This study used a retrospective cohort study design, and the data were obtained from a retrospective cohort study previously undertaken by Chinese researchers (Chen et al.) from a computerized database in China (23). The target-independent variable was BMI at baseline. The outcome variable was reversion to normoglycemia from prediabetes at follow-up.

The raw data were obtained free of charge from DATADRYAD (www.datadryad.org) and provided by Ying Chen et al. With reference to the terms of service of the Dryad database, the dataset can be used by researchers to be able to share, remix, modify and create derivative works for non-commercial purposes, as long as the author and source are credited. Data information was obtained from an open access article published in 2018 – “Association of body mass index and age with diabetes onset in Chinese adults: a population-based cohort study” (http://dx.doi.org/10.1136/bmjopen-2018-021768). The data can be downloaded at: https://doi.org/10.5061/dryad.ft8750v (23).

The initial researchers took information from a computerized database created by the Rich Healthcare Group in China. This database contains all medical records of participants who underwent a health check from 2010 to 2016, spanning 32 regions and 11 cities in China. The Rich Healthcare Group Review Board initially approved the original study, and the information was retrieved retrospectively. For the retrospective study, no informed consent or approval was required by the institutional ethics committee (23). Therefore, the current secondary analysis did not require ethical approval. Additionally, the initial study was conducted in accordance with the Helsinki Declaration (23). So did this secondary analysis.

The original study enrolled 685,277 participants older than 20 who passed at least two health examinations. 473,444 participants meeting the exclusion criteria were excluded. The following were the original study’s exclusion criteria: The original study’s exclusion criteria were as follows: (i) participants with a visit interval of less than two years; (ii) participants with extreme BMI values (15 kg/m² or > 55 kg/m²); (iii) participants with no information about weight, height, sex, and fasting plasma glucose(FPG) value at baseline; (iv) participants with diabetes at enrollment; and (v) participants whose diabetes status at follow-up was unknown. Finally, the analysis of the initial study comprised a total of 211,833 people (23). In the current study, we first further included 26,018 participants with baseline FPG of 5.6-6. 9 mmol/L. We then excluded participants with missing FPG information at follow-up (n = 12) as well as those with abnormal and extreme BMI (three standard deviations greater or less than three standard deviations from the mean) (n = 132). Finally, the current study included 25,874 people in total. The procedure for choosing participants is shown in Figure 1 . It is important to highlight that according to the American Diabetes Association 2022 criteria, prediabetes is defined as the presence of IFG (FPG level of 5.6–6.9 mmol) and/or IGT and/or hemoglobin A1c(HA1c) (24). Our definition of prediabetes is therefore based on FPG. To make the study more accurate, our study population is reported as patients with IFG.

We divided the individuals into three categories based on the World Health Organization’s BMI values for Chinese patients: “normal”, “overweight”, and “obesity”. The means and standard deviations were presented for continuous variables with Gaussian distributions, medians were reported for skewed distributions, and frequencies and percentages were presented for categorical variables. We used the Kruskal-Wallis H test (skewed distribution), the One-Way ANOVA test (normal distribution), or χ2 (categorical variables) to test for differences among different BMI groups.

Following collinearity screening, we used univariate and multivariate Cox proportional-hazards regression models to examine the relationship between BMI and the reversion rate to normoglycemia in individuals with IFG, including a crude model with no covariates adjusted, a model with just minimal covariates adjusted (Model I with adjusted sex and age), and a model with full covariate adjustments (Model II: adjusted DBP, age, sex, SBP, AST, BUN, ALT, LDL-c, Scr, TG, HDL-c, family history of diabetes, drinking status, and smoking status). Effect sizes (HR) with 95% confidence intervals (CI) were recorded. We adjusted for confounding factors based on clinical experience, literature reports, and the results of univariate analysis. Additionally, the final multivariate Cox proportional hazards regression equation did not include TC since it was collinear with other variables ( Supplementary Table S1 ).

