Dose–response relationship between the postload–fasting gap and the risk of developing diabetes: a cohort study from multiple centers in China

Abstract

Introductive

Early impairments in post-challenge glucose regulation are not fully captured by fasting measures alone. The postload-fasting gap, defined as the difference between 2-hour postload plasma glucose (2hPG) and fasting plasma glucose (FPG), may reflect dynamic dysregulation, yet its relation with glycaemic deterioration and remission in Chinese populations remains unclear. To characterise the dose-response relation between the postload-fasting gap and four glycaemic outcomes: incident diabetes, incident prediabetes, progression from prediabetes to diabetes, and reversion to normal glucose tolerance in a large multicentre Chinese cohort.

Research design and methods

We analyzed 3094 adults free of diabetes at baseline with two revisits over a mean follow-up of 3.24 years. Outcomes were ascertained at each visit by oral glucose tolerance test (OGTT) using World Health Organization (WHO) 1999 criteria, with sensitivity analyses using American Diabetes Association (ADA) definitions that include HbA1c. Primary associations were estimated on person-period data using discrete-time hazard models with a complementary log-log link, modeling the postload–fasting gap with restricted cubic splines after adjusting for demographic, clinical, and lifestyle covariates; cluster robust SEs accounted for repeated observations. Spline knots (K=3, 4, or 5) were placed at recommended percentiles and selected by Akaike information criterion, treating delta Akaike information criterion less than or equal to 2 as equivalent and favoring the more parsimonious model. Multiplicity was controlled using the false discovery rate. Internal validation used cluster bootstrap resampling. We further assessed prediction with six nested models (A–F), reporting area under the curve (AUC) with bootstrap CIs, net reclassification improvement and integrated discrimination improvement, and evaluated clinical utility by decision curve analysis.

Results

Higher postload–fasting gaps were associated with more adverse metabolic profiles at baseline and with higher risks of incident diabetes, incident pre-diabetes, and progression; lower postload–fasting gaps were associated with reversion to normal glucose tolerance. Dose–response curves showed that for incident diabetes, risk was flat close to a postload–fasting gap of 0 and increased beyond 2 mmol/L; for incident pre-diabetes, risk increased in a generally monotonic fashion; for progression, the increase was steeper; for reversion, risk decreased as postload–fasting gap increased. Findings were robust to alternative covariate sets, knot choices, and diagnostic definitions. In prediction analyses, the model that combined FPG with the postload—fasting gap (model F) provided the greatest incremental value across outcomes. For incident diabetes, the optimism-corrected AUC was 0.686, continuous net reclassification improvement was up to 0.349, and integrated discrimination improvement was 0.005; decision curve analysis indicated a higher net benefit for model F across clinically relevant thresholds.

Conclusions

The postload–fasting gap is an independent and non-linear marker of glycemic risk and remission potential. Incorporating this measure, particularly together with FPG, improves risk stratification and clinical utility, supporting its use as a practical OGTT-derived metric for early identification of people at risk of developing diabetes and targeted prevention.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• The postload–fasting gap, defined as the difference between 2-hour postload plasma glucose and fasting plasma glucose, reflects the dynamic balance between insulin secretion and peripheral glucose utilization. Previous studies have suggested its potential to indicate early dysglycemia, but most used simple categorical groupings or cross-sectional designs, leaving the shape and independence of the dose–response relationship with diabetes risk unclear, especially in large, longitudinal Chinese cohorts.

WHAT THIS STUDY ADDS

• Using a multicenter Chinese cohort with repeated oral glucose tolerance tests, this study characterizes the non-linear dose–response association between the postload–fasting gap and multiple glycemic outcomes, including incident diabetes, pre-diabetes, progression, and reversal. Postload–fasting gaps showed heterogeneous metabolic profiles across the range, while clinically meaningful risk increases were mainly observed at higher positive values. Adding the postload–fasting gap significantly improved risk discrimination, reclassification, and clinical utility compared with conventional glycemic markers alone.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• The postload–fasting gap can be readily calculated from a standard oral glucose tolerance test without additional testing. Incorporating this metric into clinical assessment could enhance early identification of individuals at risk for diabetes, support targeted prevention, and inform future research on dynamic glucose regulation and personalized risk stratification.

Introduction

Type 2 diabetes (T2D) has emerged as a growing global health concern, with an estimated 589 million adults affected worldwide in 2024 and projections exceeding 853 million by 2050.¹ In China, the burden is particularly severe, with more than 140 million people living with diabetes, placing enormous pressure on healthcare systems and the economy.^{2 3} This is accompanied by a rising prevalence of pre-diabetes, a condition that frequently progresses to diabetes within a few years.^{4 5} Despite widespread use of traditional glycemic markers such as fasting plasma glucose (FPG), 2-hour postload plasma glucose (2hPG), and glycated hemoglobin (HbA1c), these indicators may not fully reflect the complexity of glucose metabolism and early-stage dysfunction.^{6 7} There is an urgent need for more sensitive and integrative biomarkers to improve early-risk stratification and enable timely interventions.

Recent research has highlighted the clinical significance of the postload–fasting gap (the value of 2hPG minus FPG) as a potential indicator of glucose metabolism abnormalities even in people with normal glucose tolerance (NGT).^{8 9} This measure integrates the characteristics of both FPG and 2hPG, offering comprehensive insights into the dynamic interplay between insulin secretion, peripheral glucose utilization, and hepatic glucose metabolism regulation.10,13 Prior studies have shown associations between the postload–fasting gap and risks of diabetes, pre-diabetes, and reversal from pre-diabetes to NGT.^{9 14 15} However, most of these studies relied on simplistic classifications, for example, by dividing the postload–fasting gap into binary categories (positive vs negative) or basic percentiles, which may obscure more complex and non-linear relationships. In addition, little evidence is available from large-scale longitudinal studies in Chinese populations, despite the urgent need for context-specific risk markers. From a clinical and public health perspective, a simple indicator that can be derived from an oral glucose tolerance test (OGTT) may provide an inexpensive and practical tool for identifying high-risk individuals and guiding prevention strategies.

In this study, we aimed to investigate the dose–response relationships between the postload–fasting gap and key glycemic outcomes, including incident diabetes, pre-diabetes, progression from pre-diabetes to diabetes, and reversal to NGT. Specifically, we hypothesized that larger postload–fasting gap values would be associated with higher risks of incident diabetes, incident pre-diabetes, and progression from pre-diabetes to diabetes, whereas smaller postload–fasting gap values were hypothesised to be associated with a higher probability of reversion to NGT. This study was designed to evaluate the potential clinical utility of the postload–fasting gap as a biomarker for early-risk stratification and intervention.

