Baseline, repeated-measure burden, and two-time-point patterns of the triglyceride-glucose index-visceral lipid ratio (TyG-VLR) and incident cardiovascular disease in cardiovascular-kidney-metabolic (CKM) syndrome stages 0–3: a longitudinal China Health and Retirement Longitudinal Study (CHARLS) analysis Authors: Ao-Men Lu, Hao-Jie Han, Qian Xu, Shuang-Ke Bai Published: 2026 ABSTRACT BackgroundCardiovascular-kidney-metabolic (CKM) syndrome stages 0–3 define a population without overt clinical cardiovascular disease (CVD) but with accumulating metabolic, adiposity, lipid, and kidney-related risk. We evaluated whether baseline triglyceride-glucose-visceral lipid ratio (TyG-VLR), repeated-measure TyG-VLR burden, and two-time-point TyG-VLR patterns were associated with incident CVD and CVD subtypes in CHARLS.MethodsData were obtained from the China Health and Retirement Longitudinal Study (CHARLS), with follow-up through the 2020 wave. CKM stages were operationalized using CHARLS-available measures of adiposity, prediabetes, diabetes, hypertension, dyslipidemia, chronic kidney disease, metabolic syndrome, predicted CVD risk, and clinical CVD status. TyG-VLR was calculated as TyG multiplied by the visceral lipid ratio (VLR; LDL-C × Chinese visceral adiposity index). Repeated-measure TyG-VLR burden was calculated as the average of Wave 1 and Wave 3 values multiplied by the interval between assessments before the landmark follow-up period. K-means clustering of standardized Wave 1 and Wave 3 TyG-VLR identified two-time-point patterns based on the available repeated measurements. Cox proportional hazards models, dose–response analyses, marker-specific comparisons, and prespecified sensitivity analyses were performed.ResultsThe baseline analytic cohort included 7,132 participants with CKM stages 0–3 and 1,688 incident CVD events; the median follow-up for CVD was 9.00 years (IQR, 7.00–9.00). The repeated-measure/two-time-point cohort included 4,111 participants and 777 incident CVD events; the median landmark follow-up was 5.00 years (IQR, 4.92–5.00). Each 1-SD increase in baseline TyG-VLR was associated with incident CVD (HR 1.15, 95% CI 1.10–1.21), heart disease (HR 1.11, 95% CI 1.05–1.18), and stroke (HR 1.21, 95% CI 1.12–1.31). Repeated-measure TyG-VLR burden was associated with CVD (HR 1.21, 95% CI 1.12–1.30), and the highest two-time-point pattern group had greater CVD risk than the lowest group (HR 1.51, 95% CI 1.22–1.85). Restricted cubic spline and Kaplan–Meier analyses showed graded CVD risk patterns, whereas marker-specific comparisons showed that TyG-VLR performed broadly similarly to VLR, CVAI, and TyG-CVAI. Incremental discrimination beyond traditional factors and adapted CKM stage was modest. Associations remained consistent in inverse probability-weighted, multiple-imputation, alternative-formulation, and competing-risk analyses.ConclusionsIn middle-aged and older adults with adapted CKM syndrome stages 0–3, higher baseline TyG-VLR, repeated-measure TyG-VLR burden, and higher two-time-point TyG-VLR patterns were associated with incident CVD, particularly stroke. TyG-VLR may be interpreted as an epidemiological marker of combined metabolic, lipid, and adiposity burden, with potential value for population-level risk characterization pending further validation.Graphical abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s12933-026-03302-0. Introduction Cardiovascular disease (CVD), encompassing coronary heart disease and stroke, remains the leading cause of premature mortality and disability worldwide, accounting for a substantial proportion of the global disease burden [1]. In China, stroke contributes disproportionately to this burden, driven by population aging and the progressive accumulation of cardiometabolic risk factors, underscoring the urgent need for early preventive strategies [2, 3]. Recent analyses from the Global Burden of Disease Study further highlight the persistent contribution of metabolic and kidney-related risk factors to vascular disease burden globally [4, 5]. The China Health and Retirement Longitudinal Study (CHARLS) provides a nationally representative platform for evaluating these risks in middle-aged and older Chinese adults [6]. The cardiovascular-kidney-metabolic (CKM) syndrome framework integrates adiposity, dysglycemia, hypertension, dyslipidemia, kidney dysfunction, and cardiovascular injury into a unified, stage-based continuum of disease progression [7]. CKM stages 0–3 represent a critical preclinical window prior to overt cardiovascular disease, during which risk factor modification may have the greatest preventive potential. These early stages are highly prevalent and heterogeneous in the general population, supporting the need for refined metabolic risk stratification using routinely available biomarkers [8, 9]. A recent CHARLS-based longitudinal study comparing nine non-insulin-based surrogate indices of insulin resistance for progression from CKM stages 0–2 to stage 3 found substantial variation in their incremental predictive performance, with TyG-WC and TyG-WHtR showing the strongest predictive value. These findings emphasize that newly proposed composite indices should be directly compared with established alternatives rather than presumed to offer superior performance [10]. The triglyceride-glucose (TyG) index is a well-established surrogate marker of insulin resistance, while the Chinese visceral adiposity index (CVAI) reflects visceral adipose dysfunction in Chinese populations [11, 12]. Low-density lipoprotein cholesterol (LDL-C) captures atherogenic lipid exposure. These three domains—insulin resistance, visceral adiposity, and atherogenic lipid accumulation—represent complementary yet biologically interconnected pathways underlying CKM progression and cardiovascular risk. Based on this rationale, the visceral lipid ratio (VLR), defined as LDL-C × CVAI, was constructed to integrate visceral adiposity and lipid burden, and TyG-VLR further combines TyG with VLR to provide a composite measure of metabolic and lipid-related exposure. The construction of TyG-VLR requires cautious interpretation. Because both CVAI and