Abstract
Background
Estimated glucose disposal rate (eGDR) is a surrogate measure of insulin sensitivity derived from routinely available clinical variables. However, whether repeated eGDR measurements provide additional information for cardiovascular risk assessment among individuals with metabolic dysfunction-associated steatotic liver disease (MASLD) remains unclear. We examined the associations of cumulative eGDR and longitudinal eGDR patterns with incident cardiovascular disease (CVD) and evaluated their incremental risk-discrimination value.
Methods
Using data from the China Health and Retirement Longitudinal Study, we conducted a prospective cohort analysis with 2015 specified as the landmark. Participants were aged ≥ 45 years in 2011, met an operational TyG-WC–based definition of MASLD in 2011, had calculable eGDR in both 2011 and 2015, were alive and free of CVD at the 2015 landmark, and had valid post-2015 outcome and follow-up data. Cumulative eGDR was calculated from the 2011 and 2015 measurements and, because all participants shared the same four-year interval, was mathematically proportional to the mean of the two measurements; K-means clustering was used exploratorily to identify longitudinal eGDR patterns, operationally characterised as two-wave eGDR level–change patterns based on the two observed measurements. Participants were followed from 2015 to 2020, and incident CVD was defined as the first report of physician-diagnosed heart disease or stroke. Cox proportional hazards models and restricted cubic splines were used to estimate associations and dose–response relationships. Competing-risk, subgroup, and sensitivity analyses, including alternative handling of CVD event times, were performed to assess robustness. Incremental predictive performance was evaluated using conventional and internally validated discrimination measures, including paired out-of-bag C-index comparisons.
Results
A total of 2,059 participants were included, of whom 457 developed incident CVD during follow-up. Each 1-SD lower cumulative eGDR was associated with a 39% higher risk of CVD after full adjustment (hazard ratio [HR] 1.39, 95% confidence interval [CI] 1.26–1.55). Compared with participants in the highest cumulative eGDR quartile (Q1), those in the lowest quartile (Q4) had an HR of 2.40 (95% CI 1.77–3.25; P for trend < 0.001). The dose–response relationship was approximately linear (P for non-linearity = 0.915). K-means clustering identified four exploratory longitudinal eGDR patterns based on the two observed measurements. Using the high eGDR–slight decline pattern as the reference, the moderate-low eGDR–slight rise, high eGDR–marked decline, and low eGDR–slight decline patterns were associated with HRs of 1.43 (95% CI 1.07–1.90), 1.64 (1.20–2.23), and 2.14 (1.67–2.74), respectively. Lower cumulative eGDR was also associated with higher risks of heart disease and stroke, and the main findings remained stable across multiple sensitivity analyses, including alternative handling of event times. Adding cumulative eGDR to the conventional risk-factor model increased the paired out-of-bag C-index by 0.033 (95% CI 0.009–0.056); however, when 2015 eGDR was already included, adding cumulative eGDR produced essentially no further improvement (ΔC-index − 0.001, 95% CI − 0.007 to 0.006).
Conclusions
Among middle-aged and older Chinese adults meeting an operational TyG-WC–based MASLD definition, lower cumulative eGDR and longitudinal eGDR patterns characterised by low levels at both observed measurements or a marked decline between measurements were associated with a higher risk of incident CVD. Cumulative eGDR provided modest incremental discrimination beyond conventional risk factors but essentially no additional discrimination beyond the 2015 eGDR measurement, indicating substantial overlap with the most recent eGDR value.
Graphical abstract
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s12933-026-03384-w.
Keywords: Metabolic dysfunction-associated steatotic liver disease, Estimated glucose disposal rate, Insulin resistance, Cardiovascular disease, Cumulative exposure, Longitudinal patterns
Research insights
What is currently known about this topic?
Insulin resistance is central to MASLD and cardiovascular risk.
Lower eGDR has been associated with adverse cardiovascular outcomes.
Evidence on repeated eGDR measurements in MASLD remains limited.
What is the key research question?
Are cumulative eGDR and two-wave eGDR level–change patterns associated with incident CVD among individuals meeting an operational TyG-WC–based MASLD definition?
Do repeated eGDR measurements provide incremental risk-discrimination information beyond conventional risk factors and a recent eGDR measurement?
What is new?
Lower cumulative eGDR was associated with a higher risk of incident CVD in an approximately linear manner.
Two-wave eGDR patterns characterised by low levels at both observed waves or a marked decline were associated with higher CVD risk.
Cumulative eGDR modestly improved discrimination beyond conventional risk factors but provided essentially no additional discrimination once 2015 eGDR was already included.
How might this study influence clinical practice?
eGDR may serve as a simple composite cardiometabolic risk marker, while repeated measurements may help characterise the context of change; however, the clinical value and optimal interval of repeated assessment require further validation.
Introduction
Metabolic dysfunction-associated steatotic liver disease (MASLD), characterized by hepatic steatosis in the presence of at least one cardiometabolic risk factor, has become a major chronic metabolic liver disease worldwide [1]. Global Burden of Disease estimates suggest that approximately 1.3 billion people had MASLD in 2023, accounting for 16.1% of the global population [2]. In China, a recent systematic review and meta-analysis estimated the prevalence of MASLD among adults to be approximately 30.4% [3]. Meanwhile, cardiovascular disease (CVD) remains a leading cause of death and disability worldwide, affecting approximately 626 million people and causing about 19.2 million deaths in 2023 [4]. MASLD commonly coexists with cardiovascular risk factors such as central obesity, hypertension, dysglycaemia, and dyslipidaemia, and its clinical consequences extend beyond liver-related outcomes [5, 6]. Previous longitudinal studies have associated steatotic liver disease with increased risks of both fatal and nonfatal cardiovascular events [7]. Identifying individuals at high risk of subsequent CVD within the large MASLD population is therefore important for long-term disease management and cardiovascular prevention.
Insulin resistance is central to the development and progression of MASLD and represents an important link between hepatic fat accumulation, systemic metabolic dysfunction, and cardiovascular injury [8]. Although the hyperinsulinaemic–euglycaemic clamp is a reference method for assessing insulin sensitivity, its complexity limits its application in large population-based studies [9]. The homeostatic model assessment of insulin resistance requires measurement of fasting insulin and is likewise impractical for routine screening [10]. Estimated glucose disposal rate (eGDR), calculated from waist circumference, hypertension status, and glycated haemoglobin, is a surrogate measure of insulin sensitivity that can be calculated without direct insulin measurement, with lower values generally indicating poorer estimated insulin sensitivity [11, 12]. Previous prospective studies and meta-analyses have associated lower eGDR with adverse cardiovascular outcomes in the general population, individuals without diabetes, and patients with diabetes, suggesting that it may provide a simple measure for cardiometabolic risk assessment [13–15].
Studies using repeated eGDR measurements from the UK Biobank, the China Health and Retirement Longitudinal Study, and the US Health and Retirement Study have reported that lower cumulative eGDR or adverse eGDR trajectories are associated with a higher risk of subsequent CVD [16–18]. In contrast, previous studies in MASLD populations have focused mainly on baseline triglyceride–glucose indices and their adiposity-related derivatives [19]. Whether repeated eGDR measures provide distinct or complementary risk information in MASLD remains insufficiently examined. Within a common exposure-assessment window, cumulative eGDR, the most recent eGDR, simple change, and longitudinal eGDR patterns may summarise different aspects of the observed eGDR profile, including average level, recent status, magnitude of change, and the combination of initial level, final level, and direction of change. Because waist circumference, hypertension, and HbA1c are components of eGDR and are also established cardiometabolic risk factors, these measures should be interpreted as related representations of a composite cardiometabolic marker rather than as independent dimensions of insulin resistance.
Using data from the China Health and Retirement Longitudinal Study, we therefore investigated middle-aged and older Chinese adults meeting an operational TyG-WC–based MASLD definition, treating cumulative eGDR during 2011–2015 as the primary exposure and using K-means clustering exploratorily to identify longitudinal eGDR patterns, operationally characterised as two-wave eGDR level–change patterns based on the two observed measurements [20]. We evaluated the associations of both measures with incident CVD after 2015. The year 2015 was specified as the landmark, ensuring that both the 2011 and 2015 eGDR measurements preceded the start of outcome follow-up. We further compared the risk information provided by 2011 eGDR, landmark eGDR, cumulative eGDR, simple change, and longitudinal eGDR patterns, and examined dose–response relationships. We hypothesized that lower cumulative eGDR and longitudinal eGDR patterns characterised by low eGDR at both observed measurements or a marked decline between measurements would be associated with a higher risk of incident CVD.
Methods
Study design and population
This study used data from the China Health and Retirement Longitudinal Study (CHARLS), a nationally representative longitudinal cohort of community-dwelling adults aged 45 years or older in China. CHARLS used a multistage probability sampling design and systematically collected information on sociodemographic characteristics, lifestyle factors, medical history, medication use, anthropometric measurements, physical examinations, and blood biomarkers. The national baseline survey was conducted in 2011, with follow-up surveys in 2013, 2015, 2018, and 2020 [20, 21]. Although the CHARLS source cohort was designed to be nationally representative, the present analyses were based on a selected analytical subset and did not incorporate survey sampling weights; therefore, the final analytical cohort should not be interpreted as nationally representative.
The present study included a primary landmark cohort and a baseline supportive cohort. For the primary analysis, 2011–2015 was defined as the exposure assessment period. Cumulative eGDR derived from the 2011 and 2015 measurements was treated as the primary exposure. Because only two measurements were available over the same four-year interval for all participants, cumulative eGDR was interpreted as a time-weighted summary that is mathematically proportional to the two-wave mean eGDR. The year 2015 was specified as the landmark separating the exposure assessment period from the outcome follow-up period, and participants were followed from 2015 to 2020 for incident cardiovascular disease (CVD). This temporal design ensured that both eGDR measurements used to define longitudinal exposure preceded the start of outcome follow-up.
Participants were eligible for the primary analysis if they were aged ≥ 45 years in 2011, met the operational TyG-WC–based definition of MASLD in 2011, had calculable eGDR in both 2011 and 2015, were alive and free of CVD at the 2015 landmark, and had valid post-2015 CVD outcome and follow-up data. Participant selection began with 17,708 individuals in the 2011 CHARLS baseline survey. We first excluded 7,709 participants who lacked data required to calculate eGDR in either 2011 or 2015, leaving 9,999 participants. We then excluded 430 participants aged < 45 years in 2011 and 5,442 participants who did not meet the operational MASLD definition in 2011, leaving 4,127 participants. A further 2,042 participants were excluded because they had died before the 2015 landmark or had reported physician-diagnosed heart disease or stroke in 2011, 2013, or 2015, leaving 2,085 participants. Finally, 26 participants without valid post-2015 CVD outcome or follow-up data were excluded, resulting in a final analytical cohort of 2,059 participants. The detailed participant selection process is shown in Fig. 1.
Fig. 1.
Flow chart of participant selection
In addition, we constructed a 2011 baseline supportive cohort to evaluate the association between a single baseline eGDR measurement and long-term CVD risk and to provide complementary evidence that did not require survival to the 2015 landmark or the availability of repeated eGDR measurements. This cohort included participants who were aged ≥ 45 years in 2011, met the operational MASLD definition, were free of heart disease and stroke at baseline, had calculable baseline eGDR, and had valid CVD outcome and follow-up data. Participants were followed from 2011 to 2020. A total of 3,464 participants were included in the baseline supportive analysis.
Definition of MASLD
Because CHARLS did not include liver imaging or histological assessments, MASLD was operationally classified using sex-specific thresholds of the triglyceride-glucose waist circumference index (TyG-WC) as a surrogate population definition, in accordance with previous CHARLS-based studies [22]. This operational classification was used for cohort identification and should not be considered equivalent to imaging- or histology-confirmed hepatic steatosis.