Besides, the Cox proportional hazards regression model with cubic spline functions and smooth curve fitting were performed to account for the nonlinear relationship between BMI and reversion to normoglycemia in participants with IFG. Furthermore, a two-piecewise Cox proportional hazards regression model was used to clarify the nonlinear association between BMI and reversion from IFG to normoglycemia. Finally, a log-likelihood ratio test was performed to choose the best model to explain the association between them in patients with IFG. Considering that patients who experience diabetes at follow-up are no longer likely to recover from IFG to normoglycemia, this may hinder the observation of prediabetes reversal to normoglycemia events or alter the likelihood of events occurring (32, 33). Therefore, competing risks multivariate Cox proportional-hazards regression was performed, as described by Fine and Gray, with progression to diabetes as the competing risk for the reversal to normoglycemia events (33, 34).

Using a stratified Cox proportional hazard regression model, subgroup analyses were performed across various groupings (age, sex, SBP, DBP, smoking status, and drinking status). First, based on clinical cut-off points, continuous data, such as SBP and age, were transformed into categorical variables (age: 30, 30 to 40, 40 to 50, 50 to 60, 60 to 70, 70 years old; SBP: 140, 140 mmHg) (35). In addition to the stratification factor itself, we adjusted each stratification for DBP, age, SBP, sex, Scr, ALT, HDL-c, AST, BUN, TG, LDL-c, drinking status, family history of diabetes, and smoking status. Ultimately, in models with and without interaction terms, the likelihood ratio test was employed to identify whether there were interaction terms or not.

To check the reliability of the findings, we ran a series of sensitivity analyses. Previous studies have suggested that drinking status, TG, and family history of diabetes are significantly related to glucose metabolism (36–38). We also conducted further sensitivity analyses to examine the connection between BMI and reversion to normoglycemia in prediabetic patients. Firstly, we performed a sensitivity analysis on participants who had never consumed alcohol (n=21,010). We also performed a sensitivity analysis after excluding patients with a family history of diabetes (n=25,244). In addition, we further explored the relationship between BMI and reversion to normoglycemia in participants with TG<1.7mmol/L (N= 15,858). The continuity covariate was also incorporated into the equation as a curve using a generalized additive model (GAM) to confirm the reliability of the findings. We also calculated E-values to examine the possibility of unmeasured confounding between BMI and reversion from IFG to normoglycemia (39).

All results were written in accordance with the STROBE statement (40). Empower Stats (X&Y Solutions, Inc., Boston, MA, http://www.empowerstats.com) and R statistical software packages (http://www.r-project.org, The R Foundation) were used for all analyses. Statistical significance was set at P values lower than 0.05(two-sided). Supplementary Figure S1 showed the analytical framework for this study.

The study participants’ demographic and clinical characteristics are presented in Table 1 . The mean age was 49.07 ± 13.82 years old, and 17,168 (66.35%) were male. The median follow-up time was 3.05 years, and 11,856 (45.82%) participants had a final reversion to normoglycemia. BMI presents a normal distribution, ranging from 15.2 to 34.9kg/m², with a mean of 24.74kg/m² ( Figure 2 ). We assigned adults into subgroups based on Chinese criteria for BMI categories (normal: < 24, overweight: 24-28, obesity: ≥28kg/m²). Compared with the normal group, age, height, weight, DBP, SBP, TG, LDL-c, TC, AST, ALT, Scr, and BUN increased significantly in the obesity group, whereas the opposite results were found in the HDL-c covariates. In addition, the proportion of men, current smokers, and current drinkers was higher in the obesity group.

Baseline characteristics according to regression and progression status of patients with IFG are shown in Supplementary Table S2 . Participants who progressed to diabetes had significantly higher levels of age, height, weight, BMI, DBP, SBP, TG, LDL-c, TC, AST, ALT, Scr, and BUN than participants with persistent IFG but significantly lower levels of HDL-c. Besides, compared with participants with persistent IFG, age, height, weight, BMI, DBP, SBP, TG, LDL-c, TC, AST, ALT, Scr, and BUN decreased significantly in participants who reverted to normoglycemia, whereas the opposite results were found in the HDL-c covariates.