TyG incorporate triglyceride-related components, structural overlap exists across the composite index, which may partly contribute to observed associations and necessitates direct empirical comparison with related indices. Previous CHARLS-based studies have examined cumulative TyG-BMI, TyG-CVAI, CVAI trajectories, and other CKM-related metabolic markers [13–20]. However, most prior work has focused on single-domain insulin resistance or adiposity measures, and it remains unclear whether integrating multiple complementary metabolic domains provides meaningful incremental information beyond established indices. Therefore, the incremental value of TyG-VLR warrants systematic evaluation. Long-term exposure may better characterize cumulative cardiometabolic burden than single baseline measurements. Repeated assessments may better capture sustained biological exposure while reducing regression dilution bias inherent in single-time-point measurements, particularly in individuals with CKM stages 0–3, in whom metabolic abnormalities accumulate prior to clinical disease onset. Prior longitudinal studies have reported associations between cumulative or repeated TyG-related measures and incident cardiovascular outcomes, including stroke [21–23]. Although multi-trajectory approaches incorporating metabolic and lifestyle factors further support the importance of persistent exposure, such analyses typically require multiple repeated measurements that are not available in the CHARLS dataset [22, 23]. Related work has also combined the TyG index with a frailty index to characterize the joint contribution of metabolic dysfunction and systemic vulnerability to cardiovascular risk [24]. In this study, we therefore investigated the associations of baseline TyG-VLR, repeated-measure TyG-VLR burden, and two-time-point TyG-VLR patterns with incident CVD, heart disease, and stroke among adults with CKM stages 0–3. Given that only two laboratory waves were available, repeated-measure analyses were conducted using a Wave 3 landmark framework, and clustering results were cautiously interpreted as two-time-point exposure patterns rather than true longitudinal trajectories. In addition, we conducted a series of sensitivity and methodological analyses to evaluate component overlap, multicollinearity, model assumptions, incremental predictive performance, missing data mechanisms, selection bias, and clustering robustness. Methods Study design and participants CHARLS is a nationally representative longitudinal survey of Chinese adults aged 45 years or older and their spouses [6]. The present analysis used Wave 1 (2011) as baseline and incorporated follow-up information through the 2020 CHARLS wave, which served as the administrative end of follow-up. Participants were followed until the first incident outcome, death, loss to follow-up, or their interview date in the 2020 CHARLS wave, whichever occurred first. Reporting followed STROBE guidance for observational studies [25]. The study adhered to the principles of the Declaration of Helsinki and received approval from the Biomedical Ethics Review Board of Peking University (IRB 00001052–11015). Written informed consent was obtained from all participants prior to their inclusion in the study. Further details about CHARLS are available on its official website (http://charls.pku.edu.cn/en). The original Wave 1 sample included 17,708 participants. After excluding individuals who were not eligible for CKM stage classification, had CKM stage 4 or baseline clinical CVD, lacked baseline TyG-VLR or required covariates, or had no follow-up information, 7,132 participants were included in the baseline TyG-VLR analysis. For repeated-measure and two-time-point pattern analyses, Wave 3 (2015) was used as the landmark time point; participants additionally required valid Wave 3 TyG-VLR, no clinical CVD before or at Wave 3, and follow-up after Wave 3, yielding 4,111 participants (Fig. 1).Fig. 1Flowchart of participant selection Flowchart of participant selection Assessment of TyG-VLR TyG was calculated as ln[fasting triglycerides (mg/dL) × fasting plasma glucose (mg/dL) / 2]. CVAI was calculated using sex-specific formulas incorporating age, BMI, waist circumference, triglycerides, and HDL-C, with triglycerides and HDL-C expressed in mmol/L [16, 26–28]. VLR was calculated as LDL-C (mg/dL) multiplied by CVAI, and TyG-VLR was calculated as TyG multiplied by VLR [29]. The multiplicative formulation was selected as an exploratory extension of the previously proposed VLR construct rather than on the assumption of an established biological interaction on the product scale. We therefore evaluated log-transformed TyG-VLR and a standardized additive formulation in sensitivity analyses. Because TyG-VLR is a multiplicative composite index with a broad raw numeric scale and no widely established clinical threshold, the main analyses used standardized per-SD increments and quantile-based groups to improve comparability across analyses. Repeated-measure TyG-VLR burden was calculated as the average of Wave 1 and Wave 3 TyG-VLR multiplied by the interval between the two assessments. The raw repeated-measure burden was expressed in arbitrary TyG-VLR-years and has no established clinical threshold; therefore, standardized per-SD values were used in Cox regression analyses. The formulas and related comparator indices are shown in Supplementary Table S1. Classification of CKM syndrome stages CKM stages were assigned hierarchically according to an adapted version of the AHA CKM framework using variables available in CHARLS [7]. Stage 0 indicated no identifiable CKM risk factor. Stage 1 included overweight/abdominal obesity or prediabetes. Stage 2 included diabetes, hypertension, metabolic syndrome, dyslipidemia, or chronic kidney disease. Stage 3 included very-high-risk kidney dysfunction, defined as eGFR < 30 mL/min/1.73 m2, or high predicted cardiovascular risk, defined as a Framingham general CVD 10-year risk ≥ 20%. Stage 4 was defined as clinical CVD and was excluded from the analytic population [30]. The full operational definition is provided in Supplementary Table S2. In line with