Burden trajectories of stroke attributable to lipid-glucose dysfunction and biomarker validation: evidence from the global burden of disease (GBD) 2023 and the China health and retirement longitudinal study (CHARLS) Authors: ZhenKun Xiao, LiangJie Deng, YongHong Duan, Bing Wang, AiHua Liu Published: 2026 ABSTRACT AbstractBackgroundMetabolic dysfunction is an important contributor to stroke burden and may also help identify individuals at higher future risk of stroke. However, the population-level burden attributable to major metabolic risks and the individual-level predictive value of lipid-glucose metabolic biomarkers have not been well characterized within a complementary analytical framework. This study assessed temporal trends in stroke burden attributable to high FPG and high LDL-C in China and compared the prognostic performance of lipid-glucose metabolic biomarkers for incident stroke.MethodsGBD 2023 estimates were used to evaluate stroke burden attributable to high FPG and high LDL-C in China from 1990 to 2023. Individual-level analyses were conducted in CHARLS (2011–2020) as the discovery cohort and HRS (2016–2022) as an external-validation cohort. Network and correlation analyses were used to describe interrelationships among predictors. Cox proportional hazards models were applied to assess associations between biomarkers and incident stroke, with component-aware covariate adjustment and FDR correction. Machine-learning models with SHAP interpretation were used to evaluate predictor contribution. The incremental predictive value of eGDR was further assessed by adding it to conventional clinical risk models, with evaluation of discrimination, calibration, and decision-curve performance.ResultsFrom 1990 to 2023, age-standardized mortality and DALY rates for stroke attributable to high FPG and high LDL-C declined in China, whereas the absolute burden remained substantial. In 2023, high FPG accounted for 216.80 thousand stroke deaths and 4,378.51 thousand DALYs, while high LDL-C accounted for 233.09 thousand deaths and 5,642.72 thousand DALYs. In CHARLS, eGDR showed a consistent inverse association with stroke risk and ranked as the leading predictor in SHAP analyses. Similar patterns were observed in HRS. Adding eGDR to conventional clinical risk models modestly improved discrimination and reclassification in both cohorts, with AUC improvements of 0.017–0.025 in CHARLS and 0.009–0.013 in HRS.ConclusionAge-standardized stroke burden attributable to high FPG and high LDL-C in China has declined, but the absolute burden remains substantial. eGDR showed consistent associations with incident stroke and modest incremental predictive value across CHARLS and HRS, supporting its potential relevance for future stroke risk-stratification research.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12933-026-03269-y.Keywords: Stroke, Machine learning, Lipid-glucose metabolism, Biomarkers, eGDR Introduction Stroke remains the second leading cause of death and the third leading cause of disability worldwide, resulting in approximately 7.3 million deaths annually, with the disease burden continuing to rise [1, 2]. Particularly in low- and middle-income countries, stroke-related deaths and disability-adjusted life years (DALYs) account for over 86% of the global total, exhibiting significant regional inequality [3]. China is one of the countries with the greatest stroke burden worldwide, with a high age-standardized incidence rate and a substantial mortality burden [4]. Owing to population aging and persistent exposure to lifestyle and metabolic risk factors, the absolute numbers of stroke cases, prevalent cases, and deaths in China remain substantial and continue to rise [5]. These trends highlight the need to identify modifiable metabolic pathways that can inform both stroke prevention and stroke risk stratification [6]. Glucose and lipid metabolic abnormalities are central components of cardiometabolic risk and are closely linked to cerebrovascular injury [7]. Insulin resistance, hyperglycemia, and atherogenic dyslipidemia may contribute to stroke through endothelial dysfunction, chronic inflammation, oxidative stress, microvascular injury, blood-brain barrier disruption, plaque formation, and thrombosis [8–14]. Therefore, integrated lipid-glucose metabolic biomarkers may help capture combined vascular risk information that is not fully reflected by single laboratory measurements [15, 16]. Several lipid-glucose metabolism and insulin-resistance surrogate markers, including the triglyceride-glucose index, TyG-related composite indices, the atherogenic index of plasma, the Chinese visceral adiposity index, the lipid accumulation product, the metabolic score for insulin resistance, and estimated glucose disposal rate, have been examined in relation to cardiovascular and cerebrovascular outcomes [17–20]. However, evidence regarding their relative predictive performance remains inconsistent. For example, Liu et al. reported that TyG-BMI was associated with 1-year stroke recurrence risk among patients with acute ischemic stroke and hypertension [21], whereas Dong et al. found that estimated glucose disposal rate outperformed several other insulin-resistance surrogates in predicting cardiovascular outcomes [22]. Because composite biomarkers differ in their required components, clinical accessibility, and overlap with conventional risk factors, their value should be evaluated within a common analytical framework and in comparison with established clinical predictors. At the population level, Global Burden of Disease (GBD) analyses can quantify temporal and demographic patterns in stroke burden attributable to metabolic risk factors [23–25]. By contrast, longitudinal cohort data can evaluate whether related biomarkers identify individuals at higher future risk of stroke. These two sources of evidence therefore address distinct but complementary questions: GBD estimates describe population-level attributable burden, whereas CHARLS and HRS allow individual-level biomarker comparison, risk modelling, and external validation. Therefore, in the present study, we used GBD 2023 estimates to describe temporal trends and age-sex patterns of stroke burden attributable to high fasting plasma glucose and high LDL cholesterol in China. We then used the China Health and Retirement Longitudinal Study (CHARLS) as a discovery cohort to compare the associations and predictive performance of multiple lipid-glucose metabolic biomarkers for incident stroke. Finally, we introduced the Health and Retirement Study (HRS) as an independent external-validation cohort to assess the reproducibility of key findings. The aim was to provide a more cautious and externally supported assessment of lipid-glucose metabolic biomarkers for stroke risk stratification, rather than to establish clinically actionable cutoffs. Method The GBD 2023 study assessed more than 370 diseases and injuries across 204 countries and territories, providing standardized estimates of incidence, mortality, DALYs, and age-standardized rates (ASRs) [25]. The GBD estimates are regularly updated using standardized modelling strategies to improve comparability across time and regions [26]. The CHARLS database is a prospective, nationally representative longitudinal survey designed to reflect the social, economic, and health status of Chinese adults aged 45 years and older [27]. This study used data from the CHARLS survey conducted between 2011 and 2020, with the 2011 survey serving as the