of hypertension, were taking prescribed antihypertensive medications, or had a BP reading of ≥140/90 mmHg (28). The BMI was calculated using the following formula: BMI (kg/m²) = body mass/height². Obesity was defined as a BMI of 28 kg/m² or higher.

Four IR indices were evaluated to predict the CVD incidence: eGDR, TyG, TG/HDL-C ratio, and METS-IR. The formula for calculating eGDR was as follows: 21.158 − (0.09 × WC) − (3.407 × hypertension) − (0.551 × HbA1c) [WC (cm), hypertension (yes = 1/no = 0), and HbA1c (%)] (18, 29). TyG was using from the formula: ln[TG (mg/dL) × FBG (mg/dL)/2] (23). The TG/HDL-C ratio was calculated by dividing TG (mg/dL) by HDL-C (mg/dL) (23). METS-IR was calculated as follows: ln[(2 × FBG (mg/dL) + TG (mg/dL))× BMI/ln[HDL-C (mg/dL)] (30).

In this study, the primary outcome was CVD, encompassing both current and future heart disease and stroke. In line with standard precedents (31–33), incident CVD was identified through self-reports in which participants confirmed receiving a definitive diagnosis of CVD from their physicians. In NHANES, incident CVD data were collected during each survey cycle. Participants in the CHARLS were followed from the baseline in 2011 until the occurrence of CVD or until the most recent survey (2020), whichever occurred first.

Continuous variables are presented as mean ± standard deviation (SD) or median (interquartile range). Multivariate logistic regression analysis is a statistical tool that can be used to select and combine input variables linked to a certain outcome. Multivariate logistic regression models were used to evaluate the odds ratios (ORs) and 95% confidence intervals (CIs) for the relationships between the eGDR, TyG, TG/HDL-C, and METS-IR indices and CVD. Three multivariate models were constructed: Model 1 was an unadjusted model; Model 2 was adjusted for age, sex, marital status, education, smoking, and alcohol consumption status; and Model 3 was further adjusted region, TC, HDL-C, TG, LDL-C, BUN, UA, hemoglobin, and obesity. Restricted cubic splines (RCS) can be advocated as a potential alternative to these modelling strategies (categorizing a continuous variable or imposing the assumption of a linear association on a continuous variable) to explore non-linear continuous associations. The RCS has an additional property: the curve is linear before the first knot and after the last knot. RCS logistic regression analysis was used to explore the linearity and dose-response relationship between eGDR, TyG, TG/HDL-C, METS-IR and CVD. P for non-linear was cleated, and P for non-linear (<0.05) was considered a non-linear relationship. Covariates of interest and potential confounders were selected a priori based on the biological rationale and preexisting knowledge of risk factors for CVD, which can be obtained from NHANES and CHARLS. In this study, covariates were adjusted across the three models, with eGDR, TyG, TG/HDL-C, or METS-IR values at an OR of one serving as the reference point. Receiver operating characteristic (ROC) curves were generated to evaluate the predictive performance of the eGDR, TyG, TG/HDL-C, or METS-IR for CVD. The C-statistic was used to quantify the predictive value (34, 35). To further assess the enhanced predictive performance relative to the basic models, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indices were calculated (36). Decision curve analysis was used to compare the clinical benefits. Subgroup analyses were performed to evaluate the effect of eGDR, TyG, TG/HDL-C, and METS-IR (continuous and categorical variables) on CVD across various subgroups, which included age (< 60/≥ 60 years), sex (male/female), smoking status (current smoker/former smoker/never smoker/non-smoker), alcohol consumption (drinker/non-drinker), and obesity (yes/no). Sensitivity analysis was performed to assess the robustness of the results. The relationship between eGDR, TyG, TG/HDL-C, METS-IRand CVD was examined among participants without diabetes mellitus in model 3. Diabetes mellitus was defined in participants who either self-reported a diagnosis of diabetes mellitus, were taking prescribed antidiabetic medications, or had elevated FBG and HbA1c levels. All analyses were performed using R, version 4.2.1. Two-sided P values <0.05 were considered statistically significant.