The TyG index was calculated only among participants with fasting blood samples [23]:
 |
TyG-WC was then calculated as:
 |
FPG denotes fasting plasma glucose, TG triglycerides, and WC waist circumference. Participants with TyG-WC ≥ 750 in men or ≥ 540 in women were classified as meeting the operational MASLD definition, whereas those below the corresponding sex-specific threshold were classified as not meeting the operational definition. Participants with missing fasting status, FPG, TG, waist circumference, or sex could not be classified.
The primary study population was defined according to the operational MASLD classification in 2011. For analyses of the course of the operational MASLD classification, 2015 status was determined using the same method. Participants meeting the definition in both 2011 and 2015 were classified as having persistent MASLD according to the operational definition, whereas those meeting the definition in 2011 but not in 2015 were classified as having remitted MASLD according to the operational definition. Because waist circumference is a component of both TyG-WC and eGDR, the population-defining criterion and the eGDR exposure share a common component. This overlap was considered in the interpretation of the findings and precludes interpreting the observed association as independent of adiposity.
Assessment of eGDR
eGDR was calculated separately for the 2011 and 2015 survey waves using waist circumference, hypertension status, and glycated haemoglobin (HbA1c) [11, 13]:
 |
Hypertension was coded as yes = 1 and no = 0. Hypertension was defined by any of the following: a previous physician diagnosis, current use of antihypertensive medication, systolic blood pressure ≥ 140 mmHg, or diastolic blood pressure ≥ 90 mmHg. Higher eGDR indicates better estimated insulin sensitivity, whereas lower eGDR indicates poorer estimated insulin sensitivity. Because eGDR is deterministically derived from waist circumference, hypertension status, and HbA1c, it was interpreted in the present study as a composite cardiometabolic marker and surrogate of insulin sensitivity rather than as a direct or component-independent measure of insulin resistance.
Cumulative eGDR
Cumulative eGDR during 2011–2015 was calculated using the trapezoidal method, consistent with previous studies of cumulative eGDR exposure [18]:
 |
Because only two eGDR measurements were available and all participants had the same four-year exposure interval, cumulative eGDR was mathematically proportional to the arithmetic mean of the 2011 and 2015 eGDR values. Thus, although the term “cumulative eGDR” was retained when referring to the conventional calculation for comparability with previous literature, this measure was interpreted primarily as the time-weighted two-wave mean eGDR rather than as a measure of cumulative dose or exposure duration.
For the primary association analyses, cumulative eGDR was reverse-coded and standardized as a z score so that higher modeled values represented lower cumulative eGDR. Because all participants had the same four-year exposure interval, this standardized exposure is mathematically equivalent to the corresponding standardized two-wave mean eGDR. Associations were therefore expressed as the HR per 1-SD lower cumulative eGDR and across quartiles, with the highest cumulative eGDR quartile as the reference.
Alternative exposure analyses also considered 2011 eGDR, 2015 eGDR, and the change between measurements:
 |
In the supportive 2011 baseline analysis, baseline eGDR was reverse-coded and standardized, and associations were expressed per 1-SD lower baseline eGDR and across quartiles, with the highest baseline eGDR quartile as the reference.
Longitudinal eGDR patterns (two-wave level–change patterns)
Paired raw eGDR values from 2011 and 2015 were entered into K-means clustering to characterise longitudinal eGDR patterns in level and change between the two assessments, consistent with previous analyses of repeated eGDR measurements [24]. Because only two eGDR measurements were available, these longitudinal patterns were operationally interpreted as two-wave eGDR level–change patterns rather than formal long-term trajectories. Candidate clustering solutions with K = 2–6 were evaluated using the average silhouette width, Calinski–Harabasz index, Davies–Bouldin index, cluster size, resampling stability, and clinical interpretability. Clustering diagnostics were also repeated after standardising eGDR values within each survey wave.
The principal internal clustering indices favored a two-cluster solution for both raw and standardized eGDR values. However, a four-cluster solution was retained as the primary exploratory representation because it provided more granular and clinically interpretable combinations of eGDR level and change while maintaining acceptable cluster sizes and bootstrap stability. CVD outcome information was not used to select the clustering solution, and two- and three-cluster solutions were examined in sensitivity analyses (Supplementary Table S6).
According to the cluster centroids, the four patterns were labelled as follows: high eGDR–slight decline, moderate-low eGDR–slight rise, high eGDR–marked decline, and low eGDR–slight decline. These pattern labels describe the two-wave profiles represented by the cluster centroids and do not imply that every participant within a cluster changed in the same direction. Cluster stability was assessed using 1,000 bootstrap resamples and Jaccard similarity coefficients.
CVD outcomes and follow-up
The primary outcome was incident CVD during 2015–2020, defined as the first report of physician-diagnosed heart disease or stroke in the 2018 or 2020 survey. Heart disease and stroke were analysed separately as secondary outcomes. These outcomes were not mutually exclusive, and a participant could contribute to both during follow-up.
In the primary Cox analyses, event time was assigned to the survey wave in which the corresponding diagnosis was first reported, corresponding to approximately 3 years after the 2015 landmark for events first reported in 2018 and 5 years for events first reported in 2020. Participants without CVD were censored, as appropriate, at death, the last confirmed follow-up, or 2020. In the primary Cox analyses, death before CVD was treated as a censoring event, whereas in Fine-Gray analyses it was treated as a competing event. Because outcomes were ascertained by survey wave, event times represent interval approximations rather than exact clinical onset dates. To assess the potential influence of this approximation, additional sensitivity analyses were performed using midpoint-assigned event times and a proportional-hazards model that explicitly accounted for interval-censored event times, as described below.
The supportive 2011 baseline analysis used the same CVD definition. An event was recorded when heart disease or stroke was first reported in the 2013, 2015, 2018, or 2020 survey. Follow-up time was calculated from 2011 to the first CVD report, death, last confirmed follow-up, or 2020. Participants who died before CVD were censored as non-CVD events.
Covariates
Covariates included sociodemographic characteristics, lifestyle factors, medical history, medication use, and laboratory measures. All covariates used in the primary association models were assessed at the 2011 baseline, corresponding to the start of the 2011–2015 exposure-assessment period. Changes in these covariates during 2011–2015 were not included in the primary adjustment models.
Three sequentially adjusted models were fitted. Model 1 adjusted for age and sex. Model 2 additionally adjusted for educational attainment, place of residence, marital status, smoking status, drinking status, and physical activity. Model 3 further adjusted for dyslipidaemia, diabetes, cancer, chronic kidney disease, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, estimated glomerular filtration rate, and use of lipid-lowering medication.
Disease status and medication use were defined using harmonised CHARLS questionnaire, physical examination, and biomarker data, as appropriate. Because waist circumference, hypertension, and HbA1c are deterministic components of the eGDR calculation, they were not included as conventional covariates in the primary association models to avoid adjustment for components of the exposure. This modelling choice was intended to avoid conditioning on variables that mathematically define eGDR and should not be interpreted as demonstrating an association independent of adiposity, hypertension, or glycaemia. The incremental value of longitudinal eGDR measures beyond conventional risk factors was evaluated separately in the prediction analyses.
Statistical analysis
Descriptive analyses, missing data, and the 2011 baseline analysis
Continuous variables are presented as means with standard deviations or medians with interquartile ranges, as appropriate, and categorical variables as numbers and percentages. Characteristics of the primary analytical cohort were described according to quartiles of cumulative eGDR and the four two-wave eGDR level–change patterns. Characteristics of the 2011 supportive cohort were described according to quartiles of baseline eGDR. Continuous variables were compared using analysis of variance or the Kruskal–Wallis test, as appropriate, and categorical variables using the chi-squared test or Fisher’s exact test.
To evaluate potential selection related to the landmark-cohort eligibility criteria, baseline characteristics were additionally compared between the 2,059 participants included in the primary analysis and 3,916 excluded participants who were aged ≥ 45 years and met the operational MASLD definition in 2011. Selection-related differences were summarized using absolute standardized mean differences (SMDs), with an absolute SMD ≥ 0.10 considered indicative of a potentially meaningful difference; no significance tests were used for this comparison (Supplementary Table S3).
Missing covariate data were handled using multiple imputation by chained equations, and regression estimates were combined according to Rubin’s rules [25]. Five imputed datasets were generated with five iterations per dataset. The imputation model included age, sex, education, residence, marital status, smoking, drinking, physical activity, hypertension, dyslipidaemia, diabetes, cancer, chronic kidney disease, antihypertensive medication, glucose-lowering medication, lipid-lowering medication, body mass index, waist circumference, systolic and diastolic blood pressure, HbA1c, LDL-C, HDL-C, and estimated glomerular filtration rate, where available. Continuous variables were imputed using predictive mean matching, binary variables using logistic regression, ordered categorical variables using proportional-odds logistic regression, and unordered categorical variables using multinomial logistic regression.
To preserve information related to the time-to-event outcome, CVD event status and the Nelson–Aalen cumulative hazard were included as predictors in the imputation model; raw follow-up time was not used as a predictor once the Nelson–Aalen estimate was included. The primary exposures, MASLD classification, outcomes, and follow-up time were not imputed. Variable-specific missing counts and percentages before imputation are reported in Supplementary Table S1. Multiple imputation was restricted to missing covariate data within the analytical cohort and therefore did not recover participants who had been excluded because eGDR or the operational MASLD classification could not be determined. Complete-case analyses were also performed to assess the influence of missing-data handling on the findings.
In the 2011 supportive cohort, Cox proportional hazards models and restricted cubic splines consistent with those used in the primary analysis were applied to evaluate the association between baseline eGDR and incident CVD during 2011–2020. Baseline eGDR was analysed continuously and in quartiles, with associations expressed per 1-SD lower baseline eGDR and with the highest baseline eGDR quartile as the reference.
Primary association analyses
Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% confidence intervals for the association between cumulative eGDR and incident CVD. As the primary continuous association was modelled in the direction of lower eGDR, estimates were expressed as the HR per 1-SD lower cumulative eGDR. Cumulative eGDR was also analysed in quartiles ordered from higher to lower values, with the highest cumulative eGDR quartile as the reference, and a test for linear trend was performed across the ordered quartiles.
The same sequentially adjusted models were used to evaluate the association between the two-wave eGDR level–change patterns and incident CVD, with the high eGDR–slight decline pattern as the reference. Kaplan–Meier curves and the log-rank test were used to compare unadjusted cumulative incidence across the pattern groups, and standardised cumulative incidence was estimated from the fully adjusted model. These pattern-based analyses were considered exploratory.
Restricted cubic splines were used to examine the dose–response association between cumulative eGDR and incident CVD. Four knots were placed at the 5th, 35th, 65th, and 95th percentiles of the exposure distribution. The median cumulative eGDR was used as the reference, and tests were conducted for both the overall association and non-linearity.
Heart disease and stroke analyses
Fully adjusted Cox proportional hazards models were used to evaluate the associations of cumulative eGDR and the two-wave eGDR level–change patterns with incident heart disease and stroke separately. For cumulative eGDR, associations were expressed per 1-SD lower value, and the high eGDR–slight decline pattern was used as the reference category for the pattern analyses. False discovery rate correction was applied to multiple testing of the prespecified secondary outcomes using the Benjamini–Hochberg procedure within the secondary-outcome family.
Alternative eGDR measures and other cumulative metabolic indices
To compare the risk information provided by different eGDR representations, we separately evaluated 2011 eGDR, 2015 eGDR, the change between the two measurements, cumulative eGDR, and the two-wave eGDR level–change patterns in relation to incident CVD. Because absolute change and annualised change provide equivalent information after standardisation, the results for absolute change are reported primarily.
For comparison, repeated measurements available in 2011 and 2015 were used to calculate cumulative values of the triglyceride–glucose index (TyG), TyG–waist circumference, TyG–waist-to-height ratio, metabolic score for insulin resistance (METS-IR), C-reactive protein–triglyceride–glucose index, atherogenic index of plasma, cholesterol–high-density lipoprotein–glucose index, and remnant cholesterol. Each index was calculated according to previously published formulae [26–29], and cumulative exposure during 2011–2015 was derived using the same trapezoidal method applied to cumulative eGDR.
