Data from the CHARLS were used. The details of the program’s objectives, design and methods can be found elsewhere (17). The CHARLS is a prestigious, ongoing, and nationally representative longitudinal survey. The project employed a multi-stage sampling approach, utilizing the probability proportional to size sampling method at both the county/district and village/community sampling stages. Its purpose is to meticulously collect a robust set of high-quality microdata that reflects the lives of households and individuals in China who are 45 years of age and above. The inaugural national baseline survey was launched in 2011, reaching into 450 villages, spanning 150 counties within 28 provinces, and engaging over 17,000 individuals. This benchmark study is meticulously repeated every 2 to 3 years to provide a snapshot of the demographic trends. The data is meticulously collected through face-to-face interviews. CHARLS collected data on body measurements and blood samples. The datasets could be acquired at http://charls.pku.edu.cn/en once the online application was approved and five national waves data are available including the baseline survey (2011y, wave 1), the first followed data (2013y, wave2), second follow‐up survey (2015y, wave 3), third follow‐up survey (2018y, wave 4), and fourth follow‐up survey (2020y, wave 5). The CHARLS project was approved by the Biomedical Ethics Committee of Peking University, and all participants were required to sign informed consents.

The data used in this study was from wave 1 and wave 3. Figure 1 presented the procedures of the study design, sampling and exclusion. Of 17,708 participants at baseline, a total of 2,655 participants with prediabetes were included in the present study.

Continuous variables in this study were presented as means ± standard deviations and compared by t-test. Categorical variables were presented as frequency (n, %) and compared by chi-square test. Multiple logistic regression analysis was conducted to explore the association between factors which showed significant difference between two groups with prediabetes regression to NGR. Stepwise regression based on the Akaike information criterion (AIC) minimum was used to select optimal feature factors as the previous research (14).

To elucidate the dose-response relationship between fasting plasma glucose (FPG) and glycated hemoglobin (HbA1c) levels in relation to the regression to NGR, we utilized Restricted Cubic Spline (RCS) analyses. These analyses were instrumental in identifying FPG and HbA1c as critical determinants of prediabetes status.

As part of our sensitivity analysis, we employed a multinomial logistic regression model. This model accounted for three distinct glycemic outcome groups observed during the follow-up period: regression to NGR, progression to diabetes, and remained as prediabetes. For the purposes of this analysis, the group that remained prediabetic at the follow-up was used as the reference category.

Attribution fraction (AF) was used to evaluate the contribution of the modifiable factors for regression to NGR. AF was the embodiment of the percentage reduction of a given outcome that was expected if there was no exposure (22). AFs and 95% CI were calculated with the ‘AF’ R package (22). In addition, individual AFs cannot be summed directly to obtain an overall AF as given the assumption that the independence of risk factors is likely biased (23) and the individual AFs may be over-evaluated.

We conducted a further analysis to assess the correlation between changes in modifiable risk factors and the regression to NGR. This evaluation involved calculating the changes in these variables by determining the difference between their follow-up values and their initial baseline measurements.

All statistical analyses were performed using SPSS version 24.0 (IBM Corp., Armonk, NY) and R (version 4.2.2, http://www.r-project.org). A two-sided p value < 0.05 was considered statistically significant.

The baseline characteristics of the participants were detailed in Table 1 . Among 2,655 participants identified with prediabetes at baseline, 570 (21.5%) reverted to NGR over the subsequent 4-year follow-up period. In comparison with those who did not revert to NGR, individuals who achieved NGR were younger, predominantly male, and exhibited lower levels of TC, LDL-C, fasting plasma glucose, HbA1c, and WC, as well as a reduced prevalence of general and abdominal obesity (all p-values < 0.05). When participants were categorized into three distinct groups (regression to NGR, progression to diabetes, and remained as prediabetes), similar patterns in baseline characteristics were observed across all groups ( Supplementary Table S1 ).