Methods

Study population and data source

We analyzed data from the SENSIBLE and SENSIBLE-Addition studies, which were designed to determine the cut-off values of advanced glycation end-products and HbA1c for diagnosing diabetes in China.^{16 17} Participants were recruited across multiple centers with consecutive enrollment at each site using harmonized protocols. People aged 20–70 years who had been living in the area for 5 years were invited to participate and responded to a comprehensive health and lifestyle questionnaire and underwent a general health examination and laboratory tests. The protocol and study design details have been described previously.⁸ A total of 7600 participants from the SENSIBLE and SENSIBLE-Addition studies were invited to join a cohort and followed up two times, called the SENSIBLE-Cohort study.⁹ Inclusion criteria were ages 20–70 years, residence in the study area for at least 5 years, completion of the baseline questionnaire, examination and laboratory testing, including OGTT, and provision of written informed consent.

For this analysis, we excluded participants with missing data on glucose values (FPG, 2hPG and HbA1c), sociodemographic information (eg, age, gender, ethnicity, family history of diabetes), and outliers (greater than the 99.9th percentile or less than the 0.1st percentile) of anthropometric examination characteristics, missing data on dietary information, and participants with self-reported diabetes or who were diagnosed with diabetes during the baseline OGTT, resulting in 3094 adults for analysis.⁹ Online supplemental figure 1 shows the flowchart of this research.

Sample size calculations indicated that approximately 3091 participants were required after allowing for 10% loss to follow-up to provide 80% power (α=0.05) to detect an HR of 1.5 for incident diabetes, assuming an exposure prevalence of 22% and a cumulative incidence of 10% over 3 years (online supplemental methods). Our final analytic sample of 3094 exceeded this requirement, supporting adequate statistical power.

Outcome definitions and study variables

The primary outcome was incident diabetes, defined as a binary variable at each follow-up visit: ‘yes’ indicating a new diagnosis of diabetes and ‘no’ indicating no diabetes (NGT to diabetes). Secondary binary outcomes were incident pre-diabetes (NGT to pre-diabetes), disease progression (pre-diabetes to diabetes), and disease reversal (pre-diabetes to NGT). Although event times were recorded, the irregular scheduling of visits resulted in interval censoring, which reduced follow-up precision and limited the applicability of conventional time-to-event methods. Therefore, outcomes were analyzed as binary repeated measures at each visit using modified Poisson regression with a log link fitted via generalized estimating equations (GEE) and robust SEs clustered by participant. As a sensitivity analysis, we fitted a discrete-time hazard model with a complementary log–log link, including interval-specific fixed effects to model the baseline hazard and the log of interval length as an offset.

An OGTT was conducted in the baseline survey and at the two follow-up revisits. FPG and 2hPG were measured at each visit. The postload–fasting gap has been defined as the value of 2hPG minus FPG.^{8 9} Diagnoses of diabetes, impaired glucose tolerance (IGT), impaired fasting glucose (IFG), and NGT were defined following the WHO 1999 criteria.¹⁸ To reflect current clinical practice, we conducted sensitivity analyses that reclassified outcomes using ADA and WHO definitions incorporating HbA1c. We compared incidence under WHO 1999 criteria with incidence under HbA1c-inclusive definitions and repeated all primary models under the alternative definitions to assess robustness.

For analyses of incident diabetes or pre-diabetes, follow-up time began at the baseline visit and ended on the day of diagnosis of diabetes or pre-diabetes. For analyses related to disease progression or disease reversal, follow-up time began at the diagnosis of pre-diabetes until the diagnosis of diabetes or reversal to NGT. For participants who remained free of pre-diabetes/diabetes, the follow-up time was censored at their last available visit (online supplemental figure 2).⁹

Other study variables were participant demographics (ie, age, gender, ethnicity, education levels, occupations), general health examinations (ie, body mass index (BMI), systolic blood pressure (SBP), diastolic blood pressure (DBP) and heart rate), the results of laboratory tests (ie, HbA1c, total cholesterol (TC), high-density lipoprotein-cholesterol (HDL-C), low-density lipoprotein-cholesterol (LDL-C) and triglyceride (TG)), lifestyle (ie, smoking, drinking, vigorous exercise and regular diet) and histories of diseases and drugs (ie, family history of diabetes, antidiabetic agent use and self-report liver disease). All variables were measured at each visit, that is, three times in total.⁹

Missing data and selection bias

We assumed missing at random for covariates and repeated measures. As sensitivity analyses, we applied inverse probability weighting (IPW) for inclusion and follow-up, with weights estimated from baseline predictors of inclusion and visit attendance. The 45 participants without follow-up postload–fasting gap measures were retained in the analysis. All had baseline data; at the first follow-up, two were diagnosed from FPG alone without a 2hPG. At the second follow-up, all 45 were confirmed to have diabetes either from prior diagnosis or elevated FPG, but none underwent 2hPG. These participants were classified as incident diabetes cases at the time of diagnosis and censored thereafter.

Statistical analysis

Participants were categorized into quintiles (Q1–Q5) based on the postload–fasting gap. Categorical variables were expressed as percentages and compared using χ² tests, while continuous variables were tested for normality and presented as mean±SD or median (IQR) accordingly. Group comparisons used parametric or non-parametric tests where appropriate. Continuous variables were first assessed for normality using the Shapiro–Wilk test; normally distributed variables were compared using analysis of variance, and skewed variables using the Kruskal–Wallis rank-sum test.

To model interval risks over irregular follow-up, we constructed a person-period dataset in which each participant contributed one record per interval (baseline to first revisit; first to second revisit). We then fitted discrete-time hazard models as binomial regressions with a complementary log–log (cloglog) link, yielding HRs for interval-specific risks. SEs were cluster robust (sandwich) with clustering by participant to account for repeated observations. The primary exposure was the postload–fasting gap (mmol/L), modeled flexibly using restricted cubic splines (RCS) to capture potential non-linear dose–response relationships. Three models were fitted: the unadjusted model included only the postload–fasting gap; model 1 further adjusted for age, sex, BMI, and FPG; and model 2 additionally adjusted for ethnicity, recruiting center, HbA1c, blood pressure (SBP and DBP), resting heart rate, lipid profile (TC, HDL-C, LDL-C, TG), education, smoking, vigorous exercise, and regular diet. Because the postload–fasting gap (2hPG−FPG) inherently includes FPG, we incorporated FPG in both model 1 and model 2 to offset the influence of baseline fasting levels and to isolate the independent effect of the postchallenge increment. Under this parameterization, the coefficient for the gap can be interpreted as the effect of 2hPG conditional on FPG, while the FPG coefficient represents the independent contribution of FPG, thereby mitigating bias from mathematical coupling and regression to the mean.