In participants with IFG, 11,856 individuals developed diabetes. The overall rate of reversion to normoglycemia was 155.33 per 1000 person-years. In particular, the reversal rate to normoglycemia among participants with IFG of BMI groups was normal group:186.84, overweight group:139.14, and obesity group:119.25 per 1000 person-years, respectively. The overall cumulative reversal rate of IFG to normoglycemia was 45.82% over a median follow-up period of 3.05 years. The cumulative reversal rate in each BMI group was normal group:54.55%, overweight group: 41.38%, and obesity group: 35.19% ( Figure 3 ). Participants with higher BMI had a significantly lower reversal rate than those with a lower BMI (p<0.001 for trend) ( Table 2 , Figure 3 ).

In the age stratification by ten intervals, the rate of reversion to normoglycemia among participants with IFG was higher in women than in men, regardless of their age group ( Figure 4 ). It was also found that the reversal rate decreased with age in both men and women.

Univariate analyses showed that reversion to normoglycemia in patients with IFG was negatively correlated with age, BMI, DBP, SBP, ALT, AST, TG, TC, LDL-c, BUN, and family history of diabetes but was positively related to HDL-c, never smoking and never drinking (all P<0.05; Table 3 ).

Figure 5 showed the Kaplan-Meier curves for the probability of reversion to normoglycemia from IFG stratified by BMI category. The probability of reversal to normoglycemia from IFG varied significantly between BMI groups (log-rank test, p<0.001). The probability of reversion to normoglycemia decreased progressively with rising BMI, meaning that patients with the highest BMI had the lowest probability of reverting from IFG to normoglycemia. Supplementary Figure S2 presented Kaplan-Meier survival curves for diabetes-free survival probability. Among BMI groups, there were statistically significant differences in the probability of diabetes-free survival (log-rank test, p<0.001). IFG patients with the greatest BMI had the highest risk of progression to diabetes.

We constructed three models using the Cox proportional-hazards regression model to investigate the association between BMI and reversion to normoglycemia in patients with IFG. In the crude model, a 1kg/m² increase in BMI was associated with a 6.5% decrease in the probability of reversion to normoglycemia (HR=0.935,95%CI 0.930-0.940, p<0.001). In the minimally-adjusted model, when we adjusted for population variables only, each 1kg/m² increase in BMI was associated with a 4.6% lower probability of reversion to normoglycemia (HR=0.954, 95%CI 0.949-0.960, p<0.001). The HR between BMI and reversion to normoglycemia from IFG was 0.977 (95% CI: 0.971-0.984, p<0.001) in the fully adjusted model. The distribution of confidence intervals suggested that the link between BMI and reversion to normoglycemia among patients with IFG obtained by the model was reliable ( Table 4 ).

Besides, we transformed BMI from a continuous variable to a categorical variable and then reintroduced the categorically transformed BMI into the model. The results of the multivariate-adjusted model showed that with reference to participants with normal BMI, the HR was 0.901(95%CI:0.863-0.939) for overweight participants and 0.831(95%CI:0.780-0.886) for obese participants. That is, compared with participants with normal BMI(<24kg/m²), overweight participants (BMI:24-28kg/m²) with prediabetes had a 9.9% lower probability of returning to normoglycemia, while obese patients (BMI ≥ 28kg/m²) had a 16.9% decreased probability of reverting from IFG to normoglycemia ( Table 4 Model II).