This study included 7,220 participants from the NHANES cohort (mean age: 61.9 ± 10.7 years, 54.0% male) and 6,426 participants from the CHARLS cohort (mean age: 57.9 ± 8.44 years, 45.2% male). In the final NHANES cohort sample, the mean values were as follows: eGDR was 6.71 ± 2.60, TyG was 8.70 ± 0.61, TG/HDL-C was 2.61 ± 2.01, and METS-IR was 7.81 ± 0.42. For the CHARLS cohort, the mean values were: eGDR was 9.48 ± 2.26, TyG was 8.67 ± 0.65, TG/HDL-C was 3.13 ± 4.02, and METS-IR was 7.59 ± 0.41. Table 1 offers further details information on the characteristics of the included participants. Based on the results of NHANES, 1,265 cases of CVD were documented, including 1,006 cases of heart disease and 413 cases of stroke. As for CHARLS, there were 789 recorded cases of incident CVD, comprising 555 cases of heart disease and 276 cases of stroke.

Table 2 showed a significant relationship between the eGDR and the current or feature incidence of CVD when eGDR was analyzed as a continuous variable. In the NHANES study, higher eGDR were related to a lower risk of current CVD across all models, with OR and 95%CI of 0.78 (0.76, 0.80), 0.79 (0.77, 0.82), and 0.83 (0.81, 0.86), respectively. The results demonstrated a significant relationship between eGDR and both current heart disease or stroke, as shown in Supplementary Tables S1 and S2 . Compared to the lowest quartile (Q1) of eGDR, the ORs and 95% CIs for Q2−Q4 were 0.75 (0.63, 0.90), 0.51 (0.41, 0.62), and 0.30 (0.23, 0.39) for current CVD risk; 0.80 (0.66, 0.98), 0.53 (0.42, 0.66), and 0.33 (0.24, 0.44) for current heart disease risk; and 0.77 (0.58, 1.02), 0.59 (0.43, 0.80), and 0.27 (0.17, 0.42) for current stroke risk. These RCS curves for model 3 revealed a nonlinear association between eGDR and the risk of current CVD and heart disease (P-values for nonlinearity < 0.05, Supplementary Figure S1 ). However, a linear relationship was observed between eGDR and current stroke in both models (P-values for nonlinearity > 0.05).

Similar findings were observed in the CHARLS study, where higher eGDR levels were associated with a lower risk of future CVD in both models. The OR and 95% CI were 0.88 (0.85, 0.91), 0.89 (0.86, 0.92), and 0.91 (0.88, 0.94), respectively. The association between eGDR and the future risk of heart disease or stroke was also evident in both models. Compared to the lowest quartile (Q1) of eGDR, the ORs and 95% CIs for Q2−Q4 were 0.75 (0.63, 0.90), 0.51 (0.41, 0.62), and 0.30 (0.23, 0.39) for future CVD incidence; 0.80 (0.66, 0.98), 0.53 (0.42, 0.66), and 0.33 (0.24, 0.44) for future heart disease incidence; and 0.77 (0.58, 1.02), 0.59 (0.43, 0.80), and 0.27 (0.17, 0.42) for future stroke incidence. The RCS regression models indicated that eGDR and the future incidence of CVD, heart disease, and stroke are linear (P for nonlinearity > 0.05, Supplementary Figure S2 ).

The results of NHANES revealed that higher TyG was related to a higher risk of current CVD. When the TyG was treated as a continuous variable, the unadjusted model revealed an OR of 1.44 (95% CI: 1.30, 1.59), and the fully adjusted model showed an OR of 1.46 (95% CI: 1.16, 1.84) ( Table 2 ). A similar relationship between TyG levels and the current risk of heart disease was observed in both models ( Supplementary Tables S1 , S2 ). However, no significant relationship was observed between TyG and the current stroke. Participants in higher quartiles of TyG were found to have a higher risk of CVD in the current. Compared to the first quartile (Q1) of eGDR, the ORs with 95% CIs for the Q2−Q4 were as follows: 0.93 (0.76, 1.15), 1.11 (0.88, 1.41), and 1.40 (1.02, 1.92) for current CVD risk; and 0.95 (0.75, 1.20), 1.13 (0.88, 1.46), and 1.44 (1.02, 2.02) for current heart disease risk. Similar patterns were observed for TyG and the risk of stroke. The RCS regression models revealed that TyG and CVD or heart disease are nonlinear (P for non-linearity < 0.05, Supplementary Figure S2 ).