recent CHARLS-based research examining progression to CKM stage 3 [10], we explicitly report the dataset-specific staging criteria used in the present analysis. Because CHARLS did not include albuminuria, coronary artery calcium, cardiac biomarkers, or the variables needed to implement PREVENT equations, the primary CKM definition should be interpreted as a CHARLS-adapted operationalization. An eGFR-only stage 3 reclassification was summarized descriptively in Supplementary Tables S6 and S7, and a sensitivity analysis excluding participants classified as CKM stage 2 solely by lipid-related criteria is presented in Supplementary Table S25. Assessment of outcomes The primary outcome was incident CVD, defined as the first report of either heart disease or stroke during follow-up. Secondary outcomes were incident heart disease and incident stroke. Outcomes were ascertained from CHARLS self-reported physician diagnosis questions and related treatment fields when available. If heart disease and stroke were first reported in the same follow-up wave, the participant was counted once in the composite CVD analysis and in both corresponding component analyses [31, 32]. Because exact diagnosis dates were unavailable, event time was assigned to the interview wave at which the outcome was first reported; midpoint event-time assignment was examined in sensitivity analyses (Supplementary Table S15). Non-events were censored at the last available follow-up assessment. Covariates Baseline descriptive variables were derived from standardized CHARLS questionnaires, physical measurements, and blood assays. Demographic and lifestyle variables included age, sex, marital status, residence, education, smoking, and drinking. Clinical variables included hypertension, diabetes mellitus, dyslipidemia, kidney disease, cancer, systolic blood pressure, HbA1c, white blood cell count, and C-reactive protein. Diabetes was defined using fasting glucose, HbA1c, self-reported physician diagnosis, or glucose-lowering medication use; the diabetes distributions are shown in Supplementary Tables S3 and S4. Age was retained in adjusted models because it is a major CVD confounder, although the conditional nature of this adjustment was considered when interpreting results involving CVAI-derived indices. Statistical analysis Continuous variables were summarized as means with standard deviations or medians with interquartile ranges, according to their distributional characteristics, whereas categorical variables were summarized as frequencies and percentages. Baseline TyG-VLR was categorized into quartiles because of the larger baseline cohort, whereas repeated-measure TyG-VLR burden was categorized into tertiles to reduce sparse-event instability in the smaller landmark cohort. Alternative baseline tertile and repeated-measure quartile categorizations were examined in sensitivity analyses (Supplementary Table S17). Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs). Model 1 was unadjusted. Model 2 adjusted for age, sex, marital status, residence, education, smoking, and drinking. Model 3 further adjusted for hypertension, diabetes mellitus, dyslipidemia, kidney disease, cancer, systolic blood pressure, HbA1c, white blood cell count, and C-reactive protein. Model 3 was interpreted as a conditional association model because several covariates overlap with cardiovascular-kidney-metabolic (CKM) staging or may lie on cardiometabolic pathways. Variance inflation factor (VIF) diagnostics did not indicate severe multicollinearity, and Schoenfeld residual tests did not suggest violations of the proportional hazards assumption for the main TyG-VLR exposure terms (Supplementary Tables S9 and S14). Because age is incorporated into the formula used to calculate the Chinese visceral adiposity index (CVAI), parallel models excluding additional adjustment for age were fitted to assess whether further age adjustment materially influenced the estimates (Supplementary Table S31). Restricted cubic splines (RCSs) were used to assess dose–response relationships between TyG-VLR and the study outcome, and Kaplan–Meier curves were used descriptively to compare cumulative incidence across exposure groups. Marker-specific Cox models separately compared standardized TyG-VLR with TyG, low-density lipoprotein cholesterol (LDL-C), CVAI, VLR, TyG-CVAI, TyG-BMI, TyG-WC, TyG-WHtR, TyG-WWI, TyG-BRI, and TyG-ABSI. These indices were modelled separately because they share mathematical components, thereby avoiding potential multicollinearity arising from their simultaneous inclusion in the same model (Supplementary Table S8). Incremental predictive performance was evaluated using the C-statistic, time-dependent area under the receiver operating characteristic curve (AUC), Brier score, calibration, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Direct comparisons of AUCs were performed using DeLong tests [33]. The prediction and calibration analyses are reported in Supplementary Tables S21–S24 and S28 and Supplementary Figure S2. To determine whether the findings depended on the multiplicative numerical scale of TyG-VLR, fully adjusted models also evaluated log-transformed TyG-VLR and a standardized additive score defined as Z(TyG) + Z(LDL-C) + Z(CVAI) (Supplementary Table S32). To identify patterns of change in TyG-VLR from Wave 1 to Wave 3 and evaluate their associations with the study outcome, K-means clustering analysis was performed using the “cluster” and “factoextra” packages in R. K-means clustering is an iterative unsupervised classification method that repeatedly updates cluster centroids and assigns individuals with similar characteristics to the same cluster by minimizing the within-cluster sum of squares (WSS). Standardized Wave 1 and Wave 3 TyG-VLR values were used as the clustering variables, and the WSS was calculated for each candidate number of clusters, K. According to the elbow method, the WSS generally decreases as the number of clusters increases. The point at which the additional reduction in WSS becomes relatively small is regarded as an appropriate number of clusters, balancing model complexity and within-cluster