For spline specification and model selection, we evaluated RCS with K=3, 4, or 5 knots placed at recommended percentiles (3-knot: 5th, 50th, 95th; 4-knot: 5th, 35th, 65th, 95th; 5-knot: 5th, 25th, 50th, 75th, 95th). The final K was chosen by minimizing Akaike information criterion (AIC); models within ΔAIC≤2 were considered equivalent, with ties favoring the more parsimonious specification. Stability of knot choice and curve shape was assessed via cluster bootstrap (eg, 500 resamples, resampling participants with replacement). Multiplicity arising from non-linearity tests and subgroup/interaction analyses was addressed using the false discovery rate (FDR), with FDR-adjusted q-values reported alongside p values in online supplemental materials. To limit overfitting, we prespecified an events-per-parameter threshold of ≥10, including spline terms. Internal validation used cluster bootstrap to estimate optimism in discrimination and to obtain CIs for performance metrics. Model calibration was evaluated using calibration plots and the Hosmer-Lemeshow test applied to predicted probabilities from the fitted cloglog models.

We quantified the added value of glycemia measures beyond conventional risk factors using six nested models: model A (baseline clinical model: age, sex, BMI, ethnicity, recruiting center, SBP, DBP, heart rate, lipid profile, education, smoking, vigorous exercise, regular meal pattern); model B=A+FPG; model C=A+2hPG; model D=A+HbA1c; model E=A+postload–fasting gap; model F=A+postload–fasting gap+FPG (a joint specification to condition on fasting levels). Because the postload–fasting gap equals 2hPG−FPG, model F deliberately includes FPG alongside the postload–fasting gap to offset baseline fasting influence, avoid misattributing fasting differences to the postload–fasting gap (mitigating mathematical coupling), and more accurately characterize the postchallenge increment’s independent effect. Discrimination was assessed using the C-statistic (area under the curve (AUC)) with bootstrap 95% CIs; calibration by calibration plots and calibration slope. We compared models using categorical and continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI). Decision curve analysis (DCA) quantified net benefit across clinically relevant threshold probabilities. Risk category definitions for categorical reclassification, the resampling scheme, and DCA threshold ranges are detailed in online supplemental methods.

For subgroup and interaction analyses, prespecified subgroups (age, sex, BMI) were analyzed by refitting the cloglog-RCS model within strata and by testing exposure-by-subgroup spline interactions (global Wald tests) in the full model with FDR control. For comparability and to avoid overfitting in smaller strata, we applied the same knot number and placement as in the main analysis rather than reselecting knots within subgroups. Two-dimensional density contour plots were used to visualize the joint distribution of age and BMI with each outcome.

To assess robustness to potential selection and follow-up bias, we refit all models using stabilized IPW for baseline inclusion and visit attendance. Weight specification, diagnostics (distributions, effective sample size, covariate balance), and IPW results are detailed in online supplemental materials.

All analyses were performed using R V.4.5.0 (R Foundation for Statistical Computing, Vienna, Austria) and SAS V.9.4 (SAS Institute). A two-sided p value <0.05 was considered statistically significant.

Ethical approval

Written informed consent was obtained from all participants at baseline and reconfirmed at each follow-up visit.¹⁶ All procedures complied with the Declaration of Helsinki.

Results

Population characteristics

Baseline characteristics from this cohort were previously reported.⁹ The analysis included 3094 adults (66% female; mean age 52 years; mean BMI 25 kg/m²). Participants were divided into quintiles according to baseline postload–fasting gap (≤−0.13 to >2.44 mmol/L). Across quintiles, higher postload-fasting gap values were associated with older age, higher BMI and blood pressure, and lower prevalence of current smoking (table 1, all p<0.05). However, SBP across quintiles showed a pattern that was not strictly monotonic but rather approximated a U-shaped distribution, and the proportion of current smokers also differed between quintiles without forming a clear linear trend. These findings indicate that the distributions varied among groups but did not follow a consistent increasing or decreasing pattern.

Table 1. Comparison of participant’s baseline information.

