Baseline characteristics analysis of two cohorts

A total of 132,125 participants were included in this study, with 88.3% of the study population from the Chinese cohort and the remaining 11.7% from the Japanese cohort. The baseline characteristics of the two cohort populations were summarized in Table 1. Overall, the distribution and proportions of most baseline characteristics were comparable between the Chinese and Japanese cohorts, such as age (median age 41 vs 42), sex composition (male 53.80% vs 54.51%; female 46.20% vs 45.49%), height, ALT and HDL-C levels, and the proportion of drinkers. However, there were some differences between the two cohorts in certain measured parameters and biochemical indicators. It can be observed that, compared to the Japanese population, the Chinese population generally had higher weight and BMI, higher baseline levels of SBP, DBP, TG, AST, and the four IR surrogates, and relatively lower levels of FPG, TC, and LDL-C.

Table 1 Characteristics regarding the study variables at baseline in two cohort studiesStandardized HRs for the associations between the four IR surrogates and diabetes

In the two independent cohorts, the average follow-up time was 3.1 years (maximum 7.56 years) in the Chinese cohort, with 2681 incident cases (2.30%) of diabetes recorded, and 6.13 years (maximum 13.14 years) in the Japanese cohort, with 373 incident cases (2.41%) of diabetes recorded.

In the screening for collinearity between the four IR surrogates and baseline variables (Additional file 1: Tables S1–4), we identified that weight, BMI, TC, TG, and drinking status were collinear and were not included in the subsequent multivariable models. Furthermore, the Kaplan–Meier curve analysis, incorporating follow-up time and exposure variables, indicated that the Cox proportional hazards models constructed to assess the associations between the four IR surrogates and diabetes were appropriate for this study (Additional file 1: Figures S1–4).

The results of univariate Cox regression analysis indicated that, except for drinking status, all other baseline indicators were significantly associated with the risk of diabetes. Among them, having a smoking habit and higher levels of age, height, weight, BMI, SBP, DBP, FPG, TC, TG, LDL-C, ALT, AST were significantly associated with an increased risk of diabetes. Conversely, higher levels of HDL-C and being female were associated with a lower risk of diabetes (Additional file 1: Table S5). Subsequently, the independent associations between the IR surrogates and diabetes were demonstrated in three sequentially adjusted multivariable Cox regression models (Table 2). Consistent with the majority of previous findings, after adjusting for potential confounders, all four IR surrogates showed a significant positive association with diabetes. Based on the standardized HRs, TyG-BMI and MetS-IR exhibited the strongest associations with diabetes among the four IR surrogates (Model 3, HR: TyG-index 1.33 vs TyG-BMI 1.51 vs TG/HDL-C ratio 1.37 vs MetS-IR 1.51). Furthermore, after stratifying by ethnicity, this stronger association was still observed in the Chinese population (Model 3, HR: TyG-index 1.33 vs TyG-BMI 1.48 vs TG/HDL-C ratio 1.37 vs MetS-IR 1.47) and the Japanese population (Model 3, HR: TyG-index 1.23 vs TyG-BMI 1.79 vs TG/HDL-C ratio 1.26 vs MetS-IR 1.86). These findings suggested that both TyG-BMI and MetS-IR may be more useful in assessing the risk of diabetes, whether in the Chinese or Japanese population.

Table 2 Cox regression analyses for the association between TyG index, TyG-BMI, TG/HDL-C ratio, MetS-IR and incident DM in different models

Furthermore, to validate the stability of the aforementioned association analysis results, we calculated the E-values for each IR surrogate based on Model 3. In the All cohort, the E-values for TyG index, TyG-BMI, TG/HDL-C ratio, and MetS-IR were 1.99, 2.39, 2.08, and 2.39, respectively. Upon further differentiation by ethnicity, in the Chinese population, the E-values for TyG index, TyG-BMI, TG/HDL-C ratio, and MetS-IR were 1.99, 2.32, 2.08, and 2.3, respectively; in the Japanese population, the E-values for TyG index, TyG-BMI, TG/HDL-C ratio, and MetS-IR were 1.76, 2.98, 1.83, and 3.12, respectively. We observed that the E-values for these IR surrogates were all relatively large and higher than the corresponding HR values, indicating that it is unlikely that unmeasured confounders were affecting the stability of the results.