For CHARLS, the higher level of TyG was related to a higher future incidence of CVD in both models, with the ORs (95% CI) being 1.27 (1.14, 1.42), 1.27 (1.14, 1.42), and 1.33 (1.04, 1.70), respectively. The relationship between TyG and the future incidence of heart disease was observed in both models. However, no significant differences were found in the comparisons of TyG and the future risk of stroke. Compared to the fQ1 of TyG, the ORs with 95% CIs for future CVD incidence were 1.16 (0.92, 1.47) for Q2, 1.34 (1.05, 1.71) for Q3, and 1.37 (1.01, 1.88) for Q4. For future incidence of heart disease, the ORs were 1.07 (0.81, 1.40) for Q2, 1.26 (0.94, 1.68) for Q3, and 1.47 (1.02, 2.14) for Q4. For future stroke risk, only the Q4 of TyG showed a significant difference compared to the Q1, with an OR of 1.47 (1.02, 2.14). The RCS regression models revealed that TyG and the future incidence of CVD, heart disease, or stroke are linear (P for non-linearity > 0.05, Supplementary Figure S2 ).

The results revealed that, according to the results of the NHANES and CHARLS, the TG/HDL-C ratio was not related to CVD, heart disease, or stroke in either the current or future context ( Table 2 ; Supplementary Tables S1 , S2 ). Similarly, when the TG/HDL-C ratio was treated as a nominal variable, most comparisons yielded consistent results. The RCS curves indicated a linear relationship between TG/HDL-C and the incidence of CVD, heart disease, or stroke, both currently and in the future, regardless of whether covariates were adjusted for (P for non-linearity > 0.05, Supplementary Figure S3 ).

NHANES results showed that higher METS-IR were related to a higher risk of current CVD. In the unadjusted model, the OR was 2.14 (95% CI: 1.85, 2.47), while in the fully adjusted model, the OR was 2.03 (95% CI: 1.51, 2.71). The relationship between METS-IR and current heart disease was also observed in both models ( Supplementary Tables S1 , S2 ). However, the fully adjusted model revealed no significant association between METS-IR and current stroke. When METS-IR was treated as a nominal variable, only the Q4 showed a significant difference compared to the Q1, with ORs of 1.49 (95% CI: 1.08, 2.05) for CVD and 1.63 (95% CI: 1.14, 2.32) for heart disease. Similar results revealed no significant relationship between METS-IR and current stroke. The RCS curves revealed a nonlinear relationship between METS-IR and current CVD or heart disease in the adjusted model (P for non-linearity < 0.05, Supplementary Figure S4 ).

In the CHARLS study, higher METS-IR were related to an increased incidence of future CVD in both models. Compared with Q1, the ORs with 95% CIs were 1.57 (1.32, 1.87) for Q2, 1.64 (1.37, 1.95) for Q3, and 2.08 (1.45, 2.99) for Q4. The association between METS-IR and the incidence of future heart disease and stroke was also revealed in both models. Compared to the Q1 of METS-IR, the fully adjusted ORs with 95% CIs for the Q2−Q4 were as follows: for future CVD: 1.25 (0.98, 1.59) for Q2, 1.70 (1.31, 2.20) for Q3, and 1.73 (1.25, 2.40) for Q4; for future heart disease: 1.06 (0.80, 1.40) for Q2, 1.50 (1.11, 2.01) for Q3, and 1.63 (1.11, 2.38) for Q4; for future stroke: 1.80 (1.18, 2.79) for Q2, 2.09 (1.34, 3.32) for Q3, and 1.91 (1.11, 3.31) for Q4. The RCS regression models revealed that METS-IR and the incidence of future CVD heart disease, and stroke are linear. (P for non-linearity > 0.05, Supplementary Figure S4 ).