When progression to incident diabetes from IFG was treated as a competing event, the competing analysis results were shown in Table 5 . In the crude model, BMI was negatively related to the probability of reversion to normoglycemia (SHR=0.93, 95% CI:0.93-0.94). In the minimally adjusted model (model I: adjusted age, sex), the result did not have a noticeable change (SHR:0.95, 95%CI: 0.95-0.96, p<0.001). In the fully adjusted model (model II) (adjusted age, sex, SBP, DBP ALT, AST, BUN, Scr, TG, LDL-c, HDL-c, family history of diabetes, drinking status, and smoking status), we could also detect a negative association between BMI and reversion to normoglycemia (SHR=0.92, 95%CI: 0.89-0.96). In addition, when BMI was used as a categorical variable, multivariate-adjusted model (fully adjusted model) results showed that overweight participants with IFG had a 10% lower probability of returning to normoglycemia compared with participants with normal BMI (SHR=0.90, 95%CI: 0.86-0.94), while obese patients had a 17.0% decreased probability of reverting from IFG to normoglycemia compared with patients with normal BMI (SHR=0.83, 95%CI: 0.78-0.89).

A series of sensitivity analyses were performed to ensure that our findings were robust. We first introduced the continuity covariate as a curve into the equation using a GAM. As shown in Table 4 , the outcome of Model III was consistent with the fully adjusted model. Referring to patients with IFG with normal BMI, obese patients had a 14.5% lower probability of reverting to normoglycemia (HR = 0.855, 95% CI: 0.802-0.912).

Furthermore, we conducted a sensitivity analysis on participants who had never consumed alcohol (n = 21,010). After adjusting for confounding variables (including DBP, age, SBP, sex, TG, ALT, AST, LDL-c, BUN, Scr, HDL-c, family history of diabetes, and smoking status), the findings indicated that BMI was also negatively linked with reversion to normoglycemia from IFG (HR=0.979, 95%CI:0.972-0.986, p<0.001). We also excluded patients with a family history of diabetes for the sensitivity analyses (n=25,244). After adjusting for confounding variables (including BUN, sex, AST, SBP, age, DBP, HDL-c, ALT, Scr, LDL-c, TG, drinking status, and smoking status), the results suggested that BMI was still negatively associated with reversion to normoglycemia from IFG (HR=0.977, 95% CI:0.971-0.984, p<0.001). Besides restricting the analysis to participants with TG<1.7mmol/L (adjusted for age, sex, SBP, DBP, ALT, AST, BUN, Scr, LDL-c, HDL-c, family history of diabetes, drinking status, and smoking status), the results suggested that the HR between BMI and probability of reverting to normoglycemia was 0.981 (95% CI:0.973-0.989, P<0.001). Similarly, when BMI was used as a categorical variable, sensitivity analyses of multivariate-adjusted models showed a significantly lower probability of recovery from IFG to normoglycemia in overweight and obese patients compared with participants with normal BMI ( Table 6 ). Based on all the sensitivity analyses, it is evident that our findings were robust. We also calculated an E-value to evaluate the sensitivity to unmeasured confounding. Unknown or unmeasured variables likely had little impact on the association between BMI and recovery from IFG to normoglycemia, as the E-value (1.53) was greater than the relative risk of BMI and unmeasured confounders (1.36).

Using a Cox proportional hazards regression model with cubic spline functions, we found that the correlation between BMI and the probability of reversal to normoglycemia in patients with IFG was nonlinear ( Figure 6 ). Additionally, using a standard binary two-piecewise Cox proportional-hazards regression model to fit the data, we chose the model that best fit the data using the log-likelihood ratio test ( Table 7 ). Less than 0.05 was the P-value for the log-likelihood ratio test. By using the recursive technique, we first established the 21.7 kg/m² as the BMI inflection point. Next, we utilized a two-piecewise Cox proportional hazards regression model to get the HR and CI for either side of the inflection point. Before the inflection point, the HR was 1.000 (95% CI: 0.978, 1.022, P=0.979), which was not statistically significant, and after the inflection point, the HR was 0.972 (95% CI: 0.964-0.980).

In all prespecified or exploratory subgroups assessed ( Table 8 ), sex, age, smoking status, SBP, and alcohol consumption did not modify the relationship between BMI and reversion to normoglycemia from IFG. That is, there was no significant interaction between these factors and BMI (P > 0.05 for interaction).