baseline. To examine the external validity of the cohort-based findings, we additionally used the Health and Retirement Study (HRS), a nationally representative longitudinal study of middle-aged and older adults in the United States [28]. The HRS 2016 wave was used as the external-validation baseline, and participants were followed through 2022 [29]. Harmonized inclusion and exclusion criteria were applied as far as allowed by data availability. The participant-selection process for CHARLS and HRS is shown in Supplementary Fig. S1. This cohort study has been reported in line with the STROBE guidelines [30]. Data collection and measurement The death count and mortality rate in the GBD data were primarily estimated using the cause-of-death ensemble model, while disease count and prevalence were estimated using Bayesian meta-regression models. Disability weights are defined as indicators for measuring the severity of health loss or non-fatal disability. Years lived with disability (YLD) were calculated by multiplying the number of patients by the duration before remission or death and by the disability weight. Years of life lost (YLL) were calculated by multiplying the number of deaths by the corresponding standard life expectancy derived from a reference life table. DALYs are defined as the total years of health lost from disease onset to death, which is the sum of YLL and YLD [31]. The CHARLS and HRS cohort data extraction focused on five domains: (1) Demographic data, including age and sex; (2) Health behaviors and medical history, including smoking and drinking status, and medical history such as diabetes and hypertension; (3) Physiological indicators, including waist circumference, BMI, systolic blood pressure (SBP), and diastolic blood pressure (DBP); (4) Hematological and biochemical indicators, including total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), glycated hemoglobin (HbA1c), fasting plasma glucose (FPG), C-reactive protein (CRP), white blood cell count (WBC), platelet count (Plt), hemoglobin (HB), blood urea nitrogen (BUN), serum creatinine (SCr), and uric acid (UA); (5) outcome-related indicators, including incident stroke, follow-up time, and all-cause mortality where available. Derived indicators reflecting lipid-glucose metabolism, insulin resistance, adiposity, and frailty included TyG, TyG-BMI, CTI, TyG-WC, TyG-CVAI, TyG-FI, TyG-WHtR, AIP, LAP, LCI, METS-IR, eGDR, CVAI, Non-HDL-C, UHR, and FI when the required component variables were available. The formulas and units of these derived indicators are provided in Supplementary Table S1. FI construction and cross-cohort harmonization are described in Supplementary Table S2, and biomarker availability across CHARLS and HRS is summarized in Supplementary Table S3. Definition In the GBD 2023 framework, stroke is defined according to the 10th edition of the International Classification of Diseases (ICD-10), encompassing intracerebral hemorrhage, ischemic stroke, and sequelae of cerebrovascular disease within the GBD cause hierarchy [1]. The GBD analysis in the present study focused on stroke burden attributable to high fasting plasma glucose and high LDL cholesterol, rather than on individually adjudicated stroke events. The primary GBD outcome metrics included incidence, mortality, DALYs, and the corresponding age-standardized rates [31]. In the CHARLS cohort, stroke status was ascertained through structured face-to-face interviews at each survey wave. Participants who reported a history of stroke at baseline were excluded. Incident stroke was defined as a new physician-diagnosed stroke reported during follow-up from 2013 to 2020, based on an affirmative response to the standardized question: “Have you ever been told by a doctor that you have had a stroke?” [32, 33]. For participants who reported incident stroke, the event time was defined as the interval from baseline to the first follow-up wave in which stroke was reported. For participants who did not report stroke, follow-up time was calculated from baseline to the last available survey wave with valid stroke information. The same outcome-ascertainment logic was applied in the HRS external-validation cohort [34]. Participants with stroke at the 2016 baseline wave were excluded. Incident stroke was defined as a newly reported physician-diagnosed stroke during follow-up through 2022. For participants with incident stroke, follow-up time was calculated from baseline to the first wave in which stroke was reported; for participants without incident stroke, follow-up time was calculated from baseline to the last available wave with valid stroke information. Machine learning model development and evaluation To evaluate the predictive performance of lipid-glucose metabolism biomarkers for incident stroke, seven machine-learning algorithms were developed in the CHARLS discovery cohort: decision tree, random forest, K-nearest neighbors, gradient boosting machine (GBM), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and logistic regression. For internal validation, CHARLS was randomly split into training and testing sets at a 7:3 ratio before preprocessing. Preprocessing parameters and SMOTE were applied within the training set only, and the testing set retained the original outcome distribution. Hyperparameters were tuned by grid search with 5-fold cross-validation, using AUC as the primary selection metric. Final model performance was evaluated in the CHARLS testing set using AUC, sensitivity, specificity, accuracy, precision, and F1 score, with 95% CIs for AUC estimated by 1,000 bootstrap resamples. Comparative performance across machine-learning models is summarized in Supplementary Table S4. External validation was performed in the independent HRS cohort using harmonized outcome definitions and available predictor variables. Model performance was assessed by discrimination, calibration, and decision-curve analysis. To examine whether eGDR provided incremental value beyond conventional clinical predictors, eGDR was added to two clinically oriented baseline models: a Framingham Stroke Risk Profile (FSRP)-type model and a pooled cohort equations (PCE)-type model. Incremental performance was assessed by changes in AUC, integrated discrimination improvement (IDI), continuous net reclassification improvement (NRI), calibration, and decision-curve analysis(DCA) [35, 36]. SHAP-based model interpretation To improve the interpretability of the GBM model, SHAP values were calculated to quantify the contribution of each predictor to the model output [37]. Global feature importance was determined by calculating the mean absolute SHAP value in the testing set, allowing predictors to be ranked according to their overall influence on model predictions. Beeswarm plots and feature-importance plots were further used to visualize how high and low feature values contributed to the predicted risk. Finally, SHAP dependence plots were used to explore potential nonlinear patterns across the range of continuous variables. These SHAP-based analyses were interpreted as model-explanation tools rather than as evidence of causal effects or clinically validated thresholds. Statistical analysis The statistical evaluation of the GBD data focused on assessing the macro-level burden of stroke attributable to lipid-glucose metabolism risks using metrics such as DALYs, mortality, ASDR, and ASMR, all reported with 95% uncertainty intervals (UIs). Joinpoint regression analysis was employed to delineate temporal trends, identify