Characteristics	All n=3094 Mean±SD (%)	Q1 (≤−0.13) n=620 (20.04%) Mean±SD (%)	Q2 (>−0.13, ≤0.74) n=621 (20.07%) Mean±SD (%)	Q3 (>0.74, ≤1.45) n=617 (19.94%) Mean±SD (%)	Q4 (>1.45, ≤2.44) n=615 (19.88%) Mean±SD (%)	Q5 (>2.44) n=621 (20.07%) Mean±SD (%)	P value*
Age at baseline (years)	52.1 (9.2)	51.2 (9.5)	51.4 (9.4)	51.0 (9.7)	52.7 (8.4)	54.1 (8.4)	<0.001
Female (%)	2053 (66.4)	317 (51.1)	424 (68.3)	441 (71.5)	437 (71.1)	434 (69.9)	<0.001
Ethnicity							<0.001
Dai	125 (4.04%)	29 (4.68%)	17 (2.74%)	18 (2.92%)	34 (5.53%)	27 (4.35%)
Han	2507 (81.0%)	483 (77.9%)	514 (82.8%)	505 (81.9%)	501 (81.5%)	504 (81.2%)
Kazakh	134 (4.33%)	21 (3.39%)	23 (3.70%)	32 (5.19%)	28 (4.55%)	30 (4.83%)
Korean	51 (1.65%)	7 (1.13%)	10 (1.61%)	6 (0.97%)	14 (2.28%)	14 (2.25%)
Uyghur	110 (3.56%)	23 (3.71%)	26 (4.19%)	30 (4.86%)	16 (2.60%)	15 (2.42%)
Zhuang	162 (5.24%)	57 (9.19%)	31 (4.99%)	25 (4.05%)	20 (3.25%)	29 (4.67%)
Other	5 (0.16%)	0 (0.00%)	0 (0.00%)	1 (0.16%)	2 (0.33%)	2 (0.32%)
Recruiting center							<0.001
Eastern	2312 (74.73%)	460 (74.19%)	488 (78.58%)	460 (74.55%)	457 (74.31%)	447 (71.98%)
Southern	287 (9.28%)	86 (13.87%)	48 (7.73%)	43 (6.97%)	54 (8.78%)	56 (9.02%)
Western	246 (7.95%)	44 (7.10%)	50 (8.05%)	62 (10.05%)	45 (7.32%)	45 (7.25%)
Northern	166 (5.37%)	19 (3.06%)	24 (3.86%)	31 (5.02%)	39 (6.34%)	53 (8.53%)
Central	83 (2.68%)	11 (1.77%)	11 (1.77%)	21 (3.40%)	20 (3.25%)	20 (3.22%)
FPG (mmol/L)	5.6 (0.6)	5.6 (0.5)	5.6 (0.5)	5.5 (0.6)	5.6 (0.6)	5.7 (0.6)	<0.001
2hPG (mmol/L)	6.8 (1.7)	4.7 (0.8)	5.9 (0.6)	6.6 (0.6)	7.5 (0.6)	9.1 (1.0)	<0.001
Postload–fasting gap (mmol/L)	1.2 (1.6)	−0.9 (0.7)	0.3 (0.2)	1.1 (0.2)	1.9 (0.3)	3.4 (0.8)	<0.001
HbA1c (%)	5.4 (0.4)	5.3 (0.4)	5.3 (0.4)	5.3 (0.4)	5.4 (0.4)	5.5 (0.5)	<0.001
Initial glycemic status (WHO 1999)							<0.001
NGT	2019 (65.3%)	505 (81.5%)	521 (83.9%)	523 (84.8%)	424 (68.9%)	46 (7.4%)
IFG	303 (9.8%)	115 (18.6%)	100 (16.1%)	81 (13.1%)	7 (1.14%)	0 (0.0%)
IGT	772 (25.0%)	0 (0.0%)	0 (0.0%)	13 (2.1%)	184 (29.9%)	575 (92.6%)
BMI (kg/m2)	25.3 (3.5)	24.9 (3.2)	24.6 (3.2)	25.2 (3.6)	25.6 (3.5)	26.4 (3.7)	<0.001
SBP (mm Hg)	134 (19.3)	134 (19.1)	131 (18.5)	130 (18.8)	135 (19.2)	138 (19.8)	<0.001
DBP (mm Hg)	82 (11.4)	82 (11.6)	81 (11.2)	80 (11.3)	83 (11.0)	84 (11.5)	<0.001
Heart rate (beats per minute)	77 (11.0)	75 (10.7)	77 (11.1)	77 (10.5)	78 (11.1)	79 (11.3)	<0.001
TC (mmol/L)	5.0 (1.0)	4.9 (1.0)	4.8 (1.0)	4.9 (1.0)	5.0 (1.0)	5.1 (1.1)	<0.001
HDL-C (mmol/L)	1.5 (0.4)	1.6 (0.4)	1.6 (0.4)	1.6 (0.4)	1.5 (0.4)	1.5 (0.4)	<0.001
LDL-C (mmol/L)	2.7 (0.8)	2.7 (0.7)	2.6 (0.7)	2.7 (0.7)	2.8 (0.8)	2.9 (0.8)	<0.001
TG (mmol/L)	1.7 (1.8)	1.4 (1.3)	1.6 (1.8)	1.6 (1.7)	1.8 (1.6)	2.2 (2.4)	<0.001
Education levels (uneducated)	583 (18.8%)	87 (14.0%)	108 (17.4%)	114 (18.5%)	135 (22.0%)	139 (22.4%)	<0.001
Occupation type							0.453
Professional	238 (7.7%)	46 (7.4%)	57 (9.2%)	51 (8.3%)	47 (7.6%)	37 (6.0%)
Manual worker	2846 (92.0%)	570 (91.9%)	562 (90.5%)	565 (91.6%)	566 (92.0%)	583 (93.9%)
Student	10 (0.3%)	4 (0.7%)	2 (0.3%)	1 (0.2%)	2 (0.3%)	1 (0.2%)
Smoking							<0.001
Never	2466 (79.7%)	454 (73.2%)	508 (81.8%)	509 (82.5%)	500 (81.3%)	495 (79.7%)
Former	104 (3.4%)	28 (4.5%)	14 (2.3%)	19 (3.1%)	15 (2.4%)	28 (4.5%)
Current	524 (16.9%)	138 (22.3%)	99 (15.9%)	89 (14.4%)	100 (16.3%)	98 (15.8%)
Drinking							0.164
Never	2347 (75.9%)	441 (71.1%)	473 (76.2%)	479 (77.6%)	479 (77.9%)	475 (76.5%)
Former	109 (3.5%)	28 (4.5%)	25 (4.0%)	20 (3.2%)	17 (2.8%)	19 (3.1%)
Current	638 (20.6%)	151 (24.4%)	123 (19.8%)	118 (19.1%)	119 (19.4%)	127 (20.5%)
Vigorous exercise (yes)	983 (31.8%)	186 (30.0%)	173 (27.9%)	203 (32.9%)	223 (36.3%)	198 (31.9%)	0.023
Regular diet (yes)	2300 (74.3%)	429 (69.2%)	445 (71.7%)	464 (75.2%)	488 (79.4%)	474 (76.3%)	<0.001
Diabetes family history (yes)	509 (16.5%)	93 (15.0%)	102 (16.4%)	100 (16.2%)	106 (17.2%)	108 (17.4%)	0.800
Antidiabetic agent use (yes)	7 (0.2%)	2 (0.3%)	2 (0.3%)	1 (0.2%)	1 (0.2%)	1 (0.2%)	0.935
Self-report liver disease (yes)	109 (3.5%)	16 (2.6%)	24 (3.9%)	25 (4.1%)	18 (2.9%)	26 (4.2%)	0.435

Q1–Q5 are sorted according to the quintile of the value of postload–fasting gap (mmol/L); Q1 is the lowest quintile and Q5 is the highest quintile.

^*All p values are two sided.

BMI, body mass index; DBP, diastolic blood pressure; FPG, fasting plasma glucose; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; 2hPG, 2-hour postload plasma glucose; IFG, impaired fasting glucose; IGT, impaired glucose tolerance; LDL-C, low-density lipoprotein cholesterol; NGT, normal glucose tolerance; SBP, systolic blood pressure; TC, total cholesterol; TG, triglyceride.

Metabolic indices varied significantly across quintiles: FPG increased modestly (5.6±0.5 to 5.7±0.6 mmol/L), 2hPG rose markedly (4.7±0.8 to 9.1±1.0 mmol/L), and HbA1c increased from 5.3±0.4% to 5.5±0.5%. By definition, the postload–fasting gap rose progressively from −0.9±0.7 mmol/L in Q1 to 3.4±0.8 mmol/L in Q5 (all p<0.001). These changes were reflected in glycemic status: the prevalence of NGT declined from 81.5% in Q1 to 7.4% in Q5, while IGT increased from 0% to 92.6%. The proportion with isolated IFG peaked in Q2–Q3 but was absent in Q5 (table 1).

Longitudinal changes in the postload–fasting gap

Over the 3.24-year follow-up, postload–fasting gap values increased progressively at both revisits, with the largest rises observed in upper quintiles. The overall distribution shifted upward, indicating a general widening of the postload-fasting gap over time (figure 1a). Participants in the extreme quintiles (Q1 and Q5) tended to remain relatively stable across visits, while those in intermediate quintiles showed more movement between categories. Many in lower quintiles moved upward to higher categories, reflecting a progressive deterioration in post-challenge glucose regulation (figure 1b).