homogeneity. This approach has been widely used in analyses of longitudinal biomarker patterns, including studies examining changes in the TyG index in relation to cardiovascular disease risk [34, 35]. As shown in Fig. 2A, the decline in WSS became substantially less pronounced and the curve began to stabilize when K = 3. Therefore, the three-cluster solution was selected based on the elbow pattern, the separation of standardized Wave 1 and Wave 3 TyG-VLR values, adequate cluster sizes, and clinical interpretability. The clustering workflow and the mean raw TyG-VLR patterns within each cluster are shown in Fig. 2 and Supplementary Figure S1. Cluster stability, alternative K-means solutions, predefined change categories, hierarchical clustering, and Gaussian mixture modelling were examined in Supplementary Tables S18, S19, and S27 and Supplementary Figure S3. Because only two biomarker measurement waves were available, these groups were interpreted as repeated-measurement patterns consistent with the available data structure rather than conventional long-term trajectories derived from multiple time points. Multi-wave trajectory modelling was considered beyond the scope of the available measurements. Additional sensitivity analyses included eGFR-only stage 3 reclassification; exclusion of participants classified as CKM stage 2 solely by lipid-related criteria; exclusion of diabetes mellitus and/or HbA1c from the adjustment models; exclusion of dyslipidemia and other potentially overlapping covariates; stabilized inverse probability weighting to account for selection into the repeated-measure cohort; comparison of included and excluded participants; exclusion of early events; midpoint assignment of event times within follow-up intervals; multiple imputation by chained equations; and tests for nonlinearity in the RCS analyses (Supplementary Tables S10–S13, S15, S20, S25, S26, S29, and S30). In wave-based sensitivity analyses, events first reported at the first post-index follow-up wave were additionally excluded (Supplementary Table S33). Stabilized inverse probability weighting diagnostics summarized the distribution of weights and standardized mean differences before and after weighting (Supplementary Table S34 and Supplementary Figure S5). Fine-Gray subdistribution hazard models were used to account for death as a competing event (Supplementary Table S35). Subgroup analyses were conducted according to age, sex, residence, smoking, drinking, hypertension, diabetes mellitus, dyslipidemia, kidney disease, and CKM stage group. Potential effect modification was assessed by including product interaction terms in the models, and false discovery rate correction was applied to interaction P values to reduce the risk of false-positive findings arising from multiple comparisons (Supplementary Table S16). All statistical analyses were conducted using R software version 4.5.2. A two-tailed P value less than 0.05 was regarded as statistically significant.Fig. 2Identification of two-time-point TyG-VLR patterns. Panel A shows the elbow plot for candidate K-means cluster numbers. Panel B shows standardized Wave 1 (2011) and Wave 3 (2015) TyG-VLR values with participant-level connections to cluster centroids. Panel C shows mean raw TyG-VLR values in 2011 and 2015 across the three clusters. The groups were interpreted as repeated-measurement patterns consistent with the two available biomarker waves and should be distinguished from multi-wave trajectory modelling Identification of two-time-point TyG-VLR patterns. Panel A shows the elbow plot for candidate K-means cluster numbers. Panel B shows standardized Wave 1 (2011) and Wave 3 (2015) TyG-VLR values with participant-level connections to cluster centroids. Panel C shows mean raw TyG-VLR values in 2011 and 2015 across the three clusters. The groups were interpreted as repeated-measurement patterns consistent with the two available biomarker waves and should be distinguished from multi-wave trajectory modelling Results Participant selection and baseline characteristics The final baseline analytic cohort included 7132 participants with adapted CKM stages 0–3, among whom 1,688 developed incident CVD during a median follow-up of 9.00 years (IQR, 7.00–9.00; 54,830.7 person-years). The repeated-measure/two-time-point cohort included 4,111 participants and 777 incident CVD events during a median landmark follow-up of 5.00 years (IQR, 4.92–5.00; 19,177.1 person-years). Mean age in the baseline cohort was 58.88 ± 9.16 years, and 3755 (52.7%) participants were women. Higher TyG-VLR quartiles were characterized by higher BMI, waist circumference, systolic blood pressure, LDL-C, triglycerides, CVAI, and a larger proportion of CKM stages 2–3. Diabetes prevalence increased from 10.2% in the lowest baseline TyG-VLR quartile to 23.9% in the highest quartile (Table 1; Supplementary Table S3). Additional baseline, diabetes, included/excluded, and missingness information is provided in Supplementary Tables S4, S12, and S20.Table 1Selected baseline characteristics according to baseline TyG-VLR quartilesVariableOverallQ1Q2Q3Q4P valueParticipants, n71321776177717761803Age, mean ± SD, years58.88 ± 9.1657.26 ± 8.9457.96 ± 8.9159.16 ± 9.0861.11 ± 9.24 < 0.001Female sex, n (%)3755 (52.7)698 (39.3)967 (54.4)1042 (58.7)1048 (58.1) < 0.001BMI, median [IQR], kg/m223.01 [20.77, 25.57]20.33 [18.98, 21.92]22.20 [20.58, 23.96]23.89 [22.15, 25.94]26.14 [24.12, 28.28] < 0.001Waist circumference, mean ± SD, cm84.88 ± 9.8875.91 ± 6.6682.09 ± 6.7887.33 ± 7.3694.07 ± 8.16 < 0.001SBP, mean ± SD, mmHg128.91 ± 21.16122.85 ± 20.34126.27 ± 20.24129.97 ± 20.46136.43 ± 21.16 < 0.001FPG, median [IQR], mg/dL102.24 [94.32, 112.86]98.82 [91.08, 108.36]100.26 [93.24, 109.44]103.14 [95.40, 113.94]106.74 [98.28, 120.06] < 0.001TG, median [IQR], mg/dL103.54 [74.34, 151.34]72.57 [56.64, 99.12]92.93 [70.80, 130.10]114.17 [84.07, 163.95]144.26 [108.86, 196.47] < 0.001HDL-C, mean ± SD, mg/dL51.66 ± 15.2259.25 ± 17.7852.93 ± 14.0448.86 ± 13.6545.68 ± 11.23 < 0.001LDL-C, mean ± SD, mg/dL116.26 ± 34.6891.87 ± 29.12107.87 ± 26.05119.77 ± 27.70145.12 ± 31.60 < 0.001eGFR, median [IQR], mL/min/1.73 m2105.46 [90.89, 122.92]109.71 [94.36, 128.56]107.25 [93.04, 124.33]104.28 [90.07, 121.17]101.04 [86.30, 116.86] < 0.001TyG index, mean ± SD8.66 ± 0.648.30 ± 0.628.53 ± 0.548.78 ± 0.589.04 ± 0.55 < 0.001CVAI, median [IQR]91.80 [66.26, 120.63]51.27 [37.42, 65.18]79.42 [67.95, 93.34]102.74 [89.48, 119.88]133.64 [114.93, 152.26] < 0.001VLR, median [IQR]10,181.03 [6656.21, 14,476.82]4754.79 [3314.33, 5738.27]8403.37 [7504.06, 9308.67]12,089.68 [11042.14, 13,160.93]18,083.92 [16004.66, 21,346.98] < 0.001TyG-VLR, median [IQR]88,140.63 [56031.97, 128,340.98]39,898.63 [27286.41, 47,869.84]71,500.30 [63947.92, 79,490.82]105,501.86 [96787.72, 115,597.54]162,345.22 [141986.36, 192,210.69] < 0.001Hypertension, n (%)3824 (53.6)684 (38.5)839 (47.2)997 (56.1)1304 (72.3) < 0.001Diabetes mellitus, n (%)1126 (15.8)182 (10.2)221 (12.4)292 (16.4)431 (23.9) < 0.001Dyslipidemia, n (%)577 (8.1)47 (2.6)114 (6.4)163 (9.2)253 (14.0) < 0.001CKM stage 0, n (%)598 (8.4)410 (23.1)139 (7.8)47 (2.6)2 (0.1) < 0.001CKM stage 1, n (%)1171 (16.4)389 (21.9)393 (22.1)270 (15.2)119 (6.6)CKM stage 2, n (%)3102 (43.5)585 (32.9)816 (45.9)909 (51.2)792 (43.9)CKM stage 3, n (%)2261 (31.7)392 (22.1)429 (24.1)550 (31.0)890 (49.4)Values are mean ± SD, median [IQR], or n (%), as appropriate. This table was reduced to variables most relevant to the TyG-VLR construct, CKM staging, and cardiometabolic risk. Additional baseline, diabetes, and missingness information is provided in Supplementary Tables S3, S4, S12, and S20 Table 1Selected baseline characteristics according to baseline TyG-VLR quartilesVariableOverallQ1Q2Q3Q4P valueParticipants, n71321776177717761803Age, mean ± SD, years58.88 ± 9.1657.26 ± 8.9457.96 ± 8.9159.16 ± 9.0861.11 ± 9.24 < 0.001Female sex, n (%)3755 (52.7)698 (39.3)967 (54.4)1042 (58.7)1048 (58.1) < 0.001BMI, median [IQR], kg/m223.01 [20.77, 25.57]20.33 [18.98, 21.92]22.20 [20.58, 23.96]23.89 [22.15, 25.94]26.14 [24.12, 28.28] < 0.001Waist circumference, mean ± SD, cm84.88 ± 9.8875.91 ± 6.6682.09 ± 6.7887.33 ± 7.3694.07 ± 8.16 < 0.001SBP, mean ± SD, mmHg128.91 ± 21.16122.85 ± 20.34126.27 ± 20.24129.97 ± 20.46136.43 ± 21.16 < 0.001FPG, median [IQR], mg/dL102.24 [94.32, 112.86]98.82 [91.08, 108.36]100.26 [93.24, 109.44]103.14 [95.40, 113.94]106.74 [98.28, 120.06] < 0.001TG, median [IQR], mg/dL103.54 [74.34, 151.34]72.57 [56.64, 99.12]92.93 [70.80, 130.10]114.17 [84.07, 163.95]144.26 [108.86, 196.47] < 0.001HDL-C, mean ± SD, mg/dL51.66 ± 15.2259.25 ± 17.7852.93 ± 14.0448.86 ± 13.6545.68 ± 11.23 < 0.001LDL-C, mean ± SD, mg/dL116.26 ± 34.6891.87 ± 29.12107.87 ± 26.05119.77 ± 27.70145.12 ± 31.60 < 0.001eGFR, median [IQR], mL/min/1.73 m2105.46 [90.89, 122.92]109.71 [94.36, 128.56]107.25 [93.04, 124.33]104.28 [90.07, 121.17]101.04 [86.30, 116.86] < 0.001TyG index, mean ± SD8.66 ± 0.648.30 ± 0.628.53 ± 0.548.78 ± 0.589.04 ± 0.55 < 0.001CVAI, median [IQR]91.80 [66.26, 120.63]51.27 [37.42, 65.18]79.42 [67.95, 93.34]102.74 [89.48, 119.88]133.64 [114.93, 152.26] < 0.001VLR, median [IQR]10,181.03 [6656.21, 14,476.82]4754.79 [3314.33, 5738.27]8403.37 [7504.06, 9308.67]12,089.68 [11042.14, 13,160.93]18,083.92 [16004.66, 21,346.98] < 0.001TyG-VLR, median [IQR]88,140.63 [56031.97, 128,340.98]39,898.63 [27286.41, 47,869.84]71,500.30 [63947.92, 79,490.82]105,501.86 [96787.72, 115,597.54]162,345.22 [141986.36, 192,210.69] < 0.001Hypertension, n (%)3824 (53.6)684 (38.5)839 (47.2)997 (56.1)1304 (72.3) < 0.001Diabetes mellitus, n (%)1126 (15.8)182 (10.2)221 (12.4)292 (16.4)431 (23.9) < 0.001Dyslipidemia, n (%)577 (8.1)47 (2.6)114 (6.4)163 (9.2)253 (14.0) < 0.001CKM stage 0, n (%)598 (8.4)410 (23.1)139 (7.8)47 (2.6)2 (0.1) < 0.001CKM stage 1, n (%)1171 (16.4)389 (21.9)393 (22.1)270 (15.2)119 (6.6)CKM stage 2, n (%)3102 (43.5)585 (32.9)816 (45.9)909 (51.2)792 (43.9)CKM stage 3, n (%)2261 (31.7)392 (22.1)429 (24.1)550 (31.0)890 (49.4)Values are mean ± SD, median [IQR], or n (%), as appropriate. This table was reduced to variables most relevant to the TyG-VLR construct, CKM staging, and cardiometabolic risk. Additional baseline, diabetes, and missingness information is provided in Supplementary Tables S3, S4, S12, and S20 Selected baseline characteristics according to baseline TyG-VLR quartiles Values are mean ± SD, median [IQR], or n (%), as appropriate. This table was reduced to variables most relevant to the TyG-VLR construct, CKM staging, and cardiometabolic risk. Additional baseline, diabetes, and missingness information is provided in Supplementary Tables S3, S4, S12, and S20 Associations with incident CVD and subtypes In fully adjusted Cox models, each 1-SD increase in baseline TyG-VLR was associated with incident CVD (HR 1.15, 95% CI 1.10–1.21; P < 0.001). Repeated-measure TyG-VLR burden was also associated with incident CVD (HR 1.21, 95% CI 1.12–1.30; P < 0.001). Compared with the lowest two-time-point pattern group, the intermediate and highest pattern groups had higher CVD risk (Cluster 2: HR 1.24, 95% CI 1.05–1.48; Cluster 3: HR 1.51, 95% CI 1.22–1.85) (Table 2).Table 2Association of baseline, repeated-measure, and two-time-point TyG-VLR patterns with incident CVDCharacteristicModel 1Model 1 PModel 2Model 2 PModel 3Model 3 PBaseline TyG-VLR (per 1-SD increase)1.32 (1.27–1.38) < 0.0011.26 (1.20–1.32) < 0.0011.15 (1.10–1.21) < 0.001Baseline TyG-VLR quartilesQ11.00 (Reference)1.00 (Reference)1.00 (Reference)Q21.29 (1.11–1.50)0.0011.23 (1.05–1.43)0.0091.14 (0.98–1.33)0.091Q31.55 (1.34–1.80) < 0.0011.41 (1.22–1.64) < 0.0011.24 (1.06–1.44)0.006Q42.12 (1.84–2.44) < 0.0011.84 (1.59–2.13) < 0.0011.44 (1.23–1.67) < 0.001P for trend < 0.001 < 0.001 < 0.001Repeated-measure TyG-VLR burden (per 1-SD increase)1.38 (1.29–1.47) < 0.0011.33 (1.24–1.42) < 0.0011.21 (1.12–1.30) < 0.001Repeated-measure TyG-VLR burden tertilesT11.00 (Reference)1.00 (Reference)1.00 (Reference)T21.53 (1.26–1.86) < 0.0011.48 (1.22–1.81) < 0.0011.34 (1.10–1.64)0.004T32.33 (1.94–2.80) < 0.0012.15 (1.78–2.60) < 0.0011.71 (1.41–2.09) < 0.001P for trend < 0.001 < 0.001 < 0.001TyG-VLR two-time-point pattern clustersCluster 11.00 (Reference)1.00 (Reference)1.00 (Reference)Cluster 21.50 (1.27–1.77) < 0.0011.44 (1.22–1.71) < 0.0011.24 (1.05–1.48)0.013Cluster 32.22 (1.84–2.68) < 0.0011.99 (1.63–2.42) < 0.0011.51 (1.22–1.85) < 0.001P for trend < 0.001 < 0.001 < 0.001Model 1 was unadjusted. Model 2 adjusted for age, sex, marital status, residence, education, smoking, and drinking. Model 3 further adjusted for hypertension, diabetes mellitus, dyslipidemia, kidney disease, cancer, SBP, HbA1c, white blood cell count, and C-reactive protein. Repeated-measure TyG-VLR was calculated from Wave 1 and Wave 3 measurements and the exposure interval. Two-time-point patterns were generated by K-means clustering of Wave 1 and Wave 3 TyG-VLR. CI; confidence interval, CVD; cardiovascular disease, HbA1c; glycated hemoglobin, HR; hazard ratio, Ref; reference, SBP; systolic blood pressure, SD; standard deviation, TyG-VLR; triglyceride-glucose-visceral lipid ratio Table 2Association of baseline, repeated-measure, and two-time-point TyG-VLR patterns with incident CVDCharacteristicModel 1Model 1 PModel 2Model 2 PModel 3Model 3 PBaseline TyG-VLR (per 1-SD increase)1.32 (1.27–1.38) < 0.0011.26 (1.20–1.32) < 0.0011.15 (1.10–1.21) < 0.001Baseline TyG-VLR quartilesQ11.00 (Reference)1.00 (Reference)1.00 (Reference)Q21.29 (1.11–1.50)0.0011.23 (1.05–1.43)0.0091.14 (0.98–1.33)0.091Q31.55 (1.34–1.80) < 0.0011.41 (1.22–1.64) < 0.0011.24 (1.06–1.44)0.006Q42.12 (1.84–2.44) < 0.0011.84 (1.59–2.13) < 0.0011.44 (1.23–1.67) < 0.001P for trend < 0.001 < 0.001 < 0.001Repeated-measure TyG-VLR burden (per 1-SD increase)1.38 (1.29–1.47) < 0.0011.33 (1.24–1.42) < 0.0011.21 (1.12–1.30) < 0.001Repeated-measure TyG-VLR burden tertilesT11.00 (Reference)1.00 (Reference)1.00 (Reference)T21.53 (1.26–1.86) < 0.0011.48 (1.22–1.81) < 0.0011.34 (1.10–1.64)0.004T32.33 (1.94–2.80) < 0.0012.15 (1.78–2.60) < 0.0011.71 (1.41–2.09) < 0.001P for trend < 0.001 < 0.001 < 0.001TyG-VLR two-time-point pattern clustersCluster 11.00 (Reference)1.00 (Reference)1.00 (Reference)Cluster 21.50 (1.27–1.77) < 0.0011.44 (1.22–1.71) < 0.0011.24 (1.05–1.48)0.013Cluster 32.22 (1.84–2.68) < 0.0011.99 (1.63–2.42) < 0.0011.51 (1.22–1.85) < 0.001P for trend < 0.001 < 0.001 < 0.001Model 1 was unadjusted. Model 2 adjusted for age, sex, marital status, residence, education, smoking, and drinking. Model 3 further adjusted for hypertension, diabetes mellitus, dyslipidemia, kidney disease, cancer, SBP, HbA1c, white blood cell count, and C-reactive protein. Repeated-measure TyG-VLR was calculated from Wave 1 and Wave 3 measurements and the exposure interval. Two-time-point patterns were generated by K-means clustering of Wave 1 and Wave 3 TyG-VLR. CI; confidence interval, CVD; cardiovascular disease, HbA1c; glycated hemoglobin, HR; hazard ratio, Ref; reference, SBP; systolic blood pressure, SD; standard deviation, TyG-VLR; triglyceride-glucose-visceral lipid ratio Association of baseline, repeated-measure, and two-time-point TyG-VLR patterns with incident CVD Model 1 was unadjusted. Model 2 adjusted for age, sex, marital status, residence, education, smoking, and drinking. Model 3 further adjusted for hypertension, diabetes mellitus, dyslipidemia, kidney disease, cancer, SBP, HbA1c, white blood cell count, and C-reactive protein. Repeated-measure TyG-VLR was calculated from Wave 1 and Wave 3 measurements and the exposure interval. Two-time-point patterns were generated by K-means clustering of Wave 1 and Wave 3 TyG-VLR. CI; confidence interval, CVD; cardiovascular disease, HbA1c; glycated hemoglobin, HR; hazard ratio, Ref; reference, SBP; systolic blood pressure, SD; standard deviation, TyG-VLR; triglyceride-glucose-visceral lipid ratio For CVD subtypes, TyG-VLR showed clearer associations with stroke than with heart disease. In the fully adjusted baseline analysis, each 1-SD increase in TyG-VLR was associated with heart disease (HR 1.11, 95% CI 1.05–1.18) and stroke (HR 1.21, 95% CI 1.12–1.31). Repeated-measure TyG-VLR burden and two-time-point patterns showed the same direction, with stronger estimates for stroke (Table 3). Reclassification using an eGFR-only stage 3 definition substantially reduced the number of participants assigned to stage 3 but did not alter eligibility for the overall CKM stages 0–3 analytic cohort (Supplementary Tables S6 and S7).Table 3Fully adjusted associations of TyG-VLR exposures with CVD subtypesOutcomeExposure/contrastnEventsModel 3 HR (95% CI)P valueCVDBaseline TyG-VLR per 1-SD713216881.15 (1.10–1.21) < 0.001CVDRepeated-measure TyG-VLR burden per 1-SD41117771.21 (1.12–1.30) < 0.001CVDCluster 2 vs Cluster 141117771.24 (1.05–1.48)0.013CVDCluster 3 vs Cluster 141117771.51 (1.22–1.85) < 0.001Heart diseaseBaseline TyG-VLR per 1-SD712012791.11 (1.05–1.18) < 0.001Heart diseaseRepeated-measure TyG-VLR burden per 1-SD41025351.17 (1.07–1.28) < 0.001Heart diseaseCluster 2 vs Cluster 141025351.00 (0.82–1.24)0.966Heart diseaseCluster 3 vs Cluster 141025351.35 (1.06–1.73)0.015StrokeBaseline TyG-VLR per 1-SD71265991.21 (1.12–1.31) < 0.001StrokeRepeated-measure TyG-VLR burden per 1-SD41063101.27 (1.14–1.42) < 0.001StrokeCluster 2 vs Cluster 141063101.96 (1.47–2.62) < 0.001StrokeCluster 3 vs Cluster 141063101.97 (1.39–2.79) < 0.001All estimates are from the fully adjusted Model 3. Cluster 1 was the reference group for two-time-point pattern analyses Table 3Fully adjusted associations of TyG-VLR exposures with CVD subtypesOutcomeExposure/contrastnEventsModel 3 HR (95% CI)P valueCVDBaseline TyG-VLR per 1-SD713216881.15 (1.10–1.21) < 0.001CVDRepeated-measure TyG-VLR burden per 1-SD41117771.21 (1.12–1.30) < 0.001CVDCluster 2 vs Cluster 141117771.24 (1.05–1.48)0.013CVDCluster 3 vs Cluster 141117771.51 (1.22–1.85) < 0.001Heart diseaseBaseline TyG-VLR per 1-SD712012791.11 (1.05–1.18) < 0.001Heart diseaseRepeated-measure TyG-VLR burden per 1-SD41025351.17 (1.07–1.28) < 0.001Heart diseaseCluster 2 vs Cluster 141025351.00 (0.82–1.24)0.966Heart diseaseCluster 3 vs Cluster 141025351.35 (1.06–1.73)0.015StrokeBaseline TyG-VLR per 1-SD71265991.21 (1.12–1.31) < 0.001StrokeRepeated-measure TyG-VLR burden per 1-SD41063101.27 (1.14–1.42) < 0.001StrokeCluster 2 vs Cluster 141063101.96 (1.47–2.62) < 0.001StrokeCluster 3 vs Cluster 141063101.97 (1.39–2.79) < 0.001All estimates are from the fully adjusted Model 3. Cluster 1 was the