Each cumulative index was standardised and entered into the same covariate-adjusted models. Associations by quartiles, dose–response relationships, and apparent C-indices were also evaluated. These between-index analyses were considered exploratory. Because the available analytical sample sizes differed across indices, between-index comparisons were interpreted descriptively and were not used to support a direct ranking of predictive performance. The Benjamini–Hochberg procedure was applied separately within the alternative-eGDR and continuous cumulative-index analysis families; quartile analyses were treated as exploratory.
Prediction and incremental analyses
Prediction analyses were conducted to determine whether cumulative eGDR and the two-wave eGDR level–change patterns provided additional risk information beyond conventional risk factors, rather than to develop a final risk score for direct clinical use. These analyses were considered exploratory model-comparison analyses.
The fully adjusted Model 3 was used as the base model. The base model was compared with models that separately added cumulative eGDR, the two-wave eGDR level–change patterns, baseline eGDR, 2015 eGDR, or the 2011 eGDR components (waist circumference, hypertension status, and HbA1c). Models containing the 2011 eGDR components with and without cumulative eGDR were additionally compared to assess whether cumulative eGDR provided information beyond its constituent variables. Model discrimination and fit were assessed using the C-index, Akaike information criterion, and likelihood ratio tests. Mutually adjusted models were used to compare the overlap in information provided by different eGDR measures, including direct comparisons of cumulative eGDR with baseline eGDR and with the most recent 2015 eGDR measurement. Because cumulative eGDR is mathematically derived from the repeated eGDR measurements, coefficients from mutually adjusted models were interpreted cautiously.
For the principal prediction comparison, candidate models were evaluated in a common sample of 2,042 participants with the required predictor information. Apparent C-indices were distinguished from internally validated time-specific performance. Three-year area under the curve (AUC) and Brier scores were internally validated using bootstrap out-of-bag estimation with 500 resamples within each imputed dataset. The paired out-of-bag C-index was used to quantify the incremental discrimination achieved by cumulative eGDR to specified reduced models, using the same bootstrap resamples for each paired comparison; these paired comparisons used 200 bootstrap resamples within each imputed dataset. Five-year time-dependent receiver operating characteristic curves, calibration curves, and decision-curve analyses were based on multiple-imputation-averaged apparent risk predictions. Continuous net reclassification improvement and integrated discrimination improvement were also evaluated at 5 years as exploratory apparent measures and were not bootstrap-corrected [30–32]. Additional comparisons involving METS-IR were conducted in a separate stricter common sample because of additional missingness in repeated METS-IR measurements.
Subgroup and MASLD course-related analyses
Prespecified subgroups were defined by age, sex, diabetes, hypertension, obesity, and MASLD course. The association between cumulative eGDR and incident CVD was re-estimated within each subgroup, with estimates expressed per 1-SD lower cumulative eGDR. Effect heterogeneity was assessed by including an interaction term between cumulative eGDR, modelled in the lower-value direction, and the subgroup variable in the model. Formal P values for interaction were reported, and the Benjamini–Hochberg procedure was used to control the false discovery rate across the prespecified subgroup-interaction family. Subgroup-specific models used the fully adjusted Model 3 covariate set, excluding the stratifying variable from the adjustment set where applicable.
Among participants whose MASLD status in 2015 could be determined, MASLD course was combined with a binary eGDR pattern classification to define joint phenotypes. The binary pattern classification grouped the high eGDR–slight decline and moderate-low eGDR–slight rise patterns together, and the high eGDR–marked decline and low eGDR–slight decline patterns together. Associations between these joint phenotypes and incident CVD were evaluated using the fully adjusted model. These joint-phenotype and corresponding interaction analyses were considered exploratory.
Sensitivity analyses and model diagnostics
Sensitivity analyses included exclusion of participants who first reported CVD in 2018, restriction to participants with persistent MASLD, exclusion of participants with baseline diabetes or use of glucose-lowering medication, complete-case analysis, and alternative definitions of the two-wave eGDR level–change patterns, including two- and three-cluster K-means solutions, clustering of eGDR values standardised within each survey wave, and sex- and wave-specific median-based level-change combinations.
To assess the influence of uncertainty in CVD onset time, two additional event-time sensitivity analyses were performed. First, event times were assigned to the midpoint of the interval in which CVD was first reported and the fully adjusted Cox model was refitted. Second, a semiparametric proportional-hazards model for interval-censored data was fitted. For participants first reporting CVD in 2018, the event interval was defined as (0, 3] years after the 2015 landmark; for those first reporting CVD in 2020 after a negative 2018 assessment, the interval was (3, 5] years; and for those first reporting CVD in 2020 with missing CVD status in 2018, the interval was (0, 5] years. Participants without reported CVD were right-censored at the same censoring time used in the primary analysis, namely death, the last confirmed follow-up, or 2020, as applicable. These analyses used the same fully adjusted Model 3 covariate set as the primary association analysis.
Fine–Gray models were fitted with death before CVD as a competing event [33]. E-values were calculated to assess the potential influence of unmeasured confounding on the primary associations [34].
The proportional hazards assumption was examined using Schoenfeld residuals, and because physical activity showed evidence of non-proportional hazards, additional Cox models were fitted with the baseline hazard stratified by physical activity while retaining the remaining Model 3 covariates. Generalised variance inflation factors were used to assess multicollinearity. Because cumulative eGDR is mathematically derived from the repeated eGDR measurements, multicollinearity was also examined in models containing baseline eGDR or the eGDR components together with cumulative eGDR. For METS-IR, robustness was further evaluated in models excluding covariates related to its component variables.
All statistical tests were two-sided. Except for analyses in which false discovery rate correction was applied, a P value < 0.05 was considered statistically significant. All analyses were conducted using R software.
Results
Baseline characteristics
The primary landmark cohort included 2,059 participants (Fig. 1). At the 2011 baseline, the mean age was 57.7 ± 8.5 years, and 1,510 participants (73.3%) were women. During follow-up from 2015 to 2020, 457 participants first reported physician-diagnosed heart disease or stroke, corresponding to a cumulative event proportion of 22.2%.
Compared with the 3,916 excluded potentially eligible participants who were aged ≥ 45 years and met the operational MASLD definition in 2011, the included participants differed in several baseline characteristics (Supplementary Table S3). Absolute SMDs ≥ 0.10 were observed for age, residence, marital status, systolic blood pressure, eGFR, hypertension, dyslipidaemia, and chronic kidney disease. Included participants were generally younger, more often married and living in villages, had lower systolic blood pressure and higher eGFR, and had lower observed prevalences of hypertension, dyslipidaemia, and chronic kidney disease.
Across quartiles ordered from higher to lower cumulative eGDR, the proportions of incident CVD were 14.4%, 16.9%, 23.3%, and 34.2%, respectively (Table 1). Participants with lower cumulative eGDR were generally older and more likely to be male and to have diabetes or dyslipidaemia and to use cardiometabolic medications. As expected from the construction of eGDR, waist circumference, hypertension, and HbA1c also showed pronounced gradients across cumulative eGDR quartiles; these differences are partly structural because these variables contribute to the eGDR calculation. Participant characteristics according to the four two-wave eGDR level–change patterns are presented in Supplementary Table S2.
Table 1.
Baseline characteristics according to quartiles of cumulative eGDR
| Characteristic | Overall N = 2,0591 |
Q1 N = 5151 |
Q2 N = 5151 |
Q3 N = 5151 |
Q4 N = 5141 |
p value2 |
|---|---|---|---|---|---|---|
| Age, years | 57.7 ± 8.5 | 55.1 ± 7.6 | 56.3 ± 8.3 | 59.6 ± 8.8 | 59.8 ± 8.4 | < 0.001 |
| Sex | < 0.001 | |||||
| Female | 1510 (73.3%) | 481 (93.4%) | 362 (70.3%) | 362 (70.3%) | 305 (59.3%) | |
| Male | 549 (26.7%) | 34 (6.6%) | 153 (29.7%) | 153 (29.7%) | 209 (40.7%) | |
| Education level | 0.178 | |||||
| College or above | 58 (2.8%) | 13 (2.5%) | 15 (2.9%) | 14 (2.7%) | 16 (3.1%) | |
| High school | 545 (26.5%) | 126 (24.5%) | 158 (30.7%) | 121 (23.5%) | 140 (27.2%) | |
| Primary school or below | 1456 (70.7%) | 376 (73.0%) | 342 (66.4%) | 380 (73.8%) | 358 (69.6%) | |
| Residence | 0.033 | |||||
| City/town | 301 (14.6%) | 56 (10.9%) | 78 (15.2%) | 79 (15.3%) | 88 (17.2%) | |
| Village | 1756 (85.4%) | 459 (89.1%) | 436 (84.8%) | 436 (84.7%) | 425 (82.8%) | |
| Marital status | < 0.001 | |||||
| Married | 1874 (91.0%) | 481 (93.4%) | 479 (93.0%) | 445 (86.4%) | 469 (91.2%) | |
| Never-married/separated/widowed | 185 (9.0%) | 34 (6.6%) | 36 (7.0%) | 70 (13.6%) | 45 (8.8%) | |
| Smoking status | < 0.001 | |||||
| Current smoker | 379 (18.5%) | 40 (7.8%) | 117 (22.8%) | 100 (19.5%) | 122 (23.8%) | |
| Ex-smoker | 119 (5.8%) | 8 (1.6%) | 29 (5.7%) | 37 (7.2%) | 45 (8.8%) | |
| Non-smoker | 1554 (75.7%) | 466 (90.7%) | 367 (71.5%) | 376 (73.3%) | 345 (67.4%) | |