Figure 1. Quintile distribution and longitudinal changes of the postload–fasting glucose gap. (a) Mean±SD values of the postload–fasting gap across quintiles at baseline, first revisit, and second revisit. (b) Longitudinal shifts in quintile cut-off values of the postload–fasting gap over time. Error bars represent SDs. Q1–Q5 were defined according to the quintile distribution of the postload–fasting gap (mmol/L) at each visit. Baseline: Q1≤−0.13; Q2>−0.13 to ≤0.74; Q3>0.74 to ≤1.45; Q4>1.45 to ≤2.44; Q5>2.44 mmol/L. First revisit: Q1≤0.19; Q2>0.19 to ≤0.95; Q3>0.95 to ≤1.70; Q4>1.70 to ≤2.78; Q5>2.78 mmol/L. Second revisit: Q1≤0.40; Q2>0.40 to ≤1.25; Q3>1.25 to ≤2.05; Q4>2.05 to ≤3.26; Q5>3.26 mmol/L.

Postload–fasting gap and different outcomes over the cohort

Table 2 shows the effects of different postload–fasting gap groups on different outcomes over time, including incident diabetes, incident pre-diabetes, disease progression, and disease reversal. All subgroup and interaction analyses were prespecified. We formally tested first-order interactions between the postload–fasting gap and age, sex and BMI within the GEE framework.

Table 2. GEE estimates for repeated measures of postload–fasting gap on incident diabetes, incident pre-diabetes, disease progression and disease reversal.

Label	Mean difference (95% CI)*†	P value*†	Mean difference (95% CI)*†	P value*†	Mean difference (95% CI)*§	P value*§
Incident diabetes
Q1 versus Q2	0.004 (−0.000 to 0.007)	0.079	0.003 (−0.001 to 0.007)	0.194	−0.000 (−0.002 to 0.001)	0.721
Q1 versus Q3	0.004 (0.000 to 0.008)	0.034	0.003 (−0.001 to 0.007)	0.120	0.001 (−0.002 to 0.003)	0.632
Q1 versus Q4	0.005 (0.001 to 0.008)	0.014	0.004 (−0.000 to 0.008)	0.060	0.001 (−0.005 to 0.007)	0.683
Q1 versus Q5	−0.076 (−0.094 to −0.058)	<0.001	−0.077 (−0.095 to −0.059)	<0.001	0.002 (0.000 to 0.004)	0.023
Q2 versus Q3	0.001 (−0.002 to 0.003)	0.657	0.001 (−0.002 to 0.003)	0.671	0.001 (−0.001 to 0.003)	0.404
Q2 versus Q4	0.001 (−0.001 to 0.003)	0.379	0.001 (−0.001 to 0.003)	0.370	0.002 (−0.004 to 0.007)	0.598
Q2 versus Q5	−0.079 (−0.097 to −0.061)	<0.001	−0.079 (−0.097 to −0.062)	<0.001	0.003 (0.000 to 0.005)	0.023
Q3 versus Q4	0.001 (−0.002 to 0.003)	0.656	0.001 (−0.002 to 0.003)	0.643	0.001 (−0.005 to 0.006)	0.781
Q3 versus Q5	−0.080 (−0.098 to −0.062)	<0.001	−0.080 (−0.098 to −0.062)	<0.001	0.002 (−0.001 to 0.005)	0.165
Q4 versus Q5	−0.080 (−0.098 to −0.062)	<0.001	−0.080 (−0.098 to −0.063)	<0.001	0.001 (−0.005 to 0.007)	0.725
Linear trend		<0.001		<0.001		0.081
Incident pre-diabetes
Q1 versus Q2	0.005 (−0.004 to 0.014)	0.260	0.007 (−0.003 to 0.016)	0.163	−0.002 (−0.007 to 0.002)	0.337
Q1 versus Q3	0.001 (−0.009 to 0.011)	0.836	0.003 (−0.008 to 0.013)	0.636	−0.001 (−0.006 to 0.004)	0.723
Q1 versus Q4	−0.183 (−0.208 to −0.159)	<0.001	−0.182 (−0.207 to −0.157)	<0.001	−0.020 (−0.033 to −0.007)	0.003
Q1 versus Q5	−0.541 (−0.571 to −0.511)	<0.001	−0.546 (−0.577 to −0.515)	<0.001	0.004 (−0.001 to 0.009)	0.122
Q2 versus Q3	−0.004 (−0.013 to 0.005)	0.368	−0.004 (−0.013 to 0.005)	0.366	0.001 (−0.004 to 0.006)	0.638
Q2 versus Q4	−0.188 (−0.213 to −0.164)	<0.001	−0.189 (−0.213 to −0.164)	<0.001	−0.018 (−0.031 to −0.006)	0.005
Q2 versus Q5	−0.546 (−0.576 to −0.516)	<0.001	−0.553 (−0.583 to −0.522)	<0.001	0.006 (0.001 to 0.012)	0.028
Q3 versus Q4	−0.184 (−0.209 to −0.160)	<0.001	−0.184 (−0.209 to −0.160)	<0.001	−0.019 (−0.031 to −0.007)	0.002
Q3 versus Q5	−0.542 (−0.572 to −0.512)	<0.001	−0.549 (−0.579 to −0.518)	<0.001	0.005 (−0.001 to 0.011)	0.124
Q4 versus Q5	−0.358 (−0.396 to −0.320)	<0.001	−0.364 (−0.402 to −0.326)	<0.001	0.024 (0.011 to 0.038)	<0.001
Linear trend		<0.001		<0.001		0.223
Disease progression
Q1 versus Q2	0.002 (−0.009 to 0.012)	0.739	−0.005 (−0.016 to 0.007)	0.428	−0.003 (−0.017 to 0.010)	0.623
Q1 versus Q3	0.003 (−0.006 to 0.013)	0.491	−0.008 (−0.018 to 0.003)	0.178	−0.007 (−0.021 to 0.006)	0.280
Q1 versus Q4	0.003 (−0.006 to 0.012)	0.535	−0.016 (−0.027 to −0.005)	0.003	−0.011 (−0.024 to 0.002)	0.096
Q1 versus Q5	−0.136 (−0.155 to −0.118)	<0.001	−0.167 (−0.188 to −0.146)	<0.001	−0.170 (−0.192 to −0.147)	<0.001
Q2 versus Q3	0.002 (−0.007 to 0.010)	0.740	−0.003 (−0.013 to 0.007)	0.577	−0.004 (−0.015 to 0.007)	0.500
Q2 versus Q4	0.001 (−0.007 to 0.010)	0.806	−0.012 (−0.021 to −0.002)	0.018	−0.008 (−0.019 to 0.004)	0.190
Q2 versus Q5	−0.138 (−0.157 to −0.120)	<0.001	−0.162 (−0.182 to −0.142)	<0.001	−0.166 (−0.188 to −0.145)	<0.001
Q3 versus Q4	−0.000 (−0.007 to 0.007)	0.902	−0.009 (−0.017 to −0.001)	0.039	−0.004 (−0.014 to 0.006)	0.442
Q3 versus Q5	−0.140 (−0.158 to −0.122)	<0.001	−0.159 (−0.179 to −0.140)	<0.001	−0.162 (−0.183 to −0.142)	<0.001
Q4 versus Q5	−0.139 (−0.157 to −0.121)	<0.001	−0.150 (−0.169 to −0.132)	<0.001	−0.158 (−0.178 to −0.139)	<0.001
Linear trend		<0.001		<0.001		<0.001
Disease reversal
Q1 versus Q2	−0.106 (−0.159 to −0.053)	<0.001	−0.092 (−0.144 to −0.041)	<0.001	−0.088 (−0.135 to −0.041)	<0.001
Q1 versus Q3	−0.116 (−0.167 to −0.064)	<0.001	−0.096 (−0.145 to −0.047)	<0.001	0.217 (0.166 to 0.268)	<0.001
Q1 versus Q4	0.198 (0.145 to 0.251)	<0.001	0.227 (0.177 to 0.277)	<0.001	0.505 (0.464 to 0.546)	<0.001
Q1 versus Q5	0.445 (0.401 to 0.489)	<0.001	0.484 (0.442 to 0.525)	<0.001	−0.011 (−0.049 to 0.027)	0.559
Q2 versus Q3	−0.010 (−0.051 to 0.032)	0.656	−0.004 (−0.044 to 0.037)	0.860	0.294 (0.253 to 0.336)	<0.001
Q2 versus Q4	0.304 (0.262 to 0.347)	<0.001	0.319 (0.278 to 0.361)	<0.001	0.582 (0.552 to 0.612)	<0.001
Q2 versus Q5	0.551 (0.520 to 0.582)	<0.001	0.576 (0.546 to 0.606)	<0.001	0.305 (0.267 to 0.344)	<0.001
Q3 versus Q4	0.314 (0.274 to 0.353)	<0.001	0.323 (0.285 to 0.361)	<0.001	0.593 (0.568 to 0.618)	<0.001
Q3 versus Q5	0.561 (0.533 to 0.588)	<0.001	0.580 (0.554 to 0.606)	<0.001	0.288 (0.257 to 0.319)	<0.001
Q4 versus Q5	0.247 (0.218 to 0.276)	<0.001	0.257 (0.228 to 0.286)	<0.001	−0.088 (−0.135 to −0.041)	<0.001
Linear trend		<0.001		<0.001		<0.001