inflection points, and calculate the Annual Percent Change (APC) and its 95% confidence intervals (CIs). To rigorously test the statistical significance of these trends, the Monte Carlo permutation method (4,499 iterations) was applied, with adjustments made using the Bonferroni correction for multiple comparisons to control for false discovery rates. A trend was formally classified as increasing if the lower bound of the 95% confidence interval was greater than zero, as decreasing if the upper bound was less than zero, and as stable if neither condition was met. For the CHARLS and HRS cohorts, continuous variables are presented as medians with interquartile ranges, and categorical variables as counts and percentages. Between-group comparisons were performed using the Mann-Whitney U test or chi-square test, as appropriate. To assess potential selection bias related to missing data, baseline characteristics were compared between included participants and those excluded because of missing data (Supplementary Table S5). Network analysis, correlation matrices, and VIF analysis were used to assess biomarker interrelationships and potential multicollinearity. E-values were calculated to evaluate the potential influence of unmeasured confounding. Associations between lipid-glucose biomarkers and incident stroke were evaluated using Cox proportional hazards models. Univariate models were unadjusted. Multivariable models were adjusted for age, sex, drinking, smoking, hypertension, diabetes, CRP, WBC, platelet count, BUN, SCr, UA, HB, SBP, and DBP. Covariates that were direct mathematical components of a given biomarker were omitted from the corresponding model to reduce duplicate adjustment. FDR-adjusted P values were calculated across biomarker-specific Cox analyses. Sensitivity analyses were performed by excluding participants with extreme BMI values, participants older than 80 years, or participants with diabetes. Multiple-imputation analyses were performed separately for selected missing covariates, with estimates pooled according to Rubin’s rules; stroke status, follow-up time, death status, and biomarker values were not imputed. Fine-Gray competing-risk models were also fitted for key biomarkers to account for death before incident stroke. Restricted cubic spline(RCS) analyses were used to explore nonlinear associations between selected biomarkers and incident stroke risk. Exploratory two-piece Cox threshold-effect analyses were additionally performed for eGDR in CHARLS and HRS, but these analyses were not interpreted as clinically validated cutoffs. All statistical analyses and figure generation were performed using R software. Unless otherwise specified, two-sided P values < 0.05 were considered statistically significant. Results Stroke burden attributable to high FPG and high LDL-C in China In 2023, stroke attributable to high fasting plasma glucose (FPG) accounted for 216.80 thousand deaths and 4378.51 thousand DALYs in China, whereas stroke attributable to high LDL cholesterol accounted for 233.09 thousand deaths and 5,642.72 thousand DALYs (Table 1). From 1990 to 2023, the age-standardized mortality rate (ASMR) attributable to high FPG decreased from 22.74 to 9.84 per 100,000, and the age-standardized DALY rate (ASDR) decreased from 448.30 to 191.70 per 100,000. For high LDL-C, the ASMR decreased from 23.71 to 10.52 per 100,000, and the ASDR decreased from 505.29 to 248.76 per 100,000. The corresponding AAPCs for ASDR were − 2.36% for high FPG and − 2.08% for high LDL-C, while the AAPCs for ASMR were − 2.42% and − 2.47%, respectively. These patterns indicate a sustained decline in age-standardized rates, although the absolute number of deaths and DALYs remained substantial. Temporal patterns and age-specific distributions are shown in Fig. 1. Table 1Number (in thousands), age-standardized rate, and AAPC of stroke burden attributable to high fasting plasma glucose and high LDL cholesterol, 1990–2023High FPG cholesterolHigh LDL cholesterol 1990 Deaths (*103,95% UI)150.19 (94.17, 241.15)160.34 (60.28, 278.23)DALYs (*103,95% UI)3499.97 (2166.96, 5698.27)4063.13 (1592.43, 6940.24)ASMR/100,000(95% UI)22.74 (14.25, 36)23.71 (8.43, 41.81)ASDR/100,000(95% UI)448.3 (283.01, 716.87)505.29 (192.89, 872.69) 2023 Deaths (*103,95% UI)216.8 (154.3, 286.75)233.09 (79.23, 405.16)DALYs (*103,95% UI)4378.51 (3218.73, 5781.73)5642.72 (2133.3, 9377.54)ASMR/100,000(95% UI)9.84 (7.04, 13.11)10.52 (3.55, 18.31)ASDR/100,000(95% UI)191.7 (140.43, 254.13)248.76 (94.59, 412.21) AAPC(1990–2023) ASDR/100,000(95% CI)− 2.36(-2.46,-2.27)− 2.08(-2.16,-2.01)ASMR/100,000(95%CI)− 2.42(-2.50,-2.32)− 2.47(-2.54,-2.39)Estimates are based on the GBD 2023 comparative risk assessment framework and refer to stroke burden attributable to GBD-defined high fasting plasma glucose and high LDL cholesterol. The asterisk in *10³ is used as a multiplication symbol, indicating that deaths and DALYs are reported in thousands; it does not denote statistical significance. DALYs, disability-adjusted life years; ASMR, age-standardized mortality rate; ASDR, age-standardized DALY rate; AAPC, average annual percent change; CI, confidence interval; UI, uncertainty interval Number (in thousands), age-standardized rate, and AAPC of stroke burden attributable to high fasting plasma glucose and high LDL cholesterol, 1990–2023 Estimates are based on the GBD 2023 comparative risk assessment framework and refer to stroke burden attributable to GBD-defined high fasting plasma glucose and high LDL cholesterol. The asterisk in *10³ is used as a multiplication symbol, indicating that deaths and DALYs are reported in thousands; it does not denote statistical significance. DALYs, disability-adjusted life years; ASMR, age-standardized mortality rate; ASDR, age-standardized DALY rate; AAPC, average annual percent change; CI, confidence interval; UI, uncertainty interval Secular trends and age-sex patterns of stroke burden attributable to metabolic risks in China, 1990–2023. A–B Temporal trends in DALYs for stroke attributable to high FPG (A) and high LDL-C (B). Both graphs reveal a significant long-term decline in DALYs for both males (blue line) and females (orange line). The shaded areas represent the 95% uncertainty intervals (UIs). C–D Temporal trends in Mortality for stroke attributable to high FPG (C) and high LDL-C (D). Consistent with the Mortality trends, a substantial decrease in Mortality is observed for both metabolic risk factors across both sexes. E–F Age-specific distributions of stroke DALY rates attributable to high FPG (E) and high LDL-C (F) in 2023. The disease burden shows an exponential increase with advancing age, with the highest concentrations observed in the elderly population (e.g., > 95 years age group). In every age group, the DALY rate for males (blue bars) is significantly higher than for females (pink bars). G–H Age-specific distributions of stroke mortality rates attributable to high FPG (G) and high LDL-C (H) in 2023. Mortality rates increase dramatically with age, exhibiting a J-shaped curve. A clear gender disparity is evident, with male mortality rates (blue bars) surpassing female rates (pink bars) across all age strata Sex-stratified GBD estimates showed a consistently higher metabolic-risk-attributable stroke burden in men than in women (Table 2). In 2023, the ASDR attributable to high FPG was 246.00 per 100,000 in men and 141.90 per 100,000 in women. For high LDL-C, the corresponding ASDRs were 308.09 and 192.26 per 100,000, respectively. Age-standardized