| Drinking status | < 0.001 | |||||
| Current drink | 313 (15.8%) | 43 (8.6%) | 81 (16.1%) | 90 (18.6%) | 99 (19.9%) | |
| Ex-drink | 133 (6.7%) | 26 (5.2%) | 40 (8.0%) | 40 (8.3%) | 27 (5.4%) | |
| Never | 1535 (77.5%) | 430 (86.2%) | 381 (75.9%) | 353 (73.1%) | 371 (74.6%) | |
| Physical activity | 0.169 | |||||
| Moderate | 512 (24.9%) | 121 (23.5%) | 125 (24.3%) | 130 (25.2%) | 136 (26.5%) | |
| No | 1247 (60.6%) | 302 (58.6%) | 310 (60.2%) | 322 (62.5%) | 313 (60.9%) | |
| Vigorous | 300 (14.6%) | 92 (17.9%) | 80 (15.5%) | 63 (12.2%) | 65 (12.6%) | |
| Hypertension | < 0.001 | |||||
| No | 1176 (57.1%) | 507 (98.4%) | 461 (89.5%) | 194 (37.7%) | 14 (2.7%) | |
| Yes | 883 (42.9%) | 8 (1.6%) | 54 (10.5%) | 321 (62.3%) | 500 (97.3%) | |
| Dyslipidemia | < 0.001 | |||||
| No | 1562 (77.4%) | 430 (84.6%) | 413 (81.5%) | 393 (78.6%) | 326 (64.7%) | |
| Yes | 457 (22.6%) | 78 (15.4%) | 94 (18.5%) | 107 (21.4%) | 178 (35.3%) | |
| Diabetes | < 0.001 | |||||
| No | 1676 (81.4%) | 478 (92.8%) | 451 (87.6%) | 411 (79.8%) | 336 (65.4%) | |
| Yes | 383 (18.6%) | 37 (7.2%) | 64 (12.4%) | 104 (20.2%) | 178 (34.6%) | |
| Cancer | 0.084 | |||||
| No | 2040 (99.5%) | 511 (99.4%) | 513 (99.8%) | 512 (99.8%) | 504 (98.8%) | |
| Yes | 11 (0.5%) | 3 (0.6%) | 1 (0.2%) | 1 (0.2%) | 6 (1.2%) | |
| Chronic kidney disease | 0.478 | |||||
| No | 1956 (95.0%) | 483 (93.8%) | 492 (95.5%) | 493 (95.7%) | 488 (94.9%) | |
| Yes | 103 (5.0%) | 32 (6.2%) | 23 (4.5%) | 22 (4.3%) | 26 (5.1%) | |
| Antihypertensive medication | < 0.001 | |||||
| No | 1677 (81.9%) | 512 (100.0%) | 507 (99.0%) | 407 (79.5%) | 251 (49.0%) | |
| Yes | 371 (18.1%) | 0 (0.0%) | 5 (1.0%) | 105 (20.5%) | 261 (51.0%) | |
| Antidiabetic medication | < 0.001 | |||||
| No | 1948 (95.4%) | 507 (98.8%) | 500 (97.7%) | 488 (96.3%) | 453 (89.0%) | |
| Yes | 93 (4.6%) | 6 (1.2%) | 12 (2.3%) | 19 (3.7%) | 56 (11.0%) | |
| Lipid-lowering medication | < 0.001 | |||||
| No | 1918 (95.2%) | 498 (98.2%) | 493 (97.2%) | 488 (97.8%) | 439 (87.6%) | |
| Yes | 96 (4.8%) | 9 (1.8%) | 14 (2.8%) | 11 (2.2%) | 62 (12.4%) | |
| BMI, kg/m2 | 24.5 ± 3.6 | 22.0 ± 3.0 | 24.7 ± 3.1 | 24.4 ± 3.1 | 27.1 ± 3.3 | < 0.001 |
| Waist circumference, cm | 88.1 ± 9.5 | 79.1 ± 6.1 | 88.7 ± 7.7 | 88.4 ± 8.2 | 96.3 ± 7.0 | < 0.001 |
| Waist-to-height ratio | 0.565 ± 0.086 | 0.517 ± 0.090 | 0.562 ± 0.046 | 0.569 ± 0.108 | 0.611 ± 0.058 | < 0.001 |
| Systolic blood pressure, mmHg | 129.4 ± 20.4 | 114.3 ± 11.2 | 120.7 ± 13.9 | 137.2 ± 19.0 | 145.8 ± 19.0 | < 0.001 |
| Diastolic blood pressure, mmHg | 75.5 ± 11.6 | 68.5 ± 8.7 | 72.0 ± 9.5 | 78.5 ± 10.7 | 83.3 ± 11.3 | < 0.001 |
| Fasting plasma glucose, mg/dL | 113.0 ± 38.5 | 101.9 ± 15.5 | 106.5 ± 21.7 | 115.7 ± 41.6 | 127.9 ± 55.7 | < 0.001 |
| Total cholesterol, mg/dL | 198.4 ± 39.1 | 191.9 ± 35.5 | 196.2 ± 37.1 | 202.0 ± 40.6 | 203.5 ± 41.8 | < 0.001 |
| HbA1c, % | 5.35 ± 0.92 | 5.03 ± 0.39 | 5.22 ± 0.57 | 5.40 ± 0.98 | 5.75 ± 1.29 | < 0.001 |
| Triglycerides, mg/dL | 150.3 ± 123.5 | 120.5 ± 93.2 | 142.9 ± 89.0 | 157.9 ± 133.1 | 180.0 ± 157.8 | < 0.001 |
| HDL-C, mg/dL | 49.2 ± 14.3 | 54.2 ± 14.3 | 48.9 ± 14.5 | 49.2 ± 14.5 | 44.5 ± 12.3 | < 0.001 |
| LDL-C, mg/dL | 119.9 ± 35.3 | 115.6 ± 32.6 | 119.0 ± 34.2 | 121.6 ± 35.5 | 123.3 ± 38.5 | 0.003 |
| Serum creatinine, mg/dL | 0.73 ± 0.17 | 0.69 ± 0.13 | 0.73 ± 0.17 | 0.74 ± 0.16 | 0.77 ± 0.19 | < 0.001 |
| eGFR, mL/min/1.73 m2 | 93.9 ± 13.7 | 95.8 ± 13.4 | 95.1 ± 13.4 | 92.7 ± 13.6 | 92.0 ± 14.3 | < 0.001 |
| C-reactive protein, mg/L | 1.04 (0.55–2.01) | 0.70 (0.42–1.30) | 0.99 (0.53–1.86) | 1.08 (0.58–2.12) | 1.44 (0.78–2.71) | < 0.001 |
| eGDR in 2011, mg/kg/min | 8.8 ± 2.2 | 11.2 ± 0.6 | 10.0 ± 1.0 | 8.1 ± 1.3 | 6.0 ± 0.9 | < 0.001 |
| eGDR in 2015, mg/kg/min | 8.1 ± 2.3 | 10.8 ± 0.7 | 9.1 ± 1.2 | 7.0 ± 1.2 | 5.4 ± 1.0 | < 0.001 |
| Change in eGDR, 2011–2015 | -0.7 ± 1.6 | -0.4 ± 0.8 | -0.8 ± 1.9 | -1.1 ± 2.2 | -0.6 ± 1.1 | < 0.001 |
| Annual change in eGDR | -0.2 ± 0.4 | -0.1 ± 0.2 | -0.2 ± 0.5 | -0.3 ± 0.5 | -0.1 ± 0.3 | < 0.001 |
| Cumulative eGDR | 33.9 ± 8.4 | 44.2 ± 2.1 | 38.2 ± 1.9 | 30.3 ± 2.3 | 22.9 ± 3.2 | < 0.001 |
| Incident cardiovascular disease | < 0.001 | |||||
| No | 1602 (77.8%) | 441 (85.6%) | 428 (83.1%) | 395 (76.7%) | 338 (65.8%) | |
| Yes | 457 (22.2%) | 74 (14.4%) | 87 (16.9%) | 120 (23.3%) | 176 (34.2%) |
Values are mean ± SD, median (P25–P75) for C-reactive protein, or n (%)
Descriptive statistics are based on the original pre-imputation observations. For variables with missing values, categorical percentages and continuous summaries use the variable-specific nonmissing observations; therefore denominators may vary
Quartiles were ordered from higher to lower cumulative eGDR (Q1–Q4), with Q1 representing the highest cumulative eGDR
P values were calculated using Welch one-way analysis of variance for approximately symmetric continuous variables, the Kruskal–Wallis test for C-reactive protein, and Pearson chi-square tests for categorical variables; Monte Carlo chi-square P values were used when expected cell counts were small
Waist circumference, hypertension, and HbA1c are components of eGDR; between-quartile differences in these variables are partly structural
Associations of cumulative eGDR and two-wave eGDR level–change patterns with incident CVD
Lower cumulative eGDR was consistently associated with a higher risk of incident CVD across all adjusted models (Table 2). For each 1-SD lower cumulative eGDR, the HRs were 1.46 (95% CI 1.32–1.61), 1.45 (1.32–1.60), and 1.39 (1.26–1.55) in Models 1, 2, and 3, respectively (all P < 0.001). Quartiles were ordered from higher to lower cumulative eGDR (Q1–Q4), with Q1 as the reference. The fully adjusted HRs for Q2, Q3, and Q4 were 1.19 (95% CI 0.87–1.64), 1.64 (1.21–2.22), and 2.40 (1.77–3.25), respectively, with a significant trend across quartiles (P for trend < 0.001). Restricted cubic spline analysis showed a significant overall association without evidence of non-linearity (P overall < 0.001; P for non-linearity = 0.915; Fig. 2).
Table 2.
Associations of cumulative eGDR and two-wave eGDR level–change patterns with incident cardiovascular disease
| Panel A. Cumulative eGDR | |||||||
|---|---|---|---|---|---|---|---|
| Exposure | Category | Model 1 | Model 2 | Model 3 | |||
| HR (95%CI) | P | HR (95%CI) | P | HR (95%CI) | P | ||
| Cumulative eGDR | Per 1-SD lower cumulative eGDR | 1.46 (1.32, 1.61) | < 0.001 | 1.45 (1.32, 1.60) | < 0.001 | 1.39 (1.26, 1.55) | < 0.001 |
| Quartiles of cumulative eGDR | Q1 | Reference | Reference | Reference | |||
| Q2 | 1.22 (0.89, 1.67) | 0.213 | 1.21 (0.89, 1.66) | 0.228 | 1.19 (0.87, 1.64) | 0.279 | |
| Q3 | 1.68 (1.25, 2.26) | < 0.001 | 1.67 (1.24, 2.25) | < 0.001 | 1.64 (1.21, 2.22) | 0.001 | |
| Q4 | 2.68 (2.02, 3.57) | < 0.001 | 2.68 (2.01, 3.56) | < 0.001 | 2.40 (1.77, 3.25) | < 0.001 | |
| P for trend | < 0.001 | < 0.001 | < 0.001 | ||||
| Panel B. K-means-defined two-wave eGDR level–change patterns | |||||||
|---|---|---|---|---|---|---|---|
| Exposure | Pattern | Model 1 | Model 2 | Model 3 | |||
| HR (95%CI) | P | HR (95%CI) | P | HR (95%CI) | P | ||
| Two-wave eGDR level–change pattern | High eGDR–slight decline | Reference | Reference | Reference | |||
| Moderate-low eGDR–slight rise | 1.47 (1.11, 1.95) | 0.008 | 1.47 (1.10, 1.95) | 0.008 | 1.43 (1.07, 1.90) | 0.014 | |
| High eGDR–marked decline | 1.64 (1.20, 2.22) | 0.002 | 1.64 (1.21, 2.23) | 0.002 | 1.64 (1.20, 2.23) | 0.002 | |
| Low eGDR–slight decline | 2.38 (1.88, 3.02) | < 0.001 | 2.38 (1.87, 3.02) | < 0.001 | 2.14 (1.67, 2.74) | < 0.001 | |
Analysis sample: N = 2059; incident cardiovascular disease events = 457
Values are hazard ratios (95% confidence intervals); P values are shown in separate columns
For the continuous analysis, HRs are expressed per 1-SD lower cumulative eGDR. Quartiles were ordered from higher to lower cumulative eGDR (Q1–Q4), with Q1 as the reference category. Lower cumulative eGDR indicates poorer estimated insulin sensitivity
The high eGDR–slight decline pattern was the reference group. Pattern names describe cluster centroids and do not imply that every individual changed in the same direction
Model 1 adjusted for age and sex
Model 2 additionally adjusted for education, residence, marital status, smoking, drinking, and physical activity
Model 3 additionally adjusted for dyslipidemia, diabetes, cancer, chronic kidney disease, LDL-C, HDL-C, eGFR, and lipid-lowering medication
Fig. 2.
Two-wave eGDR level–change patterns and adjusted cumulative incidence of CVD. A Mean eGDR levels in 2011 and 2015 across the four K-means-defined two-wave eGDR level–change patterns; error bars represent 95% confidence intervals. B Model-adjusted cumulative incidence of CVD from the 2015 landmark according to the two-wave eGDR level–change patterns. Pattern labels describe cluster centroids and do not imply that every participant experienced the same individual direction of change. CVD, cardiovascular disease; eGDR, estimated glucose disposal rate
For both raw and wave-standardised eGDR values, the principal internal clustering indices favored K = 2 rather than K = 4. Nevertheless, K = 4 was retained as the primary exploratory representation because it provided more granular level–change patterns while maintaining acceptable cluster sizes and bootstrap stability; CVD outcome information was not used in selecting the number of clusters (Supplementary Table S6 and Supplementary Fig. S2). The four longitudinal eGDR patterns, operationally characterised as two-wave eGDR level–change patterns, were high eGDR–slight decline in 839 participants (40.7%), moderate-low eGDR–slight rise in 381 participants (18.5%), high eGDR–marked decline in 266 participants (12.9%), and low eGDR–slight decline in 573 participants (27.8%). The corresponding cluster centroids for eGDR in 2011 and 2015 were 10.78 and 10.38, 7.68 and 8.11, 10.27 and 6.43, and 6.11 and 5.55 mg/kg/min, respectively (Fig. 3A). The corresponding mean changes were − 0.40, + 0.42, − 3.83, and − 0.57 mg/kg/min. In the moderate-low eGDR–slight rise pattern, 43.8% of participants showed an individual increase and 56.2% showed a decrease, underscoring that the pattern labels describe cluster centroids rather than uniform individual changes. The mean Jaccard similarity coefficients for the four patterns were 0.978, 0.967, 0.931, and 0.839, respectively, based on 1,000 bootstrap resamples.