The associations between baseline characteristics and regression to NGR were presented in Table 2 . In the stepwise logistic regression model, variables including age, gender, abdominal obesity, LDL-C, FPG, and HbA1c were identified as significant predictors of the regression from prediabetes to NGR. The results indicated that older age (OR, 0.98; 95%CI, 0.97 to 0.99), female sex (OR, 0.79; 95%CI, 0.65 to 0.97), abdominal obesity (OR, 0.70; 95%CI, 0.57 to 0.86), higher LDL-C (OR, 0.69; 95%CI, 0.48 to 0.97), higher FPG (OR, 0.68; 95%CI, 0.52 to 0.90), and higher HbA1c (OR, 0.23; 95%CI, 0.18 to 0.30) were associated with a reduced likelihood of reverting to NGR.

The sensitivity analyses based on multinomial logistic regression model did not influence the primary analysis outcomes ( Table 3 ).

The dose-response relationship evaluated by RCS models showed nonlinear relationships between FPG and HbA1C and regression to NGR ( Figure 2 ).

The AFs of modifiable factors associated with regression to NGR were presented in Table 4 . After adjusting for covariates, individuals with low HbA1c levels (<5.2%) were found to have the highest AF for regression to NGR, at 46.0% (95%CI, 38.5% to 53.5%). This was followed by those with low LDL-C levels (<4.14 mmol/L), with an AF of 22.7% (95%CI, 3.0% to 42.3%), absence of abdominal obesity with an AF of 14.60% (95%CI, 6.2% to 23.0%), and low FPG levels (<5.95 mmol/L) with an AF of 11.20% (95%CI, 4.1% to 18.2%).

The associations between changes of modifiable factors and regression to NGR were presented in Table 5 . After adjusting for covariates, elevated levels of LDL-C (OR, 0.71; 95%CI, 0.60 to 0.85), TC (OR, 0.72; 95%CI, 0.62 to 0.83), and TG (OR, 0.85; 95%CI, 0.75 to 0.96) were inversely associated with the likelihood of regression to NGR. In contrast, changes in BMI, WC, and HDL-C did not show a significant association with the regression to NGR.

In the present study, 570 (21.5%) participants with prediabetes at baseline regressed to NGR. The estimated rate of prediabetes regression to NGR in this study was consistent with our previous study using CHARLS (24), but diverged from other previous studies, which reported rates ranging from 9.3% to 27.3% (13, 14, 25). For example, the German study reported 9.3% of participants regressed from glucose-defined prediabetes to NGT, and 27.3% from HbA1c-defined prediabetes to normal HbA1c (13). Moreover, during 10 years of follow-up, 17.1% of participants in Japan (25) and 36% in Korea regressed from prediabetes to NGT (14). These studies utilized varying criteria for defining prediabetes, with disparities in the study populations and durations of follow-up periods. Therefore, a large-scale, multi-regional study adhering to standardized criteria is essential to accurately estimate the prevalence and modifiable factors influencing the regression from prediabetes to NGR, thereby enhancing clinical insights.

Aging was widely acknowledged to be associated with prediabetes and T2DM. Consistent with the majority of previous studies (9, 12, 26), older age was one of the most important factors that decreased odds of regression to NGR in our study. The aging process adversely affects β-cell function, exacerbating the deficiency in insulin production and heightening insulin resistance, potentially through mechanisms such as obesity, reduction in pancreatic parenchymal volume, and other contributing risk factors (27–30). These physiological changes collectively increased the risk of prediabetes and reduced the chances of regression to NGR (27–30).

The role of gender in prediabetes regression remained mixed. A DPPOS study demonstrated that female sex increased odds of regression from prediabetes to isolated IFG and decreased odd of regression to isolated IGT (9), another two follow-up studies reported no significant effect of gender (14, 31). The present study also showed that being female decreased the odds of regression to NGR, possibly due to the decline in estrogen levels in menopausal and postmenopausal women. Estrogen plays a pivotal role in augmenting insulin sensitivity and in promoting the secretion and utilization of insulin (32, 33). Previous studies indicated that following menopause, estrogen levels in women decrease with age, resulting in impaired glucose metabolism and an increased risk of developing diabetes (32, 33).