rates declined in both sexes from 1990 to 2023, but the male excess persisted across both metabolic risk categories. Joinpoint analyses further characterized the temporal changes in these age-standardized trends (Fig. 2). Table 2Number (in thousands), age-standardized rate, and AAPC of stroke burden attributable to high fasting plasma glucose and high LDL cholesterol by sex, 1990–2023High fasting plasma glucoseHigh LDL cholesterolFemaleMaleFemaleMale 1990 Deaths (*103,95% UI)70.68 (41.47, 114.02)79.51 (45.93, 130.02)78.55 (24.67, 141.07)81.79 (30.01, 132.53)DALYs (*103,95% UI)1544.82 (897.41, 2498.28)1955.15 (1090.82, 3298.24)1875.54 (669.3, 3303.07)2187.59 (830.77, 3530.79)ASMR/100,000(95% UI)19.98 (11.81, 31.75)26.11 (15.77, 41.66)21.87 (6.68, 39.13)25.45 (8.85, 43.42)ASDR/100,000(95% UI)383.7 (225.79, 618.9)519.96 (300.11, 856.21)455.01 (157.77, 786.03)554.77 (205.62, 909.55) 2023 Deaths (*103,95% UI)91.55 (61.09, 126.06)125.25 (89.47, 165.84)100.66 (31.72, 181.54)132.43 (47.13, 230.63)DALYs (*103,95% UI)1725.15 (1213.34, 2287.31)2653.37 (1919.4, 3546.71)2280.64 (822.56, 3860.37)3362.07 (1310.56, 5559.88)ASMR/100,000(95% UI)7.6 (5.03, 10.44)12.44 (8.91, 16.56)8.37 (2.62, 15.06)12.88 (4.54, 22.7)ASDR/100,000(95% UI)141.9 (99.53, 188.23)246 (179.25, 326.64)192.26 (70.09, 321.89)308.09 (120.12, 509.41) AAPC(1990–2023) ASDR/100,000(95% CI)− 2.55(− 2.61,− 2.49)− 1.74(− 1.82,− 1.66)− 2.95(− 3.00,− 2.90)− 2.21(− 2.32,− 2.10)ASMR/100,000(95%CI)− 2.82(− 2.90,− 2.75)− 2.00(− 2.00,− 1.89)− 2.82(− 2.82,− 2.75)− 2.17(− 2.30,− 2.04)Estimates are based on the GBD 2023 comparative risk assessment framework and should be interpreted as sex-specific attributable burden estimates rather than individual-level validation of sex-specific biomarker effects. The asterisk in *10³ is used as a multiplication symbol, indicating that deaths and DALYs are reported in thousands; it does not denote statistical significance. DALYs, disability-adjusted life years; ASMR, age-standardized mortality rate; ASDR, age-standardized DALY rate; AAPC, average annual percent change; CI, confidence interval; UI, uncertainty interval Number (in thousands), age-standardized rate, and AAPC of stroke burden attributable to high fasting plasma glucose and high LDL cholesterol by sex, 1990–2023 Estimates are based on the GBD 2023 comparative risk assessment framework and should be interpreted as sex-specific attributable burden estimates rather than individual-level validation of sex-specific biomarker effects. The asterisk in *10³ is used as a multiplication symbol, indicating that deaths and DALYs are reported in thousands; it does not denote statistical significance. DALYs, disability-adjusted life years; ASMR, age-standardized mortality rate; ASDR, age-standardized DALY rate; AAPC, average annual percent change; CI, confidence interval; UI, uncertainty interval Joinpoint regression analysis of temporal trends in stroke burden attributable to major metabolic risks in China, 1990–2023. A Trends in Age-Standardized DALY Rate (ASDR). B Trends in Age-Standardized Mortality Rate (ASMR). Joinpoint regression analysis was performed to identify significant inflection points (joinpoints) in the temporal trends of age-standardized rates. The Annual Percentage Change (APC) for each segment between joinpoints is calculated, with the ° symbol denoting statistical significance (p < 0.05). Solid lines with circular markers represent the fitted trend segments from the joinpoint model Baseline characteristics of the CHARLS and HRS cohorts The CHARLS discovery cohort included 7,690 participants, among whom 420 developed incident stroke during follow-up. Baseline characteristics are summarized in Table 3. Participants who developed stroke were older than those who did not develop stroke (61.49 ± 8.44 vs. 59.14 ± 9.21 years, P < 0.001), whereas sex distribution did not differ significantly between groups (P = 0.613). Hypertension and diabetes were more frequent among participants with incident stroke, and the distributions of drinking and smoking status also differed between groups. The stroke group had higher levels of FPG, HbA1c, waist circumference, BMI, SBP, DBP, FI, TyG-related indices, LAP, METS-IR, non-HDL-C, and CVAI, whereas eGDR was lower in the stroke group than in the non-stroke group (P < 0.001). Baseline comparisons between included participants and those excluded because of missing data are provided in Supplementary Table S5. The HRS external-validation cohort included 4652 participants, with 233 incident stroke events during follow-up. Baseline characteristics of the HRS cohort are shown in Supplementary Table S7. As in CHARLS, participants with incident stroke were older (61.89 ± 8.71 vs. 58.87 ± 9.14 years, P < 0.001) and had a less favorable cardiometabolic profile, including higher FPG, HbA1c, waist circumference, BMI, SBP, DBP, FI, TyG-related indices, LAP, METS-IR, and AIP, as well as lower eGDR (P < 0.001). Sex distribution did not differ significantly between stroke and non-stroke groups in HRS (P = 0.776). Table 3Baseline characteristics of the CHARLS discovery cohortVariablesTotal (n = 7690)Non-stroke (n = 7270)Stroke (n = 420) P Age, years, Mean ± SD59.26 ± 9.1959.14 ± 9.2161.49 ± 8.44< 0.001 Sex, n(%) 0.613Female4120 (53.58)3900 (53.65)220 (52.38)Male3570 (46.42)3370 (46.35)200 (47.62) Drinking, n(%) 0.030Never drinker4547 (59.13)4321 (59.44)226 (53.81)Former drinker2516 (32.72)2370 (32.60)146 (34.76)Current drinker625 (8.13)577 (7.94)48 (11.43)Smoking, n(%)2 (0.03)2 (0.03)0 (0.00)0.035Never smoke4668 (60.70)4429 (60.92)239 (56.90)Former smoke2345 (30.49)2217 (30.50)128 (30.48)Current smoke676 (8.79)623 (8.57)53 (12.62) Hypertension, n(%) < 0.001No5737 (74.60)5516 (75.87)221 (52.62)Yes1953 (25.40)1754 (24.13)199 (47.38) Diabetes, n(%) < 0.001No7196 (93.58)6821 (93.82)375 (89.29)Yes446 (5.80)404 (5.56)42 (10.00)TC, M (Q1, Q3)190.59 (167.01, 215.34)190.21 (167.01, 214.95)194.65 (170.88, 222.39)0.003TG, M (Q1, Q3)105.32 (74.34, 153.10)104.43 (74.34, 152.22)113.28 (83.19, 157.97)< 0.001HDL-C, M (Q1, Q3)49.48 (40.59, 59.92)49.48 (40.59, 59.92)47.55 (39.82, 56.93)0.010LDL-C, M (Q1, Q3)114.05 (93.17, 137.24)114.05 (93.17, 136.86)119.07 (95.88, 141.11)0.007CRP, M (Q1, Q3)0.99 (0.54, 2.09)0.98 (0.53, 2.05)1.31 (0.66, 2.73)< 0.001HbA1c, M (Q1, Q3)5.10 (4.90, 5.40)5.10 (4.90, 5.40)5.20 (4.90, 5.50)0.001FPG, M (Q1, Q3)102.24 (94.14, 112.86)102.06 (94.14, 112.68)104.94 (95.76, 119.38)< 0.001WBC, M (Q1, Q3)5.94 (4.90, 7.20)5.91 (4.90, 7.20)6.00 (5.10, 7.50)0.088Plt, M (Q1, Q3)207.00 (163.00, 255.00)207.00 (163.00, 255.00)214.00 (164.00, 266.50)0.070BUN, M (Q1, Q3)15.13 (12.55, 18.23)15.13 (12.55, 18.21)15.28 (12.54, 18.35)0.547Scr, M (Q1, Q3)0.76 (0.64, 0.88)0.76 (0.64, 0.88)0.77 (0.67, 0.88)0.111UA, M (Q1, Q3)4.28 (3.56, 5.12)4.28 (3.56, 5.11)4.35 (3.59, 5.25)0.168HB, M (Q1, Q3)14.20 (13.00, 15.50)14.20 (13.00, 15.50)14.50 (13.20, 15.72)0.009WC, M (Q1, Q3)84.60 (77.80, 92.00)84.20 (77.50, 92.00)88.15 (81.00, 96.00)< 0.001BMI, M (Q1, Q3)23.16 (20.91, 25.80)23.10 (20.88, 25.72)24.24 (21.68, 26.95)< 0.001SBP, M (Q1, Q3)126.00 (114.00, 141.50)125.50 (113.50, 140.50)136.50 (121.50, 151.50)< 0.001DBP, M (Q1, Q3)74.50 (67.00, 83.00)74.00 (67.00, 82.50)78.00 (71.00, 87.50)< 0.001FI, M (Q1, Q3)0.14 (0.09, 0.22)0.13 (0.09, 0.21)0.17 (0.11, 0.26)< 0.001TyG, M (Q1, Q3)8.60 (8.22, 9.05)8.59 (8.21, 9.04)8.71 (8.38, 9.12)< 0.001TyG-BMI, M (Q1, Q3)199.47 (175.91, 229.12)198.84 (175.52, 228.49)213.40 (184.74, 241.78)< 0.001CTI, M (Q1, Q3)2.16 (1.89, 2.48)2.15 (1.88, 2.47)2.30 (2.01, 2.60)< 0.001TyG-WC, M (Q1, Q3)727.86 (650.27, 817.37)724.79 (648.41, 814.77)777.01 (686.89, 853.61)< 0.001TyG-CVAI, M (Q1, Q3)818.43 (569.71, 1111.88)808.93 (564.42, 1100.13)996.11 (703.34, 1278.28)< 0.001TyG-FI, M (Q1, Q3)1.18 (0.77, 1.88)1.16 (0.76, 1.84)1.54 (0.96, 2.37)< 0.001TyG-WHtR, M (Q1, Q3)4.62 (4.10, 5.18)4.60 (4.09, 5.17)4.88 (4.40, 5.39)< 0.001AIP, M (Q1, Q3)0.32 (0.12, 0.55)0.32 (0.12, 0.55)0.36 (0.18, 0.60)< 0.001LAP, M (Q1, Q3)26.75 (14.60, 47.40)26.40 (14.42, 46.69)34.46 (18.85, 56.41)< 0.001LCI, M (Q1, Q3)13.59 (7.43, 25.84)13.49 (7.33, 25.36)16.28 (8.81, 30.93)< 0.001METS-IR, M (Q1, Q3)34.37 (29.76, 40.35)34.23 (29.67, 40.10)36.73 (31.44, 42.65)< 0.001eGDR, M (Q1, Q3)10.42 (8.39, 11.27)10.47 (8.64, 11.29)8.81 (6.77, 10.59)< 0.001Non-HDLC, M (Q1, Q3)138.60 (115.59, 164.69)138.02 (115.21, 163.92)146.13 (122.36, 173.29)< 0.001UHR, M (Q1, Q3)0.09 (0.06, 0.12)0.09 (0.06, 0.11)0.09 (0.07, 0.12)0.006CVAI, M (Q1, Q3)95.06 (68.23, 125.20)94.08 (67.38, 123.70)112.86 (83.38, 140.95)< 0.001SD, Standard Deviation; TC, Total Cholesterol; TG, Triglycerides; HDL-C, High-Density Lipoprotein Cholesterol; LDL-C, Low-Density Lipoprotein Cholesterol; CRP, C-Reactive Protein; HbA1c, Glycated Hemoglobin; FPG, Fasting Plasma Glucose; WBC, White Blood Cell; Plt, Platelet; BUN, Blood Urea Nitrogen; Scr, Serum Creatinine; UA, Uric Acid; HB, Hemoglobin; WC, Waist Circumference; BMI, Body Mass Index; SBP, Systolic Blood Pressure; DBP, Diastolic Blood Pressure; FI, Frailty Index; TyG, Triglyceride-glucose Index; TyG-CVAI, Triglyceride-glucose-Chinese Visceral Adiposity Index; TyG-WHtR, Triglyceride-glucose-Waist-to-Height Ratio; AIP, Atherogenic Index of Plasma; LAP, Lipid Accumulation Product; LCI, Lipid Core Index; METS-IR, Metabolic Score for Insulin Resistance; eGDR, Estimated Glucose Disposal Rate; Non-HDL-C, Non-High-Density Lipoprotein Cholesterol; UHR, Uric Acid to High-density lipoprotein Cholesterol Ratio Baseline characteristics of the CHARLS discovery cohort SD, Standard Deviation; TC, Total Cholesterol; TG, Triglycerides; HDL-C, High-Density Lipoprotein Cholesterol; LDL-C, Low-Density Lipoprotein Cholesterol; CRP, C-Reactive Protein; HbA1c, Glycated Hemoglobin; FPG, Fasting Plasma Glucose; WBC, White Blood Cell; Plt, Platelet; BUN, Blood Urea Nitrogen; Scr, Serum Creatinine; UA, Uric Acid; HB, Hemoglobin; WC, Waist Circumference; BMI, Body Mass Index; SBP, Systolic Blood Pressure; DBP, Diastolic Blood Pressure; FI, Frailty Index; TyG, Triglyceride-glucose Index; TyG-CVAI, Triglyceride-glucose-Chinese Visceral Adiposity Index; TyG-WHtR, Triglyceride-glucose-Waist-to-Height Ratio; AIP, Atherogenic Index of Plasma; LAP, Lipid Accumulation Product; LCI, Lipid Core Index; METS-IR, Metabolic Score for Insulin Resistance; eGDR, Estimated Glucose Disposal Rate; Non-HDL-C, Non-High-Density Lipoprotein Cholesterol; UHR, Uric Acid to High-density lipoprotein Cholesterol Ratio Cox regression analyses In the CHARLS cohort, multiple lipid-glucose metabolic biomarkers were associated with incident stroke in univariate analyses (Table 4). After multivariable adjustment and component-aware covariate handling, TC, LDL-C, HbA1c, FPG, waist circumference, BMI, TyG, TyG-BMI, CTI, TyG-WC, TyG-CVAI, TyG-FI, TyG-WHtR, AIP, LAP, METS-IR, non-HDL-C, and CVAI remained positively associated with stroke risk after FDR correction. In contrast, higher eGDR was associated with lower stroke risk (HR, 0.794; 95% CI, 0.756–0.833; FDR P < 0.001). TG, HDL-C, LCI, and UHR did not retain clear evidence of association after FDR correction in the fully adjusted models. HRS external-validation Cox analyses showed a broadly consistent pattern for several key biomarkers (Supplementary Table S8). eGDR remained inversely associated with incident stroke in HRS (HR, 0.820; 95% CI, 0.768–0.875; FDR P < 0.001). HbA1c, FPG, BMI, TyG, TyG-BMI, TyG-WC, AIP, and METS-IR also retained evidence of association after adjustment and FDR correction, whereas TyG-FI, TyG-WHtR, and LAP were not consistently associated with stroke in HRS. Component-aware cross-cohort Cox analyses and multiple-imputation sensitivity analyses supported the robustness of the eGDR findings in both CHARLS and HRS (Supplementary Table S9). Table 4Multivariable cox regression analysisUnivariate analysisMultivariate analyseFDR-adjusted PHR (95%CI) P HR (95%CI) P TC1.004 (1.002–1.007)< 0.0011.003 (1.000–1.005)0.0280.036TG1.001 (1.001–1.002)< 0.0011.001 (1.000–1.002)0.0410.052HDL-C0.989 (0.982–0.996)0.0010.993 (0.986–1.000)0.0480.054LDL-C1.005 (1.002–1.007)< 0.0011.003 (1.000–1.006)0.0270.036HbA1c1.359 (1.267–1.458)< 0.0011.269 (1.160–1.388)< 0.001< 0.001FPG1.006 (1.004–1.007)< 0.0011.004 (1.003–1.006)< 0.001< 0.001WC1.042 (1.032–1.051)< 0.0011.027 (1.017–1.036)< 0.001< 0.001BMI1.067 (1.047–1.088)< 0.0011.051 (1.027–1.075)< 0.001< 0.001FI1.348 (1.068–2.539)< 0.0011.123 (1.016–2.277)0.0060.011TyG1.538 (1.350–1.753)< 0.0011.337 (1.155–1.548)< 0.001< 0.001TyG-BMI1.007 (1.006–1.009)< 0.0011.006 (1.004–1.008)< 0.001< 0.001CTI1.900 (1.573–2.295)< 0.0011.524 (1.158–2.007)< 0.001< 0.001TyG-WC1.004 (1.003–1.004)< 0.0011.003 (1.002–1.003)< 0.001< 0.001TyG-CVAI1.001 (1.001–1.001)< 0.0011.001 (1.001–1.001)< 0.001< 0.001TyG-FI1.403 (1.297–1.518)< 0.0011.181 (1.071–1.301)< 0.0010.002TyG-WHtR1.691 (1.515–1.888)< 0.0011.457 (1.284–1.653)< 0.001< 0.001AIP1.861 (1.417–2.445)< 0.0011.501 (1.109–2.029)0.0080.013LAP1.005 (1.004–1.007)< 0.0011.004 (1.002–1.006)< 0.001< 0.001LCI1.006 (1.003–1.009)< 0.0011.003 (0.999–1.006)0.0950.109METS-IR1.034 (1.025–1.044)< 0.0011.027 (1.016–1.038)< 0.001< 0.001eGDR0.746 (0.716–0.776)< 0.0010.794 (0.756–0.833)< 0.001< 0.001Non-HDLC1.006 (1.003–1.008)< 0.0011.004 (1.001–1.006)0.0030.005UHR1.385 (1.044–2.161)< 0.0011.132 (0.399–2.575)0.2770.300CVAI1.012 (1.010–1.014)< 0.0011.008 (1.006–1.011)< 0.001< 0.001Univariate Cox models were unadjusted. Multivariable Cox models were adjusted for age, sex, drinking status, smoking status, hypertension, diabetes, CRP, WBC, platelet count, BUN, serum creatinine, uric acid, hemoglobin, SBP, and DBP. For each biomarker-specific model, covariates that were direct mathematical components of the corresponding biomarker were omitted from the primary multivariable adjustment set where applicable. Specifically, CRP was omitted for CTI; hypertension was omitted for eGDR; sex was omitted for LAP; age and sex were omitted for CVAI and TyG-CVAI; and uric acid was omitted for UHR. FDR-adjusted P values were calculated using the Benjamini-Hochberg procedure across the multivariable biomarker-specific Cox tests in the CHARLS cohort. Multivariable models used complete-case model-analysis samples. The formulas and component variables are listed in Supplementary Table S1. HR, hazard ratio; CI, confidence interval; FDR, false-discovery rate Multivariable cox regression analysis Univariate Cox models were unadjusted. Multivariable Cox models were adjusted for age, sex, drinking status, smoking status, hypertension, diabetes, CRP, WBC, platelet count, BUN, serum creatinine, uric acid, hemoglobin, SBP, and DBP. For each biomarker-specific model, covariates that were direct mathematical components of the corresponding biomarker were omitted from the primary multivariable adjustment set where applicable. Specifically, CRP was omitted for CTI; hypertension was omitted for eGDR; sex was omitted for LAP; age and sex were omitted for CVAI and TyG-CVAI; and uric acid was omitted for UHR. FDR-adjusted P values were calculated using the Benjamini-Hochberg procedure