This study employed a retrospective cohort design, utilizing data from a Chinese computer database by researchers Chen et al. (21). The original data for this study were sourced from the DATADRYAD database (www.datadryad.org). The data concerning Chinese individuals came from a published article entitled “Association of body mass index and age with incident diabetes in Chinese adults: a population-based cohort study” referred to as the Dryad dataset (https://doi.org/10.5061/dryad.ft8750v) (21). This research respected the principles outlined in the Declaration of Helsinki, with all procedures aligning with the relevant protocols and rules as specified in the declaration segment. As a result of its retrospective design, ethical consent or informed approval was not necessary from the institutional review board for the analysis of this secondary dataset. Dryad’s terms of service permit the secondary analysis of data by other researchers without infringing upon the authors’ rights.

Figure 1 illustrates the initial inclusion of 685,277 participants in the Chinese cohort, with 473,744 excluded from the primary study, leaving 211,833 for analysis. The exclusion criteria for the current analysis were as follows: (i) participants lacking baseline TG data were excluded (n=5,748); (ii) participants lacking baseline HDL-C data were excluded (n=88,001); (iii) participants had BMI≥25kg/m2 (Non-obesity is defined as BMI <25 kg/m²⁾ (20) were excluded (n=35,108); The final participant count was 82,977 comprising 1041 with DM and 81,936 with non-DM.

The exposure variable in this study is the initial AIP, which is calculated based on the participants’ TG and HDL-C levels using the formula: AIP = log (TG/HDL-c) (22, 23). The criteria for incident DM included self-report and a FPG level ≥7.0 mmol/l measured at the last follow-up evaluation. The earliest result was used as the basis for diagnosis. Additionally, the follow-up period spanned five years, with a minimum follow-up duration of 2 years (21). The primary outcome was determined during the follow-up period and was whether participants were diagnosed with diabetes, which was recorded as a binary variable (0 = non-DM, 1 = DM).

In this study, we collected data from both Chinese populations, focusing on shared variables such as demographic characteristics (age and gender), fasting plasma glucose (FBG), BMI, alanine aminotransferase (ALT), systolic blood pressures (SBP), TG, total cholesterol (TC), diastolic blood pressures (DBP), HDL-c, blood urea nitrogen (BUN), serum creatinine (Scr), family history of diabetes, as well as follow-up duration. We also collected smoking status, categorized as current smoker, former smoker, never smoker, and unknown. Additionally, we collected alcohol consumption status, categorized as current drinker, former drinker, never drinker, and unknown. Fasting venous blood samples were collected after a minimum 10-hour fast at each visit. Blood pressure was measured using a standard mercury sphygmomanometer. BMI was calculated as weight in kilograms divided by the square of height in meters. Covariates were selected based on clinical experience and published literature, leading us to include the following variables: continuous variables such as DBP, FBG, BMI, HDL-C, SBP, TG, TC, and ALT; and categorical variables such as gender, smoking status, drinking status and family history of diabetes.

In this study, the number of participants with missing data for DBP, SBP, TC, LDL, ALT, BUN, and Scr were 9 (0.00%), 9 (0.00%), 1 (0.00%), 0 (0.00%), 301 (0.36%), 2170 (2.62%), and 979 (1.17%), respectively. The study used multiple imputation to handle the missing data to reduce the variability caused by missing variables. DBP, SBP, age, ALT, gender, LDL-c, TG, HDL, Scr, BUN, TC, FPG, alcohol consumption status, smoking status, and family history of diabetes were all included in the imputation model (24).

The study included participants who were classified into four groups based on their AIP values. The mean ± standard deviation (for normally distributed data) or medians (with interquartile ranges) (for skewed data) were reported for continuous variables. Categorical data was presented as frequencies and percentages. The analysis involved using the χ2 test for categorical variables, and either one-way analysis of variance (ANOVA) (for normal data) or the Kruskal-Wallis H test (for skewed data) to compare differences between the AIP groups. Survival rates and time-to-event variables were determined through the Kaplan-Meier method, and the log-rank test was employed to compare diabetes-free survival among the AIP groups.