reference group for two-time-point pattern analyses Fully adjusted associations of TyG-VLR exposures with CVD subtypes All estimates are from the fully adjusted Model 3. Cluster 1 was the reference group for two-time-point pattern analyses Dose–response, marker comparison, and incremental prediction Restricted cubic spline tests supported overall associations for CVD, heart disease, and stroke. Nonlinearity was not evident for CVD or heart disease, whereas stroke showed evidence of nonlinearity for both baseline and repeated-measure TyG-VLR burden (Supplementary Table S30; Supplementary Figure S4). The subtype-specific curves were consistent with the Cox models, showing more pronounced nonlinear dose–response patterns for stroke than for heart disease. Figure 3 presents the CVD dose–response curves and descriptive Kaplan–Meier curves for baseline TyG-VLR, repeated-measure TyG-VLR burden, and two-time-point patterns. Detailed marker-discrimination, prediction, calibration, and RCS results are provided in Supplementary Tables S21-S24 and S28 and Supplementary Figure S2.Fig. 3Dose–response and cumulative incidence patterns for TyG-VLR and incident CVD. Panels A and B show restricted cubic spline curves for baseline TyG-VLR and repeated-measure TyG-VLR burden, respectively, with histograms showing exposure distributions. Panels C-E show Kaplan–Meier cumulative incidence curves with numbers at risk for baseline TyG-VLR quartiles, repeated-measure TyG-VLR burden tertiles, and two-time-point TyG-VLR patterns Dose–response and cumulative incidence patterns for TyG-VLR and incident CVD. Panels A and B show restricted cubic spline curves for baseline TyG-VLR and repeated-measure TyG-VLR burden, respectively, with histograms showing exposure distributions. Panels C-E show Kaplan–Meier cumulative incidence curves with numbers at risk for baseline TyG-VLR quartiles, repeated-measure TyG-VLR burden tertiles, and two-time-point TyG-VLR patterns In marker-specific standardized models, TyG-VLR was associated with CVD (HR 1.15, 95% CI 1.10–1.21), heart disease (HR 1.11, 95% CI 1.05–1.18), and stroke (HR 1.21, 95% CI 1.12–1.31). The estimates were similar to those for VLR, CVAI, and TyG-CVAI, supporting comparable performance of TyG-VLR and related composite indices (Supplementary Table S8). Direct discrimination comparisons showed small improvements for TyG-VLR compared with TyG or LDL-C, but little difference compared with CVAI, VLR, or TyG-CVAI (Supplementary Table S28). Incremental prediction metrics indicated modest additional information: adding baseline TyG-VLR to traditional factors plus adapted CKM stage increased the C-index for CVD by 0.0017 and for stroke by 0.0044; exploratory 5-year NRI and IDI analyses suggested limited incremental improvement for overall CVD (Supplementary Tables S21-S24). Subgroup and sensitivity analyses Associations were directionally consistent across most age, sex, residence, smoking, drinking, hypertension, diabetes, dyslipidemia, kidney disease, and CKM-stage subgroups (Supplementary Table S16). Several nominal interaction P values were observed before adjustment for multiple testing; after FDR adjustment, these signals were attenuated, supporting cautious interpretation of subgroup patterns. Sensitivity analyses supported the main findings after excluding diabetes and/or HbA1c from the adjustment model, applying inverse probability weighting to the repeated-measure cohort, excluding early events, using midpoint event-time assignment, changing exposure categorization, and testing alternative cluster solutions (Supplementary Tables S10-S19). Additional analyses showed stable associations after excluding participants classified as CKM stage 2 solely by lipid-related criteria (CVD: baseline TyG-VLR per 1-SD HR 1.14; repeated-measure TyG-VLR burden HR 1.20; Cluster 3 vs Cluster 1 HR 1.52) and after excluding dyslipidemia from the fully adjusted model (CVD: baseline TyG-VLR per 1-SD HR 1.18; repeated-measure TyG-VLR burden HR 1.24; Cluster 3 vs Cluster 1 HR 1.60) (Supplementary Tables S25 and S26). Hierarchical clustering and Gaussian mixture modelling produced broadly similar CVD risk gradients, with agreement with K-means varying across methods (Supplementary Table S27). Multiple imputation produced results similar to the complete-case analysis (Supplementary Table S29). Associations were also consistent in models omitting additional age adjustment (CVD: baseline TyG-VLR HR 1.18, repeated-measure burden HR 1.23, and Cluster 3 HR 1.61; Supplementary Table S31). Multiplicative, log-transformed, and standardized additive formulations produced directionally similar associations and closely comparable C-indices (Supplementary Table S32). After excluding events first reported at the first post-index follow-up wave, continuous exposure associations remained statistically significant; estimates for two-time-point patterns remained directionally consistent but were less precise for heart disease (Supplementary Table S33). Stabilized IPW showed a mean weight of 0.999, and all absolute standardized mean differences were below 0.10 after weighting (Supplementary Table S34 and Supplementary Figure S5). Fine-Gray models treating death as a competing event yielded results comparable to the primary Cox models, including CVD SHRs of 1.17 for baseline TyG-VLR, 1.19 for repeated-measure burden, and 1.49 for Cluster 3 (Supplementary Table S35). Discussion In this longitudinal CHARLS analysis of middle-aged and older adults with adapted CKM syndrome stages 0–3, higher baseline TyG-VLR, greater repeated-measure TyG-VLR burden, and higher two-time-point TyG-VLR patterns were associated with incident CVD. The fully adjusted association was modest for overall CVD at baseline (HR 1.15 per 1-SD), whereas repeated-measure burden showed a slightly stronger association (HR 1.21 per 1-SD). The subtype analysis showed larger estimates for stroke than for heart disease, including nearly two-fold higher stroke risk in the higher two-time-point pattern groups. These findings suggest that TyG-VLR captures vascular risk burden at the population level and may help characterize risk heterogeneity