Association analysis of the four IR surrogates with diabetes in predefined subgroups

In predefined subgroups representing common phenotypes, we further explored the associations between the four IR surrogates and diabetes in the Chinese and Japanese populations. Detailed results were shown in Table 3. Across all subgroups, a consistent finding was observed: TyG-BMI and MetS-IR had the highest associations with diabetes among the four IR surrogates, which aligned with the results obtained from the overall population analysis and the analysis stratified by ethnicity (Table 2). This finding further supported the notion that TyG-BMI and MetS-IR may be the strongest IR surrogates for assessing the risk of diabetes.

Table 3 Stratified association between TyG index, TyG-BMI, TG/HDL-C ratio, MetS-IR and diabetes by age, BMI, sex, and hypertensionROC analysis to evaluate the predictive value of the four IR surrogates for diabetes

Table 4 presents the AUC values of the four IR surrogates for predicting diabetes. It can be observed that in the overall population (Fig. 2), TyG-BMI had the highest AUC value for predicting diabetes compared with other IR surrogates (0.7707, all DeLong test P < 0.05), followed by MetS-IR (0.7596), TyG-index (0.7589), and TG/HDL-C ratio (0.7021). After stratifying by ethnicity, we found that in the Chinese population, TyG-BMI had the highest AUC value (0.7741, all DeLong test P < 0.05), while in the Japanese population, although TyG-BMI and MetS-IR had relatively higher AUC values (0.7738 and 0.7796) (Fig. 3), after performing the DeLong test for comparison, we found that they were only significantly higher than TG/HDL-C ratio, with no statistically significant difference from TyG index.

Table 4 Predictive value of four IR surrogates for the diabetesFig. 2

ROC curve for predicting diabetes using TyG index, TyG-BMI, TG/HDL-C ratio, MetS-IR. TyG index the triglyceride-glucose index, TyG-BMI triglyceride glucose-body mass index, TG/HDL-C ratio triglyceride/high-density lipoprotein cholesterol ratio, METS-IR metabolic score for insulin resistance

Fig. 3

ROC curves for predicting diabetes using TyG index, TyG-BMI, TG/HDL-C ratio, MetS-IR in Chinese cohort and Japanese cohort. TyG index the triglyceride-glucose index, TyG-BMI triglyceride glucose-body mass index, TG/HDL-C ratio triglyceride/high-density lipoprotein cholesterol ratio, METS-IR metabolic score for insulin resistance

Analysis of the predictive value of the four IR surrogates for diabetes in predefined subgroups

To evaluate whether the predictive value of the four IR surrogates for diabetes varied across different populations, we further conducted ROC analysis in predefined subgroups. Detailed results are presented in Table 5. Overall, consistent with the findings from the stratification by ethnicity in Table 4, we observed similar patterns in the ROC analysis of the subgroups. In the Chinese cohort, TyG-BMI demonstrated the best predictive value for diabetes in almost all subgroups, with statistically significant differences observed in the majority of AUC values after conducting the DeLong test for comparison; while in the Japanese cohort, MetS-IR had the highest AUC values for predicting diabetes in all subgroups, followed by TyG-BMI. However, the differences between the two were statistically insignificant in almost all subgroups, and only in the male subgroup, obese subgroup, age < 50 years subgroup, and no hypertension subgroup, both MetS-IR and TyG-BMI exhibited significantly higher AUC values than TyG index and TG/HDL-C ratio. Additionally, in both cohorts, the predictive accuracy for predicting diabetes was higher in females compared to males, in non-obese individuals compared to obese individuals, in younger individuals (Age < 50 years) compared to older individuals (Age ≥ 50 years), and in individuals with normal baseline blood pressure compared to those with hypertension.

Table 5 The best threshold, sensitivities, specificities, and area under the curve of four IR surrogates for predicting diabetes in different subgroupsTime-dependent ROC analysis of IR surrogates for prediction of future diabetes

To further compare the predictive abilities of TyG index, TyG-BMI, TG/HDL-C ratio, and MetS-IR for future diabetes at different time intervals, we conducted time-dependent ROC analysis in the Japanese population and presented the results in Additional file 1: Table S6. By comparing the AUC values of the four IR surrogates for predicting diabetes occurrence over 1–13 years, we found that MetS-IR had the highest AUC values at almost all time points, and TyG index and TyG-BMI exhibited similar predictive abilities, while TG/HDL-C ratio had relatively weaker predictive ability, consistent with the previous analyses. Therefore, in the Japanese population, MetS-IR may be a better surrogate marker for diabetes prediction.