The numbers of incident CVD events in the four pattern groups were 124, 83, 62, and 188, corresponding to event proportions of 14.8%, 21.8%, 23.3%, and 32.8%, respectively. Kaplan–Meier curves showed significant differences in cumulative incidence across the pattern groups (log-rank P < 0.001; Supplementary Fig. S3), and the fully adjusted standardised cumulative incidence showed the same ordering (Fig. 3B).
Fig. 3.
Dose–response association between cumulative eGDR and incident CVD. Hazard ratios and 95% confidence intervals were estimated using restricted cubic splines, with the median cumulative eGDR as the reference. Cumulative eGDR is displayed in its conventional direction, with higher values indicating better estimated insulin sensitivity. The histogram shows the exposure distribution. CI, confidence interval; CVD, cardiovascular disease; eGDR, estimated glucose disposal rate
Using the high eGDR–slight decline pattern as the reference, the fully adjusted HRs were 1.43 (95% CI 1.07–1.90; P = 0.014) for the moderate-low eGDR–slight rise pattern, 1.64 (1.20–2.23; P = 0.002) for the high eGDR–marked decline pattern, and 2.14 (1.67–2.74; P < 0.001) for the low eGDR–slight decline pattern (Table 2).
Associations of cumulative eGDR and two-wave eGDR level–change patterns with incident heart disease and stroke
During follow-up, 308 incident heart disease events and 187 incident stroke events were reported. For each 1-SD lower cumulative eGDR, the fully adjusted HRs were 1.33 (95% CI 1.17–1.51) for heart disease and 1.44 (1.22–1.70) for stroke. Both associations remained statistically significant after Benjamini–Hochberg false-discovery-rate correction within the secondary-outcome family (both FDR-adjusted P < 0.001; Supplementary Table S7).
Compared with the high eGDR–slight decline pattern, the high eGDR–marked decline and low eGDR–slight decline patterns were associated with higher risks of both heart disease and stroke, and the associations remained significant after false discovery rate correction. For the moderate-low eGDR–slight rise pattern, the HR was 1.42 (95% CI 1.01–2.01) for heart disease (nominal P = 0.045; FDR-adjusted P = 0.052) and 1.50 (0.94–2.39) for stroke, with relatively wide confidence intervals. Effect estimates for overall CVD, heart disease, stroke, and the prespecified sensitivity analyses are summarised in Fig. 4.
Fig. 4.
Associations of cumulative eGDR and two-wave eGDR level–change patterns with incident CVD. Forest plot summarising the primary cumulative eGDR association, two-wave eGDR level–change pattern estimates, secondary outcomes, and sensitivity analyses. Continuous cumulative eGDR estimates are expressed per 1-SD lower value; the quartile comparison is Q4 versus Q1, with Q1 representing the highest cumulative eGDR. Estimates were derived from fully adjusted models. The Fine–Gray competing-risk estimate is reported as a subdistribution hazard ratio (sHR); all other effect estimates are hazard ratios (HRs). CI, confidence interval; CVD, cardiovascular disease; HR, hazard ratio; sHR, subdistribution hazard ratio
Associations of alternative eGDR measures with incident CVD
In the 2015 landmark cohort, higher eGDR in both 2011 and 2015 was associated with a lower risk of incident CVD. The fully adjusted HRs per 1-SD increase were 0.76 (95% CI 0.69–0.84) for 2011 eGDR and 0.72 (0.65–0.80) for 2015 eGDR. In contrast, the association between the absolute change in eGDR from 2011 to 2015 and incident CVD was weaker, with an HR of 0.92 (95% CI 0.84–1.01; P = 0.092) (Supplementary Table S8).
The 2011 baseline supportive cohort included 3,464 participants, among whom 987 developed incident CVD during follow-up to 2020. Each 1-SD lower baseline eGDR was associated with a fully adjusted HR of 1.33 (95% CI 1.24–1.43; P < 0.001). Using Q1 as the reference, the HRs for Q2, Q3, and Q4 were 1.19 (95% CI 0.96–1.46), 1.54 (1.26–1.88), and 2.11 (1.72–2.59), respectively (P for trend < 0.001; Supplementary Table S5). Restricted cubic spline analysis showed a significant overall association without clear evidence of non-linearity (P overall < 0.001; P for non-linearity = 0.847; Supplementary Fig. S1).
Associations of other cumulative metabolic indices with incident CVD
Among the other cumulative metabolic indices, the fully adjusted HRs per 1-SD increase were 1.26 (95% CI 1.12–1.41) for cumulative TyG-WC and 1.13 (1.04–1.23) for cumulative TyG-WHtR. In quartile analyses, the HRs comparing the highest with the lowest quartile were 1.76 (1.28–2.42) for cumulative TyG-WC, 1.61 (1.19–2.19) for cumulative TyG-WHtR, and 1.53 (1.08–2.18) for cumulative METS-IR.
Continuous cumulative TyG, METS-IR, C-reactive protein–triglyceride–glucose index, atherogenic index of plasma, cholesterol–high-density lipoprotein–glucose index, and remnant cholesterol were not significantly associated with incident CVD in the fully adjusted models. The corresponding dose–response analyses are shown in Supplementary Fig. S4.
Analytical sample sizes ranged from 1,739 to 2,059 across the cumulative indices. The apparent C-indices of the index-specific models were 0.647 for cumulative eGDR, 0.632 for cumulative TyG-WC, and 0.628 for cumulative TyG-WHtR; those for the remaining indices ranged from 0.618 to 0.623 (Table 3). Because the analytical samples were not identical across indices, between-index comparisons were primarily descriptive.
Table 3.
Comparative associations and apparent discrimination of cumulative insulin-resistance and cardiometabolic indices
| Cumulative index | Class | N/CVD events | Per 1-SD change | Q4 vs Q1 | P for trend | C-index | ||
|---|---|---|---|---|---|---|---|---|
| HR (95%CI) | P | HR (95%CI) | P | |||||
| Cumulative eGDR | Insulin resistance | 2,059/457 | 1.39 (1.26, 1.55) | < 0.001 | 2.40 (1.77, 3.25) | < 0.001 | < 0.001 | 0.647 |
| Cumulative TyG | Insulin resistance | 2,053/456 | 1.06 (0.94, 1.20) | 0.343 | 1.13 (0.82, 1.56) | 0.459 | 0.483 | 0.623 |
| Cumulative TyG-WC | Insulin resistance | 2,053/456 | 1.26 (1.12, 1.41) | < 0.001 | 1.76 (1.28, 2.42) | < 0.001 | < 0.001 | 0.632 |
| Cumulative TyG-WHtR | Insulin resistance | 2,042/454 | 1.13 (1.04, 1.23) | 0.004 | 1.61 (1.19, 2.19) | 0.002 | 0.007 | 0.628 |
| Cumulative METS-IR | Insulin resistance | 1,739/400 | 1.03 (0.95, 1.12) | 0.534 | 1.53 (1.08, 2.18) | 0.017 | 0.024 | 0.618 |
| Cumulative CTI | Inflammation-metabolism | 2,053/456 | 1.10 (0.98, 1.24) | 0.098 | 1.24 (0.89, 1.71) | 0.204 | 0.410 | 0.623 |
| Cumulative AIP | Residual lipid risk | 2,053/456 | 1.06 (0.91, 1.22) | 0.471 | 1.14 (0.78, 1.67) | 0.495 | 0.927 | 0.623 |
| Cumulative CHG | Residual lipid risk | 2,053/456 | 1.04 (0.89, 1.22) | 0.595 | 1.16 (0.76, 1.76) | 0.504 | 0.500 | 0.623 |
| Cumulative Residual cholesterol | Residual lipid risk | 2,047/454 | 1.04 (0.94, 1.15) | 0.477 | 1.29 (0.94, 1.77) | 0.110 | 0.053 | 0.622 |
Associations are from the fully adjusted Model 3. For cumulative eGDR, the continuous association is expressed per 1-SD lower value; for the other cumulative indices, associations are expressed per 1-SD increase. Quartile comparisons are shown as Q4 versus Q1
The C-index represents apparent discrimination from the index-specific Cox model. N and event counts are retained because the available analytic sample may differ across indices
Comparisons are descriptive; no claim of superiority should be made without a formal paired comparison in a strict common sample
Detailed 3- and 5-year AUC, Brier score, reclassification, and incremental model comparisons are reported in Supplementary Table S9
Incremental predictive performance
The common prediction sample included 2,042 participants. The apparent C-index over the full follow-up period was 0.622 (95% CI 0.594–0.649) for Model 3, 0.647 (0.620–0.674) after adding cumulative eGDR, and 0.645 (0.618–0.672) after adding the two-wave eGDR level–change patterns.
The internally validated 3-year out-of-bag AUCs for Model 3, the model including cumulative eGDR, and the model including the two-wave eGDR level–change patterns were 0.576, 0.608, and 0.603, respectively. The corresponding Brier scores were 0.131, 0.129, and 0.130. Compared with Model 3, adding cumulative eGDR increased the paired out-of-bag C-index by 0.033 (95% CI 0.009–0.056; P = 0.006). The corresponding ΔC-indices for adding cumulative eGDR to models already containing baseline eGDR, 2015 eGDR, or the 2011 eGDR components were 0.011 (95% CI − 0.002 to 0.023; P = 0.087), −0.001 (− 0.007 to 0.006; P = 0.867), and 0.010 (− 0.002 to 0.022; P = 0.105), respectively (Supplementary Table S9 and Supplementary Fig. S5).
In the full analytical cohort of 2,059 participants, the apparent C-indices were 0.638 (95% CI 0.610–0.665) for Model 3 + baseline eGDR, 0.648 (0.620–0.674) for Model 3 + 2015 eGDR, 0.647 (0.619–0.674) for Model 3 + cumulative eGDR, and 0.648 (0.621–0.675) for Model 3 + 2015 eGDR + cumulative eGDR (Supplementary Table S9).
The apparent 5-year time-dependent receiver operating characteristic curves, calibration curves, and decision curves are shown in Fig. 5. The corresponding net reclassification improvement, integrated discrimination improvement, and detailed incremental model comparisons are presented in Supplementary Table S9.
Fig. 5.
Apparent 5-year predictive performance of models incorporating eGDR information. A Five-year time-dependent receiver operating characteristic curves. B Calibration of predicted and observed 5-year CVD risk. For each model, participants were grouped into deciles of the multiple-imputation-averaged predicted 5-year risk. Within each decile, the mean predicted risk was plotted against the observed 5-year CVD risk estimated as 1 minus the Kaplan–Meier survival estimate. C Decision-curve analysis. AUC, area under the curve; CVD, cardiovascular disease; eGDR, estimated glucose disposal rate; ROC, receiver operating characteristic. Panels A–C show multiple-imputation-averaged apparent 5-year performance and are not internally validated estimates. Internally validated performance and paired out-of-bag C-index comparisons are reported in Supplementary Table S9 and Figure S5
Subgroup and MASLD course-related analyses
The association of lower cumulative eGDR with higher incident CVD risk was generally consistent across subgroups defined by age, sex, diabetes, hypertension, obesity, and MASLD course. The formal P values for interaction were 0.179 for age, 0.753 for sex, 0.332 for diabetes, 0.893 for hypertension, 0.237 for obesity, and 0.752 for MASLD course; none remained statistically significant after Benjamini–Hochberg correction (Supplementary Table S10 and Supplementary Fig. S6).