across the multivariable biomarker-specific Cox tests in the CHARLS cohort. Multivariable models used complete-case model-analysis samples. The formulas and component variables are listed in Supplementary Table S1. HR, hazard ratio; CI, confidence interval; FDR, false-discovery rate Prediction performance and incremental value of eGDR Network analysis was used to describe the interrelationships among the 39 candidate predictors (Fig. 3). Lipid-glucose metabolic indices formed a closely connected cluster with their component anthropometric, lipid, glycemic, and blood pressure variables, indicating substantial biological and mathematical interdependence among these markers. The correlation matrix further showed moderate correlations among several derived biomarkers and their source variables, without evidence of extreme pairwise collinearity (Fig. 4). VIF analysis also did not indicate severe multicollinearity among the predictors (Supplementary Fig. S2). Network analysis of interrelationships among 39 predictors. The network graph illustrates each predictor and the stroke outcome as nodes, with connections (edges) representing statistically significant associations after controlling for all other variables in the network. Node colors correspond to their respective variable groups, as indicated in the legend. The central positioning and extensive connectivity of the Stroke node (red) highlight its strong and direct associations with numerous predictors, emphasizing the multifactorial nature of stroke risk. Dense clusters of metabolic biomarkers (e.g., TyG-index, eGDR, lipid profiles) reveal synergistic effects and shared pathways in stroke pathogenesis. The thickness of the edges reflects the strength of the partial correlations, with thicker lines representing stronger associations Correlation matrix of predictor variables in the CHARLS cohort. The matrix illustrates the complex interdependencies among the predictors. Strong positive correlations (indicated by red clusters) are observed among closely related metabolic indices, such as those within the TyG series and other composite measures of adiposity and insulin resistance. In contrast, these indices exhibit moderate negative correlations (blue cells) with protective factors like eGDR (estimated glucose disposal rate). The overall pattern reveals significant multicollinearity among specific variable groups, highlighting the necessity of using machine learning techniques capable of managing such complex, high-dimensional data. This also underscores the need for Shapley value-based interpretation to isolate the individual contributions of each variable Among the seven machine-learning models evaluated in CHARLS, GBM showed the most favorable overall performance in the internal test set, with an AUC of 0.815 (95% CI, 0.774–0.854), sensitivity of 92.0%, specificity of 80.1%, and F1 score of 91.4% (Supplementary Table S4). Residual distribution and training/test ROC analyses showed broadly consistent model-performance patterns (Supplementary Fig. S3). GBM was therefore used for subsequent model-interpretability analyses. ROC analyses showed that eGDR had the strongest discrimination among the evaluated metabolic biomarkers in both CHARLS and HRS (Fig. 5). To assess incremental predictive value, eGDR was added to two available-component clinical risk models. In CHARLS, adding eGDR increased the AUC of the FSRP-type model from 0.668 to 0.693 (delta AUC, 0.025; P = 0.002) and increased the AUC of the PCE-type model from 0.675 to 0.692 (delta AUC, 0.017; P = 0.006) (Table 5). In HRS, eGDR also improved model discrimination, increasing the AUC from 0.707 to 0.720 for the FSRP-type model (delta AUC, 0.013; P = 0.007) and from 0.714 to 0.722 for the PCE-type model (delta AUC, 0.009; P = 0.016). IDI and continuous NRI were positive in both cohorts, supporting a modest but consistent incremental contribution of eGDR. Decision-curve analyses suggested small net-benefit differences favoring eGDR-augmented models across parts of the threshold range (Fig. 5), while calibration plots are provided in Supplementary Fig. S4. Predictive performance and incremental value of eGDR in the CHARLS and HRS cohorts. This figure summarizes the discrimination and clinical utility of lipid-glucose metabolic biomarkers and eGDR-augmented clinical prediction models. (A-B) ROC curves comparing the predictive performance of individual metabolic biomarkers in CHARLS and HRS. eGDR shows the highest biomarker-only AUC in both cohorts. (C-D) ROC curves comparing FSRP-type and PCE-type clinical models before and after the addition of eGDR in CHARLS and HRS. (E-F) Decision curve analyses evaluating the net benefit of the corresponding clinical models across threshold probabilities. The results indicate that adding eGDR provides modest incremental predictive information beyond conventional clinical risk factors Table 5Incremental predictive value of eGDR added to available-component clinical risk models in the CHARLS and HRS cohortsCohortBase modelPrediction modelAUC(95% CI)ΔAUC(95% CI)P for ΔAUCBrier scoreCalibration interceptCalibration slopeIDIContinuous NRICHARLSFSRPFSRP-type model0.668 (0.641–0.695)Reference0.050− 0.1390.948FSRP-type + eGDR0.693 (0.665–0.720)0.025 (0.012–0.039)0.0020.049− 0.1090.9580.0110.276PCEPCE-type model0.675 (0.648–0.701)Reference0.050− 0.1770.934PCE-type + eGDR0.692 (0.665–0.719)0.017 (0.004–0.029)0.0060.049− 0.1380.9470.0090.240HRSFSRPFSRP-type model0.707 (0.670–0.742)Reference0.047− 0.0941.145FSRP-type + eGDR0.720 (0.683–0.754)0.013 (0.009–0.018)0.0070.049− 0.0791.0290.0110.302PCEPCE-type model0.714 (0.681–0.748)Reference0.047− 0.1251.136PCE-type + eGDR0.722 (0.686–0.757)0.009 (0.002–0.017)0.0160.049− 0.0141.0300.0100.227ΔAUC, IDI, and continuous NRI were calculated relative to the corresponding base model. P values for ΔAUC were estimated using bootstrap resampling. FSRP-type and PCE-type models were constructed using available CHARLS/HRS variables corresponding to components of the Framingham Stroke Risk Profile and Pooled Cohort Equations risk-factor frameworks, respectively; therefore, they should be interpreted as available-component clinical risk models rather than exact replications of the original equations. Because this table evaluates incremental prediction rather than etiologic covariate-adjusted associations, eGDR was added to the clinical model to assess incremental predictive performance; potential overlap between eGDR components and clinical risk-model components should be interpreted in that predictive context. AUC, area under the receiver operating characteristic curve; CI, confidence interval; eGDR, estimated glucose disposal rate; IDI, integrated discrimination improvement; NRI, net reclassification improvement; FSRP, Framingham Stroke Risk Profile; PCE, Pooled Cohort Equations; CHARLS, China Health and Retirement Longitudinal Study; HRS, Health and Retirement Study Incremental predictive value of eGDR added to available-component clinical risk models in the CHARLS and HRS cohorts ΔAUC, IDI, and continuous NRI were calculated relative to the corresponding base model. P values for ΔAUC were estimated using bootstrap resampling. FSRP-type and PCE-type models were constructed using available CHARLS/HRS variables corresponding to components of the Framingham Stroke Risk Profile and Pooled Cohort Equations risk-factor frameworks, respectively; therefore, they should be interpreted as available-component clinical risk models rather than exact replications