in early CKM stages. The CKM framework provides a useful context for these findings because stages 0–3 represent a preclinical continuum of adiposity, dysglycemia, hypertension, dyslipidemia, and kidney impairment before stage 4 clinical CVD [7]. Within this continuum, a repeated metabolic-lipid-adiposity marker may reflect accumulated exposure across multiple CKM domains. This interpretation is consistent with prior CHARLS evidence showing that cumulative TyG-BMI was associated with incident CVD [13]. Our analysis extends that longitudinal-exposure approach by incorporating LDL-C and CVAI, while indicating that the incremental discrimination of the more complex metric was modest. The biological rationale for TyG-VLR is plausible and is best interpreted in the context of its component structure. Insulin resistance is linked to endothelial dysfunction, arterial stiffness, and CVD risk [36]. LDL-C has a causal role in atherosclerotic cardiovascular disease, supporting the inclusion of an atherogenic lipid component [37]. CVAI was developed to evaluate visceral adipose dysfunction in Chinese adults, and visceral or ectopic fat is strongly related to cardiometabolic disease [11, 38]. Inflammation and immune activation further connect metabolic dysfunction, adiposity, and atherosclerotic vascular injury [39, 40]. These pathways provide a rationale for studying an integrated burden marker, while TyG-VLR serves as a proxy for these interrelated processes rather than a direct mechanistic measure. A central result is that TyG-VLR showed performance comparable to closely related adiposity-lipid indices. In marker-specific analyses, its estimates were similar to those for CVAI, VLR, and TyG-CVAI, and adding TyG-VLR to traditional factors plus adapted CKM stage modestly increased the C-index by 0.0017 for CVD and 0.0044 for stroke. A recent CHARLS-based study comparing nine non-insulin-based insulin resistance surrogates for progression to CKM stage 3 reported that TyG-WC and TyG-WHtR provided the strongest incremental predictive performance [10]. Together with previous CHARLS studies of TyG-CVAI and CVAI-based measures [15, 16], these findings indicate that combining insulin resistance with central adiposity may provide useful risk information, but greater mathematical complexity does not necessarily translate into superior prediction. Consistent with this interpretation, TyG-VLR did not clearly outperform CVAI, VLR, or TyG-CVAI in our analyses. TyG-VLR should therefore be regarded as a complementary epidemiological exposure-burden marker rather than a clinically superior prediction tool. Recent research combining the TyG index with a frailty index suggests that integrating metabolic dysfunction with systemic vulnerability may also contribute to cardiovascular risk characterization [24]. This construct differs conceptually from TyG-VLR, which integrates insulin resistance, visceral adiposity, and LDL-related lipid burden rather than functional vulnerability. Both approaches nevertheless require comparison with simpler component markers before routine clinical application. Component overlap remains an important methodological issue. CVAI includes triglycerides and HDL-C, TyG includes triglycerides and glucose, and CKM staging partly uses dyslipidemia-related criteria. This shared-component structure may influence exposure construction, stage assignment, and covariate adjustment. We therefore used marker-specific models, VIF diagnostics, exclusion of participants classified as CKM stage 2 solely by lipid-related criteria, and alternative adjustment models excluding dyslipidemia, diabetes, and HbA1c. The associations remained directionally stable, although shared-component considerations should still be kept in mind. Associations also remained stable after omitting additional age adjustment and across multiplicative, log-transformed, and standardized additive formulations, with closely comparable discrimination, suggesting that the findings were not confined to a single adjustment choice or numerical scale (Supplementary Tables S31 and S32). TyG-VLR is therefore best viewed as an integrated burden marker reflecting its components rather than as a wholly independent mechanistic pathway. The stronger association with stroke is compatible with the risk-factor profile of older Chinese adults. Stroke burden in China has been closely linked to hypertension and other cardiometabolic risks [2]. Prior CHARLS analyses also reported associations between TyG-related adiposity markers and stroke [16, 41, 42]. In the present study, RCS analyses suggested nonlinearity for stroke but not for overall CVD or heart disease, implying that cerebrovascular risk may rise more clearly at higher TyG-VLR burden. This subtype pattern warrants further evaluation in datasets with adjudicated outcomes and more precise event timing. The two-time-point pattern analysis adds descriptive information beyond baseline TyG-VLR by summarizing repeated measurements over time. Studies with multiple biomarker waves can characterize temporal shapes more directly [43, 44]. The K-means groups summarized relative position and persistence between Wave 1 and Wave 3. Figure 2 illustrates this two-wave structure by displaying standardized 2011 and 2015 values, cluster centroids, and mean raw TyG-VLR levels across clusters. Alternative clustering methods produced broadly similar risk gradients, although agreement was not identical across algorithms, supporting cautious interpretation of pattern-based findings. This study has several strengths, including a nationally representative longitudinal cohort, focus on adapted CKM stages 0–3, repeated biomarker assessment, dose–response and subtype analyses, multicollinearity diagnostics, overadjustment sensitivity analyses, inverse probability weighting with covariate-balance diagnostics, multiple imputation, competing-risk modelling, and supplementary robustness checks. Several issues should be considered when interpreting these results. The observational design precludes causal inference. Outcomes were based on self-reported physician diagnosis and treatment information rather than