In the joint analysis of MASLD course and the binary eGDR pattern classification, the P values for interaction were 0.752 for cumulative eGDR × MASLD course and 0.211 for binary eGDR pattern × MASLD course. Detailed results are presented in Supplementary Table S7.
Sensitivity analyses and model diagnostics
The primary associations remained stable across the sensitivity analyses. The HRs per 1-SD lower cumulative eGDR were 1.41 (95% CI 1.17–1.70) after excluding participants who first reported CVD in 2018, 1.29 (1.15–1.45) when restricting the analysis to participants with persistent MASLD according to the operational definition, 1.34 (1.19–1.51) after excluding participants with diabetes at baseline, 1.37 (1.23–1.53) after excluding users of glucose-lowering medication, and 1.41 (1.27–1.57) in the complete-case analysis (Supplementary Table S11).
Results were also robust to alternative handling of CVD event times. The fully adjusted HR per 1-SD lower cumulative eGDR was 1.39 (95% CI 1.25–1.55) when the midpoint of the ascertainment interval was assigned as the event time and 1.40 (95% CI 1.25–1.56) in the interval-censored proportional-hazards model, compared with 1.39 (95% CI 1.26–1.55) in the primary Cox analysis (Supplementary Table S14).
In Fine–Gray models treating 66 deaths before CVD as competing events, the subdistribution HR per 1-SD lower cumulative eGDR was 1.35 (95% CI 1.23–1.49). Compared with the high eGDR–slight decline pattern, the subdistribution HRs were 1.38 (1.05–1.82) for the moderate-low eGDR–slight rise pattern, 1.57 (1.18–2.10) for the high eGDR–marked decline pattern, and 2.01 (1.59–2.54) for the low eGDR–slight decline pattern.
The direction of the results remained consistent when K-means clustering was repeated using eGDR values standardised within each survey wave and when level–change classifications were defined using sex- and wave-specific medians (Supplementary Table S11). Alternative clustering solutions also produced directionally consistent findings (Supplementary Table S12). The E-values corresponding to the primary point estimate and the lower bound of its 95% CI were 2.14 and 1.82, respectively.
Tests based on Schoenfeld residuals did not indicate violation of the proportional hazards assumption for cumulative eGDR, the eGDR pattern exposure, or the model globally, although physical activity showed evidence of non-proportional hazards. After stratifying the baseline hazard by physical activity, the HR per 1-SD lower cumulative eGDR was 1.40 (95% CI 1.26–1.55), and the HRs for the three non-reference patterns were 1.43, 1.64, and 2.14, respectively, which were essentially unchanged from the primary analysis.
As expected, moderate collinearity was observed in mutually adjusted models containing either 2011 eGDR or the individual eGDR components together with cumulative eGDR, whereas the primary cumulative-eGDR model showed acceptable collinearity diagnostics (Supplementary Table S13). After HDL cholesterol was removed from Model 3, the association for cumulative METS-IR remained similar to that in the original model (HR 1.04, 95% CI 0.96–1.12; P = 0.357; Supplementary Table S8).
Discussion
Principal findings
Among middle-aged and older Chinese adults meeting the operational TyG-WC–based MASLD definition, lower cumulative eGDR during 2011–2015, which is mathematically proportional to a lower mean of the 2011 and 2015 eGDR measurements, was associated with a higher risk of incident CVD after 2015, with an approximately linear dose–response relationship. The exploratory longitudinal eGDR patterns, operationally characterised as two-wave eGDR level–change patterns, further showed that participants with low eGDR at both observed time points had the highest risk, while those with a marked decline from a higher baseline level also had increased risk; by contrast, simple change in eGDR provided weaker risk information. Associations with heart disease and stroke were directionally consistent, and the main findings remained robust across competing-risk, complete-case, early-event exclusion, diabetes-related, physical-activity-stratified, and alternative event-time analyses. Adding cumulative eGDR to the conventional risk-factor model modestly improved internally validated discrimination, whereas little additional discrimination was observed once the 2015 eGDR measurement was already included. Taken together, these findings suggest that repeated eGDR measurements can provide information on the observed level–change context, but much of the incremental discriminatory information captured by cumulative eGDR overlaps with the most recent eGDR measurement.
Comparison with previous studies and interpretation of the findings
Previous studies in the general population and among individuals without diabetes have consistently associated lower eGDR with a higher risk of incident CVD [35, 36]. Similar associations with cardiovascular mortality and major cardiovascular events have been observed across different states of glucose tolerance and among patients with non-diabetic chronic kidney disease [37, 38]. Among individuals with diabetes, lower eGDR has also been associated with atrial fibrillation, heart failure, and cardiovascular mortality, while a recent meta-analysis confirmed inverse associations with CVD, coronary heart disease, and stroke [39, 40]. More recently, research has moved beyond single eGDR measurements to cumulative exposure and longitudinal change. In a study using repeated eGDR data from the UK Biobank and CHARLS, Zheng et al. reported that unfavourable eGDR patterns and lower cumulative eGDR were associated with a higher risk of incident CVD [16]. In a separate analysis of CHARLS and the Health and Retirement Study, Zhang et al. found a linear inverse association between cumulative eGDR and incident CVD among adults aged over 50 years without diabetes [18]. Similar findings have subsequently been reported in CHARLS participants across the cardiovascular–kidney–metabolic spectrum, among whom persistently low or cumulatively low eGDR was associated with incident CVD or stroke [24, 41]. Our findings are consistent with this evidence and provide complementary evidence among individuals meeting an operational TyG-WC–based MASLD definition. In the 2011 supportive cohort, a single low baseline eGDR was already associated with a clear gradient in long-term CVD risk, indicating that one eGDR measurement can capture important cardiometabolic risk. The primary landmark analysis further showed that repeated measurements helped characterise whether eGDR remained low across the two observed waves or changed substantially between the two assessments.
Each 1-SD lower cumulative eGDR was associated with a 39% higher risk of CVD, and participants with the lowest cumulative eGDR (Q4) had approximately 2.4 times the risk of those with the highest cumulative eGDR (Q1), without evidence of a distinct nonlinear threshold. Given that cumulative eGDR in the present study is proportional to the mean of the two observed measurements, this approximately linear association indicates a graded relationship between lower average eGDR across the two waves and subsequent CVD risk rather than a distinct risk threshold. eGDR integrates waist circumference, hypertension status, and HbA1c, which reflect central adiposity, haemodynamic load, and longer-term glycaemic exposure, respectively. Central fat accumulation may increase the release of free fatty acids and ectopic lipid deposition, promote the production of triglyceride-rich lipoproteins, and contribute to adipose tissue inflammation and adipokine imbalance [42, 43]. Sustained hypertension increases vascular and myocardial load and contributes to arterial remodelling, stiffness, and endothelial injury [44, 45]. Chronic glycaemic exposure may further promote vascular injury through advanced glycation end products, oxidative stress, and endothelial dysfunction [46, 47]. The coexistence of these abnormalities may create an adverse metabolic and vascular environment, providing a plausible biological context for the observed gradient between lower cumulative eGDR and CVD risk.
This association may be particularly relevant in the cardiometabolic context of individuals meeting the operational TyG-WC–based MASLD definition. Insulin resistance is fundamental to the development and progression of MASLD and is a central link between hepatic fat accumulation and cardiovascular injury [48, 49]. Selective hepatic insulin resistance may impair insulin-mediated suppression of gluconeogenesis while relatively preserving lipogenic signalling, resulting in the coexistence of increased hepatic glucose production, triglyceride synthesis, and secretion of triglyceride-rich lipoproteins [50, 51]. Consequently, even among individuals meeting the operational TyG-WC–based MASLD definition, substantial heterogeneity may remain in central adiposity, blood pressure burden, glycaemic dysfunction, and estimated insulin sensitivity. The observed association between lower cumulative eGDR and CVD may therefore reflect variation in this broader cardiometabolic profile. Because waist circumference contributes to both the TyG-WC–based MASLD definition and the eGDR calculation, this association should not be interpreted as evidence of an adiposity-independent effect of insulin resistance.
Risk implications of longitudinal eGDR patterns
The risk ordering of the four longitudinal eGDR patterns suggests that absolute eGDR level and the level–change profiles represented by the cluster centroids convey distinct but interrelated information. The low eGDR–slight decline pattern had the lowest eGDR levels in both 2011 and 2015 and the highest event proportion and adjusted CVD risk, suggesting that an adverse cardiometabolic profile present at both observed waves was associated with greater subsequent CVD risk. Although the high eGDR–marked decline pattern began at a relatively favourable level, its eGDR had fallen markedly by 2015, and its CVD risk was 64% higher than that of the reference pattern. This finding suggests that a marked decline between the two assessments may identify higher subsequent CVD risk even among participants with a relatively favourable initial eGDR level. Because eGDR jointly reflects waist circumference, hypertension status, and HbA1c, this pattern may reflect concurrent worsening across multiple cardiometabolic domains, highlighting the prognostic importance of transition from a relatively favourable state to a high-risk cardiometabolic phenotype.
The moderate-low eGDR–slight rise pattern, whose cluster centroid increased slightly, remained at higher risk than the high eGDR–slight decline pattern. This centroid-level increase should not be interpreted as improvement in every participant, because individual changes within the cluster were heterogeneous. The mean eGDR in 2015 remained substantially lower than that of the reference group, and the relatively low absolute eGDR level may therefore have contributed to the higher observed CVD risk. The weaker association between simple eGDR change during 2011–2015 and CVD further indicates that the magnitude of change should be interpreted in relation to both the initial and final levels. An equivalent change may occur in very different metabolic contexts, such as a slight increase from an extremely low level or a marked decline from a high level, and therefore carry different risk implications. Cumulative eGDR, the most recent measurement, and two-wave level–change patterns therefore provide related but not identical information. In the present study, cumulative eGDR primarily reflects the average eGDR level across the two observed waves, landmark eGDR reflects cardiometabolic status at the 2015 assessment, and the level–change patterns retain information on the combination of initial level, final level, and direction of change. Together, these findings suggest that both a low eGDR level at the observed assessments and a marked decline between assessments may identify individuals at higher subsequent CVD risk.
Heart disease, stroke, and other metabolic indices
The directionally consistent associations of lower cumulative eGDR and adverse two-wave eGDR level–change patterns with heart disease and stroke are consistent with shared cardiometabolic risk processes underlying these outcomes. Insulin resistance-related endothelial dysfunction, coronary atherosclerosis, abnormal myocardial substrate utilisation, and ventricular remodelling may contribute to the development of heart disease [52, 53]. Hypertension, arterial stiffness, cerebral vascular injury, atherosclerosis, and thrombosis are important pathological pathways underlying stroke [44, 45]. Because eGDR integrates central adiposity, blood pressure status, and longer-term glycaemic exposure, its associations with both outcomes further support its role as a composite marker of cardiometabolic risk.
Among the other cumulative metabolic indices, cumulative TyG-WC and TyG-WHtR were positively associated with CVD risk, suggesting that indices incorporating central adiposity may capture relevant cardiometabolic risk information in this operationally defined MASLD population [54, 55]. The TyG index primarily captures metabolic information derived from fasting triglyceride and glucose concentrations, whereas combining TyG with waist circumference or waist-to-height ratio additionally reflects central fat distribution. Previous comparative studies have shown that TyG-derived indices incorporating central adiposity generally provide stronger cardiovascular risk information than the TyG index alone [56]. eGDR differs from these indices by incorporating waist circumference, hypertension status, and HbA1c, and previous studies have reported favourable predictive performance compared with several other insulin resistance surrogates [57]. In the present study, cumulative eGDR had the numerically highest apparent C-index among the index-specific models. However, because the availability of repeated measurements and the resulting analytical sample sizes differed across indices, these between-index comparisons were descriptive and do not establish the superiority of eGDR over the other metabolic indices.