We used univariate and multivariate Cox proportional hazards regression models to investigate the association between AIP and diabetes risk. This included a crude model without adjusted covariates, a model adjusted for the minimum covariates (Model I: adjusted for gender and age), and a fully adjusted model (Model II: adjusted for gender, age, SBP, DBP, BMI, family history of diabetes, drinking status, smoking status, LDL-C, ALT, Scr, BUN, and FPG). We recorded the effect size (HR) and its 95% confidence interval (CI). We adjusted for confounding factors based on clinical experience, literature reports, and the results of univariate analysis. Additionally, to reduce the impact of variables on model stability, we performed collinearity screening and found that the VIF (variance inflation factor) for TC was 8.1, which is greater than 5, indicating a potential collinearity issue (25). Therefore, TC was excluded from the final multivariate Cox proportional hazards regression equation ( Supplementary Table S1 ).

Furthermore, using cubic spline functions and smooth curve fitting methods within the Cox proportional hazards regression model, we considered the nonlinear association between AIP and diabetes risk. We also employed a piecewise Cox proportional hazards regression model to elucidate the nonlinear association between AIP and diabetes risk. Finally, we conducted a likelihood ratio test to select the best model to explain their association in non-obese populations. We performed subgroup analyses using stratified Cox proportional hazards regression models across different groups (age, gender, BMI, SBP, DBP, family history of diabetes, smoking status, and drinking status). First, continuous data (such as age) were converted into categorical variables based on clinical cutoffs (age: <65 years, ≥65 years). In addition to the stratifying factors themselves, we adjusted for gender, age, SBP, DBP, BMI, family history of diabetes, drinking status, smoking status, LDL-C, ALT, Scr, BUN, and FPG. Ultimately, in models with and without interaction terms, we used the likelihood ratio test to identify the presence of interaction terms. To examine the reliability of the results, we conducted a series of sensitivity analyses. Continuous covariates were included in the equation and were modeled using generalized additive models (GAM) to confirm the reliability of the findings (26). Additionally, we calculated E-values to assess the potential impact of unmeasured confounding on the association between AIP and diabetes risk (27).

The R software package (http://www.r-project.org, R Foundation) and Empower Stats (X&Y Solutions, Inc., Boston, MA, http://www.empowerstats.com) were utilized for the conducted analyses. Statistical significance was determined with a P-value below 0.05.

Table 1 illustrates the demographic and clinical features of the study participants. The mean age was found to be 43.04 ± 12.82 years, with 38,011 individuals (45.81%) classified as male. The median duration of follow-up was 3.10 years, during which time 1041 subjects (1.25%) received a diabetes diagnosis. The AIP ranged from -1.96 to 2.06, resulting in a mean level of -0.42 ( Supplementary Figure S1 ). Table 1 summarizes the baseline characteristics of 82,977 participants, categorized by quartiles of the AIP. The first group included 20,740 participants, the second group had 20,735, the third group consisted of 20,755, and the fourth group comprised 20,747 participants. As AIP increased, the average age of participants significantly rose, with a clear change observed from the first group to the fourth group. Additionally, BMI also increased, reflecting a corresponding trend. SBP and DBP were significantly higher in the fourth group compared to the first group. The FBG levels also showed an upward trend, while TG levels increased substantially. Meanwhile, HDL-C levels exhibited a downward trend across the four groups. The incidence of diabetes increased from 0.70% in Q1 to 2.48% in Q4 (p < 0.001), highlighting a strong association between high AIP and diabetes risk. In addition, we divided the population into a non-DM group and a DM group based on diabetes status. Supplementary Table S2 presents the baseline characteristics of participants in both groups. Except for alcohol consumption status, all other variables showed significant differences.

Table 2 and Supplementary Figure S2 illustrate the incidence rates of diabetes. Among the study participants, 1,041 individuals (1.25%) were reported to have developed diabetes. The participants were categorized into subgroups according to the quartiles of the AIP. The diabetes incidence rates per 10,000 person-years were recorded as 22.29, 24.52, 34.95, and 80.52 for the respective AIP quartiles. Specifically, the diabetes incidence rates for each quartile were: Q1: 0.70%, Q2: 0.76%, Q3: 1.08%, and Q4: 2.48%. Those participants with the highest AIP (Q4) exhibited an increased risk of diabetes onset in comparison to those with the lowest AIP (Q1) (trend P< 0.001). As depicted in Supplementary Figure S2 , the analysis revealed a significant increase in diabetes prevalence corresponding to the ascending AIP quartiles (P < 0.001). Figure 2 presents the Kaplan-Meier curves, which depict the likelihood of diabetes development based on AIP levels. The transition probabilities were found to differ markedly according to AIP (P < 0.001), exhibiting a steady rise in likelihood as the AIP values increased.