In addition to the unmodifiable factors of age and gender, our study identified several modifiable predictors that are significantly associated with the regression to NGR. Specifically, we found that blood glucose levels, obesity, and lipid profiles were independently linked to the likelihood of reverting from prediabetes to NGR. Therefore, from the perspective of public health prevention, individuals with prediabetes can promote the regression to NGR by controlling these modifiable factors, thereby preventing complications and improving the quality of life.

Previous studies demonstrated that baseline glucose status was a main predictor for regression to NGR (9, 14, 31, 34). Studies from Sweden (31), Korea (14), and America (9) demonstrated that both FPG and 2-h glucose were associated with regression to NGT. An Australian cohort study focusing on women reported that FPG levels were linked to a reduced probability of reverting to normoglycemia over a decade-long follow-up period (34). Our study corroborated these findings, indicating a significant association between FPG and regression to NGR. However, the relationship between HbA1c and regression to NGR has been less explored. Our study revealed a strong association between HbA1c levels and the regression to NGR. Notably, a lower initial HbA1c level was found to be the most influential factor in promoting regression to NGR. It’s worth noting that HbA1c can effectively reflect an individual’s blood glucose level over the past three months (35), independent of several factors including whether fasting before the test and the use of insulin or metformin (36), and smaller biological within‐subject variation for than FPG (37). As such, HbA1c may emerge as a pivotal prognostic indicator for the regression to NGR from prediabetes in clinical practice. It is important for individuals with prediabetes to regularly monitor their HbA1c levels, with the optimal target being below 5.2%.

Previous research had reported inconsistent findings on the link between baseline BMI and WC and the regression to NGR from prediabetes. For example, Kowall et al. found that baseline BMI and WC were associated with regression from HbA1c-based or glucose-based prediabetes to HbA1c-based or glucose-based normoglycemia, respectively (13), while Paprott et al. did not (26). Our study emphasized that abdominal obesity, more than general obesity, hindered regression to NGR. Abdominal obesity and glucose metabolism disorder was related to ectopic adipose tissue (EAT) more than subcutaneous adipose tissue (SAT) (38, 39). Even in individuals of normal weight, an excess of EAT can elevate the risk of insulin resistance (IR) and impair insulin secretion (38). The flow of non-esterified fatty acids produced by excess abdominal fat to the liver may impair hepatic metabolism and lead to increased hepatic glucose production (27, 39). Furthermore, an overabundance of EAT can trigger endocrine dysfunction, dysregulation of inflammatory factors, and a decrease in adiponectin levels, all of which contribute to IR (40).

Lipid levels have been implicated in the regression from prediabetes to NGR. A Korean 10-year follow-up study reported that high TG was against to regression to NGT (14). Similarly, an Australian 10-year cohort study identified high TG levels and their increase as having adverse effects on regression (17). Rebecca et al. reported that high HDL-C increased odds of regression to NGR (26). The present study reported that low LDL-C and reduced lipid levels contributed to regression to NGR. Lipid and glucose metabolism are intricately connected; excessive TG can induce inflammation and endoplasmic reticulum stress, impairing insulin activity, glucose uptake in muscles, and β-cell function (41, 42).

Some limitations in this study should be mentioned. Firstly, the definitions for glycemic outcomes were based on FPG and HbA1c, while the lack of data on oral glucose tolerance test due to the CHARLS design might cause diagnostic bias though a previous study demonstrated nearly 90% prediabetes could be identified by only FPG and HbA1c (43). Secondly, variables such as diet habits, drug use, and body fat, which might affect the glycemic outcomes, were not available in the CHARLS dataset, and their exclusion could potentially impact the final conclusions. Thirdly, FPG and HbA1c were from single measurements instead of repeated measurements in CHARLS sample measurements, which may bias the actual results. Fourthly, the study population were derived from CHARLS aged 45 and older, therefore the conclusions might not be generalizable to other populations.

This study also had strengths. First, the study population with a large size came from a nation-wide, representative and community-based population. Second, we tried to explore the factors associated with prediabetes regression to NGR among middle-aged and elder Chinese adults, and different models were used to verify the results.