Stratified analyses were conducted to determine if the relationship between eGDR, TyG, TG/HDL-C, or METS-IR and current or feature CVD was not influenced by some of the subgroups. The association between eGDR, TyG, TG/HDL-C, or METS-IR and CVD was consistent with the main results across most subgroups ( Figure 2 ). According to the NHANES results, the relationship between eGDR and current CVD risk varied across different subgroups of alcohol consumption (P for interaction = 0.009). According to CHARLS results, the relationship between METS-IR and future CVD incidence varied across different subgroups of obesity (P for interaction = 0.036). Similar patterns were observed for heart disease and stroke outcomes ( Supplementary Figures S5 , S6 ). No significant interactions were observed between quartiles of eGDR, TyG, TG/HDL-C, or METS-IR and the current risk and future incidence of various endpoints in most subgroups ( Supplementary Figures S7 - S9 ). In the sensitivity analyses, excluding participants with diabetes, as defined by FBG and HbA1c measurements, did not substantially alter the results ( Supplementary Table S3 ).

In this study, the basic models were established to include age, sex, rural residence, marital status, education, smoking status, alcohol consumption status, TC, HDL-C, LDL-C, TG, LDL, BUN, UA, hemoglobin, and obesity. Incorporating eGDR, TyG, TG/HDL-C, and METS-IR significantly improved the predictive performance of the basic model for CVD in both NHANES and CHARLS ( Table 3 ; Figures 3A, B ). The decision curve analysis further confirmed the clinical relevance of these additions ( Figures 4A, B ). According to the results, eGDR outperformed TyG, TG/HDL-C, and METS-IR in predicting both current and future CVD. Notably, the area under the curve (AUC) for predicting current CVD was higher than for predicting future CVD. Incorporating eGDR, TyG, TG/HDL-C, and METS-IR significantly improved the predictive ability of the basic model for heart disease in both NHANES and CHARLS ( Supplementary Table S4 ; Supplementary Figure S10 ). The decision curve analysis further confirmed the clinical relevance of these additions ( Supplementary Figure S11 ). For stroke, only the addition of eGDR significantly improved the predictive performance of the basic model for current stroke risk in NHANES ( Supplementary Table S5 ; Supplementary Figure S12 ). The decision curve analysis confirmed its clinical relevance ( Supplementary Figure S13 ). Similar findings indicated that the eGDR index outperformed other indices, with the AUC for predicting heart disease or stroke being higher than for predicting their incidence. Additionally, all net NRI and IDI metrics for CVD and heart disease were significant (P < 0.001) ( Table 3 ; Supplementary Table S4 ), and most NRI and IDI metrics for stroke were also significant ( Supplementary Table S5 ). Although combined hypertension and the basic model improved the predictive performance for all endpoints, it was still less effective compared to the inclusion of eGDR.

This study had several strengths First, this study was the first study to assess the predictive performance of eGDR, TyG, TG/HDL-C, or METS-IR for current or feature incidence of CVD among middle-aged and elderly participants. Second, our results revealed that the eGDR, TyG, and METS-IR significantly improved the predictive value of the basic models for current or future incidents of CVD, which is expected to inform future CVD predictors. Furthermore, eGDR significantly improved the predictive value of the basic models for incident stroke. Third, both AUCs for predicting CVD were higher than the participants’ CVD incidence. These results reveal the superior value of the combined IR indices compared to that of the basic model. Furthermore, eGDR can be considered the most effective index (vs. basic model, basic model plus hypertension, TyG, TG/HDL-C, and METS-IR) for the current and feature incidences of CVD in middle-aged and elderly individuals.

This study had several limitations should be noted. First, as with any observational study, causality cannot be establish and there is a possibility of reverse causality. However, this limitation appears to have a minimal impact as our analysis included individuals who experienced the endpoint both currently and in the future. Second, although the model included many covariates, residual confounding factors cannot be entirely ruled out. However, thishis is a common challenge in observational cohort studies. Third, CVD in this study was self-reported by the participants based on physician diagnoses, which may have led to recall bias and misclassification. However, this approach is commonly used in cohort studies, and evidence suggests that its impact is minimal (64). Fourth, due to the lack of fasting insulin data, we could not compare the risk of CVD using HOMA-IR with the four proposed IR indices. Finally, the IR indices are known to change dynamically over time. However, owing to cost and other limitations, we could only obtain baseline data for our study.