of the original equations. Because this table evaluates incremental prediction rather than etiologic covariate-adjusted associations, eGDR was added to the clinical model to assess incremental predictive performance; potential overlap between eGDR components and clinical risk-model components should be interpreted in that predictive context. AUC, area under the receiver operating characteristic curve; CI, confidence interval; eGDR, estimated glucose disposal rate; IDI, integrated discrimination improvement; NRI, net reclassification improvement; FSRP, Framingham Stroke Risk Profile; PCE, Pooled Cohort Equations; CHARLS, China Health and Retirement Longitudinal Study; HRS, Health and Retirement Study SHAP-based interpretation and nonlinear analyses SHAP analyses were performed to characterize the contribution of individual predictors to the GBM model. In both CHARLS and HRS, eGDR ranked as the leading predictor according to mean absolute SHAP values (Fig. 6). SBP, LCI, LDL-C, smoking, drinking, TyG-FI, weight, platelet count, hypertension, and TyG also contributed to model predictions, although their relative importance was lower than that of eGDR. The consistency of eGDR as the top-ranked predictor across the two cohorts supported its stability as a model-informative lipid-glucose metabolic marker. Additional SHAP results are presented in Supplementary Fig. S7. SHAP-based interpretation of stroke prediction models in CHARLS and HRS. This figure presents the contribution of individual predictors to the GBM model using SHAP values. A, C SHAP beeswarm plots in CHARLS and HRS, respectively, showing the distribution of SHAP values for the top-ranked predictors. Each point represents an individual participant, with color indicating the corresponding feature value. B, D Mean absolute SHAP value rankings in CHARLS and HRS, respectively, summarizing global feature importance. eGDR ranks as the leading predictor in both cohorts, while SBP, LCI, LDL-C, behavioral factors, TyG-related indices, and other metabolic variables also contribute to model output SHAP dependence plots and RCS analyses were then used to explore the shape of associations for selected high-ranking metabolic biomarkers (Fig. 7). In both CHARLS and HRS, eGDR showed an inverse pattern, with lower eGDR values corresponding to higher model-predicted stroke risk and higher estimated stroke probability. Other biomarkers showed more heterogeneous patterns across cohorts, including positive or nonlinear associations for TyG-related indices, LAP, TG, AIP, LDL-C, and TyG. Sex-stratified SHAP dependence and RCS analyses are presented as exploratory analyses in Supplementary Figs. S5 and S6. SHAP dependence and restricted cubic spline analyses of high-ranking metabolic biomarkers. This figure illustrates the relationships between selected high-ranking metabolic biomarkers and predicted stroke risk in CHARLS and HRS. (A1-A5) SHAP dependence plots for the top-ranked biomarkers in CHARLS. (B1-B5) SHAP dependence plots for the corresponding high-ranking biomarkers in HRS. The plots show how biomarker values relate to SHAP contributions across the observed range. (C1-C5) Restricted cubic spline analyses in CHARLS. (D1-D5) Restricted cubic spline analyses in HRS. The spline curves describe potential nonlinear associations between biomarkers and estimated stroke probability, with shaded areas representing 95% confidence intervals. These analyses are intended to characterize association patterns rather than define clinically validated thresholds Exploratory two-piece Cox analyses were further conducted to formally evaluate potential threshold effects for eGDR (Supplementary Table S10). The overall CHARLS and HRS analyses suggested possible nonlinearity, with selected eGDR thresholds of 11.60 in CHARLS and 11.25 in HRS. Sensitivity and competing-risk analyses Sensitivity and subgroup interaction analyses are summarized in Supplementary Table S6. Additional sensitivity analyses excluding participants with extreme BMI values, participants older than 80 years, or participants with diabetes yielded results broadly consistent with the primary analyses for key biomarkers, particularly eGDR, TyG-WC, TyG-BMI, METS-IR, and AIP (Supplementary Table S11). Event distributions for all-cause mortality and competing-risk analyses are summarized in Supplementary Table S12. Fine-Gray competing-risk models showed that the inverse association of eGDR with stroke remained evident after accounting for death before stroke in both CHARLS and HRS (Supplementary Table S13). These analyses supported the robustness of the principal cohort-based findings. Discussion This study investigated the stroke burden attributable to major metabolic risk factors in China and evaluated the predictive value of lipid-glucose metabolic biomarkers for incident stroke in two longitudinal aging cohorts. The GBD 2023 estimates showed that age-standardized mortality and DALY rates attributable to high FPG and high LDL-C declined from 1990 to 2023, whereas the absolute burden remained substantial. In the cohort analyses, several lipid-glucose metabolism and insulin-resistance-related biomarkers were associated with incident stroke. Among them, eGDR showed an inverse association with stroke risk and comparatively favorable predictive performance, and this pattern was broadly consistent in the HRS external-validation cohort. The coexistence of declining age-standardized rates and a persistent absolute stroke burden is consistent with previous global and Chinese burden-of-disease analyses [1, 23, 32]. Improvements in age-standardized rates may reflect progress in risk-factor control, acute stroke care, and population-level prevention. However, population aging and continued exposure to metabolic and lifestyle risk factors may sustain or increase the absolute number of stroke events. This pattern is particularly relevant in China, where the large aging population means that even modest metabolic risk at the individual level may translate into a substantial public-health burden. The GBD findings therefore provide an epidemiological context for evaluating metabolic biomarkers in middle-aged and older adults, whereas CHARLS and HRS address individual-level risk stratification. Insulin resistance has long been a focus of metabolic stroke research [8, 38, 39]. Since TyG was proposed as a surrogate marker of insulin resistance, several related indices, including TyG-BMI, TyG-FI, CTI, TyG-CVAI, TyG-WC, and TyG-WHtR, have been examined and linked to stroke risk [40–46]. Other metabolic biomarkers, such as eGDR, AIP, and METS-IR, have also been associated with stroke or cardiovascular outcomes [17, 38, 39]. However, the strength and consistency of these associations vary across indices and cohorts [39, 41]. This variation may reflect differences in the biological domains captured by each index, the extent of overlap with conventional risk factors, and cohort-specific differences in anthropometric profiles, laboratory distributions, and vascular risk characteristics. These findings support the evaluation of composite metabolic biomarkers within a shared analytical framework, rather than treating them as interchangeable measures. Against this background, the present study compared multiple lipid-glucose metabolic biomarkers, including both basic laboratory indicators and derived composite indices, using conventional regression, prediction modelling, and model-interpretation approaches. Across these analyses, eGDR showed the most consistent