Incremental predictive value and clinical implications
Previous cohort studies have shown that lower eGDR is associated with incident CVD across cardiovascular–kidney–metabolic stages and with a broad range of cardiovascular outcomes, including myocardial infarction, heart failure, atrial fibrillation, and ischaemic stroke [58, 59]. Building on this evidence, our prediction analyses showed that adding cumulative eGDR to the conventional risk-factor model increased the paired out-of-bag C-index by 0.033 and improved the 3-year out-of-bag AUC, suggesting modest incremental discrimination beyond the conventional risk-factor model. However, when 2015 eGDR was already included, adding cumulative eGDR did not further improve the paired out-of-bag C-index (ΔC-index −0.001, 95% CI −0.007 to 0.006), indicating substantial overlap in discrimination between the two-wave summary measure and the most recent eGDR measurement. Thus, the present findings do not support superior short-term discrimination from cumulative eGDR beyond a recent eGDR measurement alone. Repeated measurements may nevertheless provide descriptive information on whether eGDR was low at both observed waves or changed markedly between the two assessments. Accordingly, landmark eGDR, cumulative eGDR, and the two-wave level–change patterns should be viewed as closely related measures that emphasise current cardiometabolic status, average eGDR across the two observed waves, and the context of change between assessments, respectively.
From an applied perspective, eGDR can be calculated from routinely available waist circumference, hypertension status, and HbA1c without fasting insulin measurement, making it practical to obtain when these measures are already collected in clinical or population-based settings. Among individuals meeting the operational TyG-WC–based MASLD definition, a low eGDR level at both observed waves or a marked decline between assessments may identify individuals with higher subsequent CVD risk and may prompt closer evaluation of body weight, blood pressure, glycaemic status, and other cardiovascular risk factors. Given the continuous and approximately linear risk gradient observed in this study, eGDR may be most appropriately interpreted as a continuous cardiometabolic marker alongside age, smoking, lipid levels, kidney function, and other conventional risk factors. However, the present study did not establish an optimal interval for repeated eGDR assessment or demonstrate that repeated measurement improves clinical outcomes; its potential role in longitudinal risk monitoring therefore requires further external validation.
Strengths and limitations
The principal strengths of this study include the use of 2015 as the landmark, ensuring that both the 2011 and 2015 eGDR measurements preceded outcome follow-up and clearly separating the exposure-assessment period from the outcome period.
Cumulative eGDR and the longitudinal eGDR patterns were evaluated together with baseline eGDR, landmark eGDR, and simple change within the same analytical framework, allowing complementary assessment of average eGDR level and change between the two observed waves. Continuous, quartile-based, and restricted cubic spline analyses provided consistent evidence, and the exploratory pattern classification was evaluated using resampling stability, clustering based on standardised values, and alternative clustering solutions. The findings also remained stable in the 2011 supportive cohort, competing-risk and complete-case analyses, analyses using alternative handling of CVD event times, after stratifying the baseline hazard by physical activity, after excluding early events, when restricting the analysis to persistent MASLD, and after excluding participants with diabetes or glucose-lowering medication use.
Several limitations should nevertheless be considered. First, MASLD was identified using an operational TyG-WC–based definition rather than imaging or histology, which may have introduced misclassification and did not permit assessment of fibrosis or inflammatory severity. Waist circumference also contributes to both the TyG-WC–based definition and the eGDR calculation. In addition, eGDR is a composite surrogate marker derived from waist circumference, hypertension, and HbA1c rather than a direct measure of insulin sensitivity; therefore, the observed associations should be interpreted as reflecting the prognostic information of a composite cardiometabolic marker rather than an adiposity- or component-independent effect of insulin resistance. Second, eGDR was measured only in 2011 and 2015. Accordingly, cumulative eGDR is mathematically proportional to the mean of the two measurements, and the K-means groups represent two-wave level–change patterns rather than formal long-term trajectories; intermediate fluctuations and exposure duration could not be assessed. CVD outcomes were based on self-reported physician diagnoses, and exact onset dates were unavailable, although the results were essentially unchanged in midpoint and interval-censored sensitivity analyses. Third, the landmark design required participants to survive without reported CVD to 2015 and to have repeated eGDR measurements. Differences between included and excluded participants indicate possible survivor and selection bias, and multiple imputation could not address exclusions caused by unavailable exposure or eligibility data. Moreover, although CHARLS used a multistage probability sampling design, the present analyses did not incorporate survey sampling weights, limiting the national generalisability of the selected analytical cohort. Finally, all covariates in the primary association models were assessed at the 2011 baseline, and changes in potential confounders during 2011–2015 were not incorporated. Residual confounding cannot therefore be excluded, and the predictive findings require further validation in independent cohorts with clinically or imaging-confirmed steatotic liver disease, more frequent repeated measurements, and adjudicated cardiovascular outcomes.
Conclusions
Among middle-aged and older Chinese adults meeting an operational TyG-WC–based MASLD definition, lower cumulative eGDR and longitudinal eGDR patterns characterised by low levels at both observed waves or a marked decline were associated with a higher risk of incident CVD. Cumulative eGDR primarily reflects the average eGDR level across the two observed waves, whereas landmark eGDR reflects more recent cardiometabolic status and the level–change patterns provide descriptive information on the context of change between assessments. Although cumulative eGDR modestly improved discrimination beyond the conventional risk-factor model, it provided essentially no additional discrimination once 2015 eGDR was already included, suggesting that repeated measurements may be most useful for characterising longitudinal context rather than providing substantial incremental short-term predictive information beyond a recent eGDR measurement.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
This study utilized data from CHARLS databases. The authors thank the CHARLS research team and all participants involved in the study.
Abbreviations
- AIC
Akaike information criterion
- AIP
Atherogenic index of plasma
- AUC
Area under the curve
- CHARLS
China Health and Retirement Longitudinal Study
- CI
Confidence interval
- CKD
Chronic kidney disease
- CTI
C-reactive protein–triglyceride–glucose index
- CVD
Cardiovascular disease
- eGDR
Estimated glucose disposal rate
- eGFR
Estimated glomerular filtration rate
- FPG
Fasting blood glucose
- HbA1c
Glycated haemoglobin
- HDL-C
High-density lipoprotein cholesterol
- LDL-C
Low-density lipoprotein cholesterol
- HR
Hazard ratio
- MASLD
Metabolic dysfunction-associated steatotic liver disease
- METS-IR
Metabolic score for insulin resistance
- TG
Triglycerides
- TYG
Triglyceride–glucose index
- TYG-WC
Triglyceride–glucose waist circumference index
- TYG-WHtR
Triglyceride–glucose waist-to-height ratio index
- WC
Waist circumference
Author contributions
YL conceived and designed the study, and YL, HG and DW analyzed the data. LR, XW, KN and XT have put forward key revision suggestions for the manuscript. GD and WG are responsible for manuscript review. All authors have reviewed and approved the final manuscript.
Funding
This work was supported by the National Natural Science Foundation of China (82204942 and 82174172), the Shandong Provincial Natural Science Foundation (ZR2022QH123), and the Shandong Provincial Health Commission Science and Technology Innovation Team Development Project.
Data availability
The data supporting the results of this study can be found on CHARLS (http://charls.pku.edu.cn).
Declarations
Ethics approval and consent to participate
The CHARLS study protocol was approved by the Ethics Committee of Peking University (IRB: 00001052–11015). All participants provided written informed consent after being fully informed prior to enrollment. The research process strictly adhered to the ethical principles outlined in the Declaration of Helsinki, and the final results reporting complied with the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Contributor Information
Guohua Dai, Email: daigh2004@163.com.
Wulin Gao, Email: gaowulin2025@163.com.
References
- 1.Rinella ME, Lazarus JV, Ratziu V, Francque SM, Sanyal AJ, Kanwal F, et al. A multisociety Delphi consensus statement on new fatty liver disease nomenclature. J Hepatol. 2023;79(6):1542–56. [DOI] [PubMed] [Google Scholar]
- 2.GBD 2023 MASLD Collaborators. Global burden of metabolic dysfunction-associated steatotic liver disease, 1990–2023, and projections to 2050: a systematic analysis for the Global Burden of Disease Study 2023. Lancet Gastroenterol Hepatol. 2026;11(6):463–94. [DOI] [PubMed] [Google Scholar]
- 3.Chai S, Zhang R, Fernandes G, Zheng Y, Wei L. Epidemiology of metabolic dysfunction-associated steatotic liver disease/metabolic dysfunction-associated steatohepatitis and associated cardiometabolic factors in adults in China (2013–2023): a systematic review and meta-analysis. World J Gastroenterol. 2025;31(46):113608. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Global Burden of Cardiovascular Diseases and Risks 2023 Collaborators. Global, Regional, and National Burden of Cardiovascular Diseases and Risk Factors in 204 Countries and Territories, 1990–2023. J Am Coll Cardiol. 2025;86(22):2167–243. [DOI] [PubMed] [Google Scholar]
- 5.Tilg H, Petta S, Stefan N, Targher G. Metabolic dysfunction-associated steatotic liver disease in adults: a review. JAMA J Am Med Assoc. 2026;335(2):163–74. [DOI] [PubMed] [Google Scholar]
- 6.Lee H, Lee HA, Kim E, Kim HY, Kim HC, Ahn SH, et al. Metabolic dysfunction-associated steatotic liver disease and risk of cardiovascular disease. Gut. 2024;73(3):533–40. [DOI] [PubMed] [Google Scholar]
- 7.Mantovani A, Csermely A, Petracca G, Beatrice G, Corey KE, Simon TG, et al. Non-alcoholic fatty liver disease and risk of fatal and non-fatal cardiovascular events: an updated systematic review and meta-analysis. Lancet Gastroenterol Hepatol. 2021;6(11):903–13. [DOI] [PubMed] [Google Scholar]
- 8.European Association for the Study of the Liver (EASL), European Association for the Study of Diabetes (EASD), European Association for the Study of Obesity (EASO). EASL-EASD-EASO clinical practice guidelines on the management of metabolic dysfunction-associated steatotic liver disease (MASLD). J Hepatol. 2024;81(3):492–542. [DOI] [PubMed] [Google Scholar]
- 9.DeFronzo RA, Tobin JD, Andres R. Glucose clamp technique: a method for quantifying insulin secretion and resistance. Am J Physiol. 1979;237(3):E214–23. [DOI] [PubMed] [Google Scholar]
- 10.Matthews DR, Hosker JP, Rudenski AS, Naylor BA, Treacher DF, Turner RC. Homeostasis model assessment: insulin resistance and beta-cell function from fasting plasma glucose and insulin concentrations in man. Diabetologia. 1985;28(7):412–9. [DOI] [PubMed] [Google Scholar]
- 11.Williams KV, Erbey JR, Becker D, Arslanian S, Orchard TJ. Can clinical factors estimate insulin resistance in type 1 diabetes? Diabetes. 2000;49(4):626–32. [DOI] [PubMed] [Google Scholar]