^*Estimates from GEE model with identity link and normal distribution.

^†Unadjusted model.

^‡Adjusted for age, gender, BMI and FPG.

^§Fully adjusted model including age, gender, BMI, FPG, ethnicity, recruiting center, HbA1c, blood pressure (SBP and DBP), heart rate, blood lipids (TC, HDL-C, LDL-C, TG), education, smoking, vigorous exercise and regular diet.

BMI, body mass index; DBP, diastolic blood pressure; FPG, fasting plasma glucose; GEE, generalized estimating equation; HDL-C, high-density lipoprotein-cholesterol; LDL-C, low-density lipoprotein-cholesterol; SBP, systolic blood pressure; TC, total cholesterol; TG, triglyceride.

Across outcomes, higher postload–fasting gap quintiles were consistently associated with greater risks of adverse glycemic outcomes. For incident diabetes and incident pre-diabetes, participants in Q1–Q4 had significantly lower risks than those in Q5 (all p<0.001). For disease progression, Q4 and Q5 showed higher risks than Q1–Q3 (all p<0.001), with Q5 demonstrating the strongest effect. In contrast, the probability of disease reversal decreased progressively across quintiles, with Q4–Q5 showing the most pronounced benefit (all p<0.001). All associations remained robust in fully adjusted models (p<0.05).

Progression from NGT to diabetes was essentially identical under ADA and WHO 1999 definitions (3.0% vs 3.1%, p=1.000). In contrast, intermediate transitions differed: more participants moved from NGT to pre-diabetes under ADA (29.7% vs 22.3%, p<0.001), fewer progressed from pre-diabetes to diabetes (12.5% vs 16.5%, p<0.001), and more regressed from pre-diabetes to NGT (63.6% vs 58.2%, p=0.001) (online supplemental figure 3). Despite these differences, GEE estimates for the postload–fasting gap were directionally consistent and of similar magnitude across definitions, indicating that the effect estimates were materially unchanged (online supplemental table S1).

The resulting goodness of fit for the unadjusted model, model 1 and model 2

Model performance was compared using Quasi-likelihood under Independence Model Criterion (QIC) and Quasi-likelihood under the Independence model Criterion (uncorrected) (QICu) criteria (online supplemental figure 4). For most outcomes, full adjustment improved model fit relative to simpler specifications, except for incident pre-diabetes where differences were minimal. Based on QICu, events-per-parameter ratio, and bootstrap stability (online supplemental tables S2-S3), the final recommended specifications were: model 1 with K=5 for incident diabetes, model 1 with K=4 for incident pre-diabetes, the unadjusted model with K=5 for disease progression, and model 1 with K=4 for disease reversal. These choices achieved strong overall and non-linear effects after FDR adjustment (online supplemental table S4). Optimism-corrected AUCs typically exceeded 0.93, indicating excellent discrimination and minimal overfitting.

Model performance and incremental value of the postload–fasting gap

Across all outcomes, model discrimination ranged from modest to good, with optimism-corrected C-statistics between 0.61 and 0.81 (online supplemental table S5). The greatest incremental predictive value was consistently observed for model F, which jointly incorporated FPG and the postload–fasting gap (online supplemental table S6). For incident diabetes, model F achieved the highest discrimination (C=0.69) and the largest reclassification improvement (NRI=0.35; IDI=0.01), followed by HbA1c (model D; NRI=0.24; IDI=0.01). Similar patterns were seen for incident pre-diabetes (C=0.70; NRI=0.46; IDI=0.04), disease progression (C=0.79; NRI=0.84; IDI=0.14), and disease reversal (C=0.71; NRI=0.54; IDI=0.09). In contrast, FPG or 2hPG alone (models B and C) offered only minor incremental gains (NRI≤0.30; IDI≤0.03).

Calibration analyses showed acceptable agreement overall, with slopes close to 1 and intercepts near 0, while model F demonstrated the best fit by Hosmer-Lemeshow tests. DCA further confirmed the superior net clinical benefit of model F across relevant threshold probabilities, followed by HbA1c and the postload–fasting gap alone (online supplemental figures 5 and 6).