As shown in Supplementary Table S3 , the univariate analysis showed that the risk of diabetes was positively associated with DBP, age, BMI, AST, SBP, TG, FPG, ALT, and TC, all with a significance level of P<0.05. In addition, AIP was positively associated with risk of diabetes. Conversely, it was negatively associated with HDL-C (all P<0.05).

Three distinct models were created utilizing Cox proportional hazards regression to examine the association between the AIP and the risk of developing diabetes. The first model, which was unadjusted, indicated that for each 1-unit increment in the AIP, there was a 5.93 rise in the probability of advancing to a diabetic condition, reflected by a HR of 6.93(95% CI 5.88-8.16, P<0.001). In the second model, which accounted for age and gender only, a 1-unit increase in the AIP corresponded to a 3.96 increase in the probability of diabetes onset, with an HR of 4.96 (95% CI 4.07-6.04, P<0.001). The third model, which was fully adjusted, illustrated that a 1-unit increment in the AIP was associated with a 1.07 rise in diabetes likelihood, yielding an HR of 2.07 (95% CI 1.47-2.48, P<0.001). The distribution of the confidence intervals underscores the strength of the association between AIP levels and diabetes risk (refer to Table 3 ). Furthermore, we converted the AIP from a continuous variable into a categorical one and reintroduced the grouped AIP into the analysis. The results from the adjusted multivariate model showed that compared to individuals in quartile 1 (Q1), the HR for those in quartiles 2 through 4 (Q2-Q4) were 0.94, 0.97, and 1.55, respectively (as shown in Table 3 , Model II).

To validate the robustness of our conclusions, we conducted a variety of sensitivity analyses. Initially, we utilized Model III of the GAM, which incorporated extra smoothing terms for various variables and resulted in a HR of 1.91 (1.47-2.48, Px<0.001) (refer to Table 3 , Model III). Subsequently, we excluded individuals with systolic blood pressure (SBP) ≥140 mmHg. Adjusting for confounding variables indicated a persistent positive association between the AIP and diabetes risk (HR = 2.15, 95% CI: 1.62-2.84, P<0.001). In another sensitivity analysis, we removed participants with diastolic blood pressure (DBP) ≥90 mmHg. Even after adjusting for confounding factors, results continued to demonstrate a strong positive association between the AIP and diabetes risk (HR=2.02, 95% CI: 1.56-2.62, P < 0.001). An additional analysis focused exclusively on participants under 60 years of age yielded an HR of 2.05 (95% CI: 1.47-2.84, P<0.001). Our thorough sensitivity analyses reinforce the reliability of our findings (see Table 4 ). Through an in-depth analysis of the original data, we reached the same conclusion ( Supplementary Table S4 ). All sensitivity analyses conducted indicated that our findings are robust. Additionally, we calculated the E-value to assess the potential impact of unmeasured confounding factors on the study results. The results showed that unknown or unmeasured variables appear to have limited influence on the relationship between AIP and diabetes risk, as the calculated E-value was 3.60, significantly higher than the relative risk value of 2.51 associated with unmeasured confounders related to AIP.

In our research, we identified a connection between the AIP and diabetes risk, as illustrated in Figure 3 and Table 5 . We initially implemented a Cox proportional hazards regression model, utilizing cubic splines to examine the association of AIP with diabetes risk. The results indicated a non-linear association between AIP and the likelihood of developing diabetes. To further explore this association, we employed a two-piecewise Cox proportional hazards regression model. In this analysis, participants with an AIP < -0.02 demonstrated a significantly heightened risk of developing diabetes (HR 2.75, 95% CI 1.97-3.85; P <0.001), while no meaningful association was observed for those with an AIP ≥ -0.02 (HR 1.12, 95% CI 0.62-2.02; P = 0.709). The p-value from the log-likelihood ratio test, which evaluated the overall fit of the two-piecewise model, was found to be 0.016, suggesting a non-linear association between AIP and diabetes risk.

In all predefined or exploratory subgroup analyses ( Figure 4 ), gender, BMI, age, smoking status, drinking, and family history of diabetes did not alter the relationship between AIP and diabetes risk. In other words, there were no significant interactions between these factors and AIP (interaction P > 0.05).