performance. This finding is biologically plausible because eGDR combines glycemic exposure, central adiposity, and hypertension status, thereby reflecting several interrelated dimensions of insulin resistance and vascular injury [38, 39]. Lower eGDR may indicate impaired glucose disposal, greater adiposity-related metabolic stress, and higher vascular load [38]. These processes are linked to endothelial dysfunction, reduced nitric oxide bioavailability, oxidative stress, chronic inflammation, arterial stiffness, and atherosclerotic plaque instability [47–49]. Previous cohort studies have also reported associations between lower eGDR and higher risks of cardiovascular disease, stroke, or dementia [38, 39, 47]. The pathophysiological relevance of eGDR may partly lie in its ability to capture a systemic metabolic phenotype rather than a single biochemical abnormality. Insulin resistance has been closely linked to ischemic stroke through dysregulation of the phosphatidylinositol 3-kinase/Akt pathway, impaired endothelial nitric oxide synthase activation, and reduced nitric oxide bioavailability [47, 48]. Chronic hyperglycemia and oxidative stress may further promote the accumulation of advanced glycation end products, arterial stiffness, endothelial injury, and plaque vulnerability [49]. Thus, lower eGDR may reflect a metabolic environment in which impaired glucose handling, central adiposity, and vascular dysfunction converge. Although the observational design cannot establish causality, these mechanisms provide biological support for the observed associations. SHAP-based interpretation and external validation further supported the relative stability of eGDR as a model-informative marker. In both CHARLS and HRS, eGDR ranked highly in model interpretation and showed an inverse pattern in nonlinear analyses. RCS analyses also suggested that lower eGDR was associated with higher estimated stroke risk. Exploratory two-piece Cox models indicated possible threshold-like patterns in the overall CHARLS and HRS analyses; however, the estimated thresholds were not fully consistent across cohorts and sex strata. Therefore, no eGDR value should be regarded as a clinically validated cutoff. These findings suggest that the association between eGDR and stroke risk may be nonlinear across the biomarker range, but formal threshold confirmation would require prespecified cut-point analyses, independent validation, and evaluation of clinical consequences. The GBD analysis showed a higher stroke burden attributable to high FPG and high LDL-C in men than in women, and previous epidemiological studies of global stroke burden have also reported sex differences in stroke occurrence and outcomes [1, 2, 31, 50]. In contrast, the CHARLS/HRS cohort analyses did not show a clear sex-related difference in biomarker-stroke associations. This discrepancy may reflect the different levels of inference captured by the two approaches. GBD estimates operate at the population level and incorporate differences in metabolic risk exposure, behavioral factors, and demographic structure, whereas CHARLS and HRS evaluate individual-level associations between biomarkers and incident stroke. Therefore, sex differences observed in the GBD analysis should not be directly translated into evidence of sex-specific biomarker effects in the cohort analyses. In the present CHARLS and HRS analyses, sex-stratified SHAP and RCS results did not show clearly divergent eGDR-related risk patterns between men and women. The sex-related findings should therefore be regarded as exploratory and should not be used to define sex-specific biomarker thresholds or screening strategies. Finally, several limitations should be considered. First, stroke ascertainment in CHARLS and HRS was based on self-reported physician diagnosis rather than adjudication by medical records, neuroimaging, or specialist review. This may have introduced false-positive or false-negative classification. The exact timing of stroke onset was also based on the first follow-up wave in which stroke was reported, and ischemic and hemorrhagic stroke subtypes could not be distinguished. These limitations may affect both association estimates and prediction performance. Second, although HRS external validation improves the assessment of generalizability, HRS differs from CHARLS in population structure, health-care context, measurement availability, and laboratory units. Some biomarkers available in CHARLS could not be fully harmonized in HRS, limiting direct comparison across all indices. Third, the primary cohort analyses relied on complete-case samples for biomarker construction. Baseline comparisons and multiple-imputation sensitivity analyses were used to evaluate the influence of missing data, but selection bias cannot be excluded. Fourth, the observational design precludes causal inference. Residual confounding from medication use, diet, physical activity, socioeconomic factors, genetic susceptibility, and unmeasured vascular comorbidities may remain. Finally, the prediction analyses were not designed to establish a deployable clinical screening tool. The incremental value of eGDR requires further evaluation in cohorts with adjudicated stroke outcomes, stroke subtype classification, longer follow-up, formal cost-effectiveness assessment, and prospective implementation studies. Conclusions Age-standardized stroke burden attributable to high FPG and high LDL-C in China has declined over recent decades, yet the absolute burden remains substantial. In the CHARLS and HRS cohorts, eGDR showed a consistent inverse association with incident stroke and provided modest incremental predictive information beyond conventional clinical risk factors. These findings support the potential relevance of lipid-glucose metabolic assessment for stroke risk stratification in middle-aged and older adults, while further studies with adjudicated stroke outcomes and prospective validation are needed before clinical application can be considered. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Acknowledgements The authors gratefully acknowledge CHARLS team for providing the open-access data used in this study. We sincerely thank all the participants for their invaluable contributions to the CHARLS project. Author contributions ZK.X conceptualized the study, conducted data analysis and drafted the manuscript. ZK.X, LJ.D, and YH.D were responsible for data validation and visualization. ZK.X and AH.L participated in manuscript editing and critical revision. W.B supervised the study, provided methodological guidance and approved thefinal version of the manuscript. All authors reviewed and approved the final manuscript. Funding This work was partly supported by the Major Project of National Clinical Key Specialty, Hunan Provincial Health Commission(grant number: 20230279), and Natural Science Foundation of China (81771233 and 82171290). Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate CHARLS was reviewed and approved by the Ethical Review Committee of Peking University (IRB 00001052–11015), and all participants provided written informed consent at the time of participation. 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. ZhenKun Xiao PhD and LiangJie Deng MD have contributed equally to this work. Contributor Information Bing Wang, Email: 1007306925@qq.com. AiHua Liu, Email: doctorliuaihua@foxmail.com. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement No datasets were generated or analysed during the current study.