Non-linear association between postload–fasting gap and incident diabetes, incident pre-diabetes, disease progression and disease reversal

Figure 2 shows the dose–response associations between the postload–fasting gap and the risk of glycemic outcomes. For incident diabetes, the risk remained relatively stable when postload–fasting gap was close to 0 but increased sharply once it exceeded approximately 2 mmol/L, with wider CIs at the extremes. A similar monotonic pattern was observed for incident pre-diabetes, where the risk rose gradually near 0 and then accelerated markedly at higher postload–fasting gap values. Disease progression from pre-diabetes to diabetes displayed an even steeper trajectory, indicating that individuals with pre-diabetes were particularly susceptible to progression as postload–fasting gap increased. In contrast, the likelihood of disease reversal declined steadily with larger postload–fasting gaps, suggesting that participants with lower or negative gaps were more likely to return to normoglycemia, whereas those with higher postload–fasting gaps had a substantially reduced probability of reversal.

Figure 2. Adjusted dose–response between the postload–fasting gap and outcome risk from discrete-time hazard models with a complementary log–log link and restricted cubic splines. Each panel represents a specific outcome: (a) incident diabetes, (b) incident pre-diabetes, (c) disease progression (pre-diabetes to diabetes), and (d) disease reversal (pre-diabetes to normoglycemia). Models for a, b, and d were adjusted for age, sex, body mass index (BMI), and fasting plasma glucose (FPG), whereas the disease progression model (c) was unadjusted. HRs were estimated relative to the median value of the postload–fasting gap. Shaded areas represent 95% CIs.

Subgroup analysis

Prespecified subgroup analyses were performed by age, sex, and BMI to assess the consistency of associations across key demographic and metabolic strata (online supplemental figures 7–11). Across BMI strata, larger postload–fasting gaps were generally associated with a higher hazard of adverse glycemic outcomes, with the steepest risk gradients in BMI≥28 and more modest increases in 24≤BMI<28. Estimates in BMI<24 were imprecise with very wide confidence bands and a sharp peak, indicating sparse/unstable data at distribution tails; these should be interpreted cautiously. Consistent with the BMI patterns, older age groups tended to exhibit stronger associations than younger groups. Sex-specific differences were also evident: men showed steeper risk increases for diabetes incidence and progression, whereas women displayed more gradual slopes. Overall, the subgroup RCS curves support the robustness of the main findings while highlighting meaningful heterogeneity by BMI, age, and sex.

Sensitivity analyses: IPW

To assess robustness to potential selection and follow-up bias, we repeated the analyses using stabilized IPW. IPW analyses yielded results that were broadly consistent with the primary findings. Weight models for baseline inclusion and the first follow-up showed good discrimination/calibration, stable weight distributions, effective sample sizes close to nominal, and substantial improvement in covariate balance. By contrast, the second follow-up weight model yielded highly variable weights, a marked reduction in effective sample size, and only partial balance, indicating residual bias and warranting cautious interpretation of estimates involving visit 2 (online supplemental tables S7–S9; online supplemental figures Sx and Sy).

Discussion

Our study demonstrates a non-linear dose–response relationship between the postload–fasting gap and glycemic outcomes, including incident diabetes, pre-diabetes, disease progression, and disease reversal. These associations persisted after adjusting for major confounders. For incident diabetes, risk was approximately flat close to a postload–fasting gap of 0 and increased at higher positive values of the postload–fasting gap. For incident pre-diabetes, risk increased in a generally monotonic fashion across the exposure range. For progression from pre-diabetes to diabetes, the increase was steeper. For disease reversal, higher values of the postload–fasting gap were associated with lower probability of reversion to NGT. Confidence intervals widened at both extremes of the distribution due to sparse data, and in the negative range the hazard estimates remained essentially flat, indicating limited information rather than a meaningful change in risk.

The spline-derived dose–response curve suggested that the risk of diabetes remained low and relatively flat when the postload–fasting gap was modest, around 0 to 2 mmol/L, while larger positive postload-fasting gaps were associated with progressively higher risk. Although no formal threshold was defined, this range may represent a relatively low-risk zone for glycaemic deterioration. Conversely, smaller or even negative postload–fasting gaps did not show evidence of improved glycaemic recovery, as hazard estimates in this range remained essentially flat with no clear indication of benefit.

The increasing burden of T2D represents a significant challenge in global healthcare systems.¹⁹ Thus, the description and understanding of early changes in glucose metabolism prior to the incidence of T2D are very important. T2D often develops gradually, with measurable changes in plasma glucose and insulin sensitivity occurring years before diagnosis. The Whitehall II study, for example, reported early elevations in FPG and 2hPG 3–6 years prior to diagnosis, with an earlier and more rapid rise in 2hPG.²⁰ Specifically, FPG began to increase linearly 3 years before diagnosis, while 2hPG began to increase rapidly 3 years before diagnosis. At the same time, insulin sensitivity decreased significantly within 5 years before diagnosis, and pancreatic β-cell function improved 4 to 3 years before diagnosis, but then continued to decline until diagnosis.²⁰ Another study supports this view, noting that HbA1c levels gradually increase in the 3 years before diabetes diagnosis, followed by a sharp increase in the year before diagnosis.²¹ These findings highlight that plasma glucose levels and related metabolic parameters alter prior to the formal diagnosis of diabetes, although within the normal range. In our cohort, similar trends were observed over 3.24 years of follow-up. We tracked changes in the postload–fasting gap longitudinally and found that individuals with larger postload–fasting gaps at baseline experienced more adverse metabolic profiles over time, including rising 2hPG and HbA1c levels. More importantly, recently published data have demonstrated that an elevated 1hPG level (≥8.6 mmol/L) was associated with increased risk of developing T2D and its long-term complications including cardiovascular disease and mortality.²²

Our study is in agreement with several previous studies and found an abrupt increase in FPG from 1 to 3 years before the diagnosis of diabetes and pre-diabetes,²³ similar to our findings on 2hPG trajectories. These have indicated that 2hPG levels can rise significantly before the diagnosis of diabetes and pre-diabetes, and the abnormality of 2hPG levels may occur earlier than changes in FPG.^{24 25} In our previous study, we observed the trajectories of FPG and 2hPG independently and defined the difference between the 2hPG value and FPG value (2hPG minus FPG) as the postload–fasting gap and followed the relationship between postload–fasting gap and diabetes/pre-diabetes incidence.^{8 9} In the trend analyses, participants with larger postload–fasting gaps tended to be older, predominantly female, with higher BMI, blood pressure, and lipid levels. Significant increases in 2hPG and HbA1c were observed across quintiles, with IGT becoming more prevalent in higher quintiles. They suggest a potential association between metabolic risk factors and the postload–fasting gap. Furthermore, over time, the overall postload–fasting gap was on the rise, particularly in higher quintiles (Q4 and Q5).

Different from the independent FPG value and 2hPG value, the postload–fasting gap can reflect both the insulin secretion function and insulin sensitivity, but also reveal the dynamic regulation state of glucose metabolism in the body.²⁰ Specifically, even within the normal range of plasma glucose, larger positive postload–fasting gap values are often associated with insulin resistance and β-cell dysfunction, which leads to exaggerated postload glucose levels, serving as significant risk factors for pre-diabetes and diabetes.²⁶ Conversely, a large negative postload–fasting gap may signal altered hepatic glucose output or abnormal insulin kinetics, which could also indicate an underlying risk of diabetes.⁹ But in our data, negative postload-fasting gap was not associated with increased glycaemic risk.

Although individuals with negative postload–fasting gaps tended to be younger and metabolically healthier, the estimated hazard ratios remained flat in this range, and there was no clear indication of elevated risk. The wider confidence intervals at the negative end reflected sparse data rather than increased hazard. This paradox may reflect early-phase dysregulation of hepatic glucose suppression^{27 28} or excessive first-phase insulin secretion followed by late-phase insufficiency^{29 30}, leading to transient post-challenge hypoglycaemia but unstable long-term glucose homeostasis. In addition, recurrent reactive hypoglycaemia can induce counter-regulatory hormonal activation, promoting hepatic glucose output and insulin resistance over time31,33. These physiological explanations describe potential origins of negative gaps, but in our cohort these mechanisms did not translate into higher observed risk, and the association remained flat across the negative range.

Several studies have shown that the postload–fasting gap may predict diabetes, with higher positive values associated with greater risk when the postload–fasting gap was classified into broad groups in prior work.^{9 14 15} Our prediction analyses extend this by showing that adding the postload–fasting gap to conventional predictors improved discrimination and reclassification and yielded higher net benefit on decision curves, while maintaining acceptable internal calibration. We deliberately modeled FPG together with the postload–fasting gap to avoid attributing baseline fasting differences to the postload–fasting gap and to more accurately characterize the impact of the postchallenge increment on outcomes. We did not pursue a single diagnostic cut-off because the exposure–risk relation is non-linear and because our focus was on risk stratification rather than dichotomous screening. Together, these findings indicate that the postload–fasting gap adds information beyond established predictors.

The postload–fasting gap can be obtained from a standard 75 g OGTT without additional testing, since it uses FPG and 2hPG that are already measured. In primary prevention, individuals with larger postload–fasting gaps had higher risks of incident diabetes and progression from pre-diabetes, whereas smaller postload–fasting gaps were associated with reversion to NGT. Accordingly, the postload–fasting gap provides a simple and interpretable indicator to flag higher or lower risk after an oral glucose challenge within routine care. In practice, postload–fasting gap can be computed automatically within electronic records and displayed alongside FPG, 2hPG and HbA1c. Compared with validated risk scores based on clinical variables, incorporating the postload–fasting gap improved model performance in our nested model framework, with gains in the C-statistic, positive net reclassification and higher net benefit across clinically relevant thresholds, while preserving reasonable calibration. This suggests that postload–fasting gap should complement rather than replace existing tools, and that it can refine risk stratification after an oral glucose challenge.

These findings suggest that the postload–fasting gap adds value beyond standard predictors and validated risk scores. The improvement in discrimination and reclassification, together with acceptable calibration, supports its potential to refine risk stratification following an oral glucose challenge. Our study highlights the significant role of age, sex, and BMI in modifying the association between the postload–fasting gap and glycemic outcomes. These factors are well-established determinants of glucose metabolism and diabetes risk. For instance, older individuals tend to have reduced insulin sensitivity and greater β-cell dysfunction, which may amplify the effect of glucose excursions.³⁴ Sex is another crucial factor, as gender differences in insulin sensitivity and glucose tolerance have been widely observed, with women often showing different metabolic responses compared with men.³⁵ Similarly, a higher BMI is strongly associated with insulin resistance and chronic inflammation, which may intensify the metabolic consequences of a widened postload–fasting gap.³⁶

This study has several strengths. We used repeated OGTT to ascertain outcomes, analyzed person-period data with a complementary log-log link to target interval risks under irregular visit spacing, and modeled the exposure flexibly with RCS to capture non-linear relations. We prespecified covariate sets, controlled multiplicity using the FDR, selected spline knots by information criteria with preference for parsimony, and used cluster bootstrap resampling for internal validation and stability checks. The consistency of the curve shapes and key contrasts across specifications supports the robustness of the conclusions.

There are also limitations. Implementation barriers include the limited routine use of OGTT in some settings, patient time burden and laboratory scheduling. A pragmatic approach is to prioritize the use of postload–fasting gap where OGTT is already undertaken for clinical reasons, or in people with intermediate FPG or HbA1c who need clarification of postchallenge dysglycemia. Although OGTT provides detailed information, some participants declined postchallenge testing at revisits, which may introduce selection effects. We therefore performed inverse probability censoring weighting as a sensitivity analysis, and the qualitative conclusions were unchanged. We conducted the primary analyses using WHO 1999 diagnostic definitions and repeated key analyses using ADA definitions that include HbA1c. As with any observational cohort, residual confounding cannot be excluded. Our reliance on glucose-based measures also means that we did not directly assess insulin sensitivity or secretion, which could clarify mechanisms underlying extreme values of postload–fasting gap. Also, we did not conduct subgroup analyses by ethnicity. In our cohort, the numbers of participants and outcome events within non-majority ethnic strata were small, which would yield unstable estimates, particularly for flexible RCS terms, reduce statistical power after multiplicity correction, and increase the risk of type I or type II errors or model non-convergence. To avoid overinterpretation, we therefore prespecified not to stratify by ethnicity. We acknowledge that glycemic responses may differ by ethnicity, and adequately powered multicenter or pooled analyses will be needed to assess potential effect modification by ethnicity. Finally, generalisability beyond Chinese populations requires confirmation in other settings. Our findings were derived from a Chinese cohort. Although non-Chinese studies such as Coronary Artery Risk Development in Young Adults (CARDIA) study have shown similar relevance of the postload–fasting difference to glycemic progression,¹⁵ external validation in other ethnicities is warranted.

Future work should seek external validation in other populations and care settings, examine feasibility and impact when postload–fasting gap is integrated into real-world pathways, and test whether modifying determinants of postload–fasting gap through lifestyle or pharmacological interventions reduces progression risk or promotes reversion. Mechanistic studies that combine postload–fasting gap with direct measures of insulin sensitivity and secretion may help disentangle organ-specific contributions across the exposure range.

In conclusion, the postload–fasting gap is a practical and informative measure derived from the OGTT. It captures clinically relevant non-linear exposure–response relations and adds meaningful value to risk prediction and clinical decision-making when combined with FPG. These properties support its use to refine risk stratification and to guide targeted prevention.

Data availability statement

Data are available upon reasonable request.