- 12.Epstein EJ, Osman JL, Cohen HW, Rajpathak SN, Lewis O, Crandall JP. Use of the estimated glucose disposal rate as a measure of insulin resistance in an urban multiethnic population with type 1 diabetes. Diabetes Care. 2013;36(8):2280–5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Zhang Z, Zhao L, Lu Y, Xiao Y, Zhou X. Insulin resistance assessed by estimated glucose disposal rate and risk of incident cardiovascular diseases among individuals without diabetes: findings from a nationwide, population based, prospective cohort study. Cardiovasc Diabetol. 2024;23(1):194. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Zabala A, Darsalia V, Lind M, Svensson A, Franzén S, Eliasson B, et al. Estimated glucose disposal rate and risk of stroke and mortality in type 2 diabetes: a nationwide cohort study. Cardiovasc Diabetol. 2021;20(1):202. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Guo L, Zhang J, An R, Wang W, Fen J, Wu Y, et al. The role of estimated glucose disposal rate in predicting cardiovascular risk among general and diabetes mellitus population: a systematic review and meta-analysis. BMC Med. 2025;23(1):234. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Zheng X, Han W, Li Y, Jiang M, Ren X, Yang P, et al. Changes in the estimated glucose disposal rate and incident cardiovascular disease: two large prospective cohorts in Europe and Asia. Cardiovasc Diabetol. 2024;23(1):403. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Yan L, Zhou Z, Wu X, Qiu Y, Liu Z, Luo L, et al. Association between the changes in the estimated glucose disposal rate and new-onset cardiovascular disease in middle-aged and elderly individuals: a nationwide prospective cohort study in China. Diabetes Obes Metab. 2025;27(4):1859–67. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Zhang J, Sun Z, Li Y, Yang Y, Liu W, Huang M, et al. Association between the cumulative estimated glucose disposal rate and incident cardiovascular disease in individuals over the age of 50 years and without diabetes: data from two large cohorts in China and the United States. Cardiovasc Diabetol. 2025;24(1):51. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Qiao Y, Wang Y, Chen C, Huang Y, Zhao C. Association between triglyceride-glucose (TyG) related indices and cardiovascular diseases and mortality among individuals with metabolic dysfunction-associated steatotic liver disease: a cohort study of UK Biobank. Cardiovasc Diabetol. 2025;24(1):12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Zhao Y, Hu Y, Smith JP, Strauss J, Yang G. Cohort profile: the China Health and Retirement Longitudinal Study (CHARLS). Int J Epidemiol. 2014;43(1):61–8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Chen X, Crimmins E, Hu PP, Kim JK, Meng Q, Strauss J, et al. Venous blood-based biomarkers in the China health and retirement longitudinal study: rationale, design, and results from the 2015 wave. Am J Epidemiol. 2019;188(11):1871–7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Jin X, Su Z, Zhong J, Wang J, Shi B, Liu Z, et al. Association of the C-reactive protein-triglyceride-glucose index with metabolic dysfunction-associated steatotic liver disease and long-term all-cause and cardiovascular mortality: evidence from two nationwide prospective cohort studies. Cardiovasc Diabetol. 2026;25(1):75. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Guerrero-Romero F, Simental-Mendía LE, González-Ortiz M, Martínez-Abundis E, Ramos-Zavala MG, Hernández-González SO, et al. The product of triglycerides and glucose, a simple measure of insulin sensitivity. Comparison with the euglycemic-hyperinsulinemic clamp. J Clin Endocrinol Metab. 2010;95(7):3347–51. [DOI] [PubMed] [Google Scholar]
- 24.Wang Y, Wei N, Li M, Liu J, Lin H, Zhang H. Cumulative exposure to the estimated glucose disposal rate and incident stroke in individuals with cardiovascular-kidney-metabolic syndrome stages 0–4: 6-year longitudinal evidence from CHARLS. Cardiovasc Diabetol. 2026;25(1):78. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.White IR, Royston P, Wood AM. Multiple imputation using chained equations: Issues and guidance for practice. Stat Med. 2011;30(4):377–99. [DOI] [PubMed] [Google Scholar]
- 26.Bello-Chavolla OY, Almeda-Valdes P, Gomez-Velasco D, Viveros-Ruiz T, Cruz-Bautista I, Romo-Romo A, et al. METS-IR, a novel score to evaluate insulin sensitivity, is predictive of visceral adiposity and incident type 2 diabetes. Eur J Endocrinol. 2018;178(5):533–44. [DOI] [PubMed] [Google Scholar]
- 27.Dobiásová M, Frohlich J. The plasma parameter log (TG/HDL-C) as an atherogenic index: correlation with lipoprotein particle size and esterification rate in apoB-lipoprotein-depleted plasma (FER(HDL)). Clin Biochem. 2001;34(7):583–8. [DOI] [PubMed] [Google Scholar]
- 28.Mo D, Zhang P, Zhang M, Dai H, Guan J. Cholesterol, high-density lipoprotein, and glucose index versus triglyceride-glucose index in predicting cardiovascular disease risk: a cohort study. Cardiovasc Diabetol. 2025;24(1):116. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Lin Y, Tao J, Wang H, Guan H, Liu X, Dong X, et al. The association of C-reactive protein-triglyceride-glucose index with cardiometabolic multimorbidity in middle-aged and older adults: evidence from two cohort studies. Cardiovasc Diabetol. 2026;25(1):90. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Pencina MJ, D’Agostino RBS, D’Agostino RBJ, Vasan RS. Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Stat Med. 2008;27(2):157–212. [DOI] [PubMed] [Google Scholar]
- 31.Heagerty PJ, Lumley T, Pepe MS. Time-dependent ROC curves for censored survival data and a diagnostic marker. Biometrics. 2000;56(2):337–44. [DOI] [PubMed] [Google Scholar]
- 32.Vickers AJ, Elkin EB. Decision curve analysis: a novel method for evaluating prediction models. Med Decis Mak. 2006;26(6):565–74. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Austin PC, Fine JP. Practical recommendations for reporting Fine-Gray model analyses for competing risk data. Stat Med. 2017;36(27):4391–400. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34.VanderWeele TJ, Ding P. Sensitivity analysis in observational research: introducing the E-value. Ann Intern Med. 2017;167(4):268–74. [DOI] [PubMed] [Google Scholar]
- 35.Tao S, Yu L, Li J, Wu J, Huang X, Xie Z, et al. Insulin resistance quantified by estimated glucose disposal rate predicts cardiovascular disease incidence: a nationwide prospective cohort study. Cardiovasc Diabetol. 2025;24(1):161. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Kong X, Wang W. Estimated glucose disposal rate and risk of cardiovascular disease and mortality in U.S. adults with prediabetes: a nationwide cross-sectional and prospective cohort study. Acta Diabetol. 2024;61(11):1413–21. [DOI] [PubMed] [Google Scholar]
- 37.Guo R, Tong J, Cao Y, Zhao W. Association between estimated glucose disposal rate and cardiovascular mortality across the spectrum of glucose tolerance in the US population. Diabetes Obes Metab. 2024;26(12):5827–35. [DOI] [PubMed] [Google Scholar]
- 38.Peng J, Zhang Y, Zhu Y, Chen W, Chen L, Ma F, et al. Estimated glucose disposal rate for predicting cardiovascular events and mortality in patients with non-diabetic chronic kidney disease: a prospective cohort study. BMC Med. 2024;22(1):411. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 39.Tan Z, Liu Y, Liu L, Li S, Xue X, Li X, et al. Association of estimated glucose disposal rate with atrial fibrillation, heart failure and cardiovascular mortality in patients with diabetes: a prospective cohort study from the UK Biobank. Front Endocrinol. 2025;16:1579836. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 40.Zhang Z, Wang B, Qu L, Huang B. Estimated glucose disposal rate and cardiovascular disease risk: a meta-analysis of cohort studies. Front Endocrinol. 2025;16:1740472. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41.Liu J, Xu W, Zhang Y, Jia Y, Bai S, Er L, et al. Changes in the estimated glucose disposal rate and incident cardiovascular disease in patients with cardiovascular-kidney-metabolic syndrome stages 0–3: a prospective cohort study in China. BMC Cardiovasc Disord. 2025;25(1):889. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 42.Neeland IJ, Ross R, Després J, Matsuzawa Y, Yamashita S, Shai I, et al. Visceral and ectopic fat, atherosclerosis, and cardiometabolic disease: a position statement. Lancet Diabetes Endocrinol. 2019;7(9):715–25. [DOI] [PubMed] [Google Scholar]
- 43.Rana MN, Neeland IJ. Adipose tissue inflammation and cardiovascular disease: an update. Curr Diabetes Rep. 2022;22(1):27–37. [DOI] [PubMed] [Google Scholar]
- 44.Powell-Wiley TM, Poirier P, Burke LE, Després J, Gordon-Larsen P, Lavie CJ, et al. Obesity and cardiovascular disease: a scientific statement from the American Heart Association. Circulation. 2021;143(21):e984-1010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45.Petrie JR, Guzik TJ, Touyz RM. Diabetes, hypertension, and cardiovascular disease: clinical insights and vascular mechanisms. Can J Cardiol. 2018;34(5):575–84. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46.Fiorentino TV, Prioletta A, Zuo P, Folli F. Hyperglycemia-induced oxidative stress and its role in diabetes mellitus related cardiovascular diseases. Curr Pharm Des. 2013;19(32):5695–703. [DOI] [PubMed] [Google Scholar]
- 47.Brownlee M. Biochemistry and molecular cell biology of diabetic complications. Nature. 2001;414(6865):813–20. [DOI] [PubMed] [Google Scholar]
- 48.Targher G, Byrne CD, Tilg H. MASLD: a systemic metabolic disorder with cardiovascular and malignant complications. Gut. 2024;73(4):691–702. [DOI] [PubMed] [Google Scholar]
- 49.Duell PB, Welty FK, Miller M, Chait A, Hammond G, Ahmad Z, et al. Nonalcoholic fatty liver disease and cardiovascular risk: a scientific statement from the American Heart Association. Arterioscler Thromb Vasc Biol. 2022;42(6):e168–85. [DOI] [PubMed] [Google Scholar]
- 50.Petersen MC, Shulman GI. Mechanisms of insulin action and insulin resistance. Physiol Rev. 2018;98(4):2133–223. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Samuel VT, Shulman GI. The pathogenesis of insulin resistance: integrating signaling pathways and substrate flux. J Clin Invest. 2016;126(1):12–22. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 52.Muniyappa R, Sowers JR. Role of insulin resistance in endothelial dysfunction. Rev Endocr Metab Disord. 2013;14(1):5–12. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 53.Wang M, Li Y, Li S, Lv J. Endothelial dysfunction and diabetic cardiomyopathy. Front Endocrinol. 2022;13:851941. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 54.Ren Q, Huang Y, Liu Q, Chu T, Li G, Wu Z. Association between triglyceride glucose-waist height ratio index and cardiovascular disease in middle-aged and older Chinese individuals: a nationwide cohort study. Cardiovasc Diabetol. 2024;23(1):247. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 55.Ahn SH, Lee HS, Lee J. Triglyceride-glucose-waist circumference index predicts the incidence of cardiovascular disease in Korean populations: competing risk analysis of an 18-year prospective study. Eur J Med Res. 2024;29(1):214. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 56.Park H, Han T, Heo S, Kwon Y. Effectiveness of the triglyceride-glucose index and triglyceride-glucose-related indices in predicting cardiovascular disease in middle-aged and older adults: a prospective cohort study. J Clin Lipidol. 2024;18(1):e70–9. [DOI] [PubMed] [Google Scholar]
- 57.Dong B, Chen Y, Yang X, Chen Z, Zhang H, Gao Y, et al. Estimated glucose disposal rate outperforms other insulin resistance surrogates in predicting incident cardiovascular diseases in cardiovascular-kidney-metabolic syndrome stages 0-3 and the development of a machine learning prediction model: a nationwide prospective cohort study. Cardiovasc Diabetol. 2025;24(1):163. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 58.Zhang H, Huang S, Fang Y, Zhang H, Yang W, Yu M. Insulin resistance assessed by estimated glucose disposal rate predicts cardiovascular disease in stages 0–3 of cardiovascular-kidney-metabolic syndrome: a UK biobank cohort study. Cardiovasc Diabetol. 2025;24(1):360. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Huang H, Xiong Y, Zhou J, Tang Y, Chen F, Li G, et al. The predictive value of estimated glucose disposal rate and its association with myocardial infarction, heart failure, atrial fibrillation and ischemic stroke. Diabetes Obes Metab. 2025;27(3):1359–68. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The data supporting the results of this study can be found on CHARLS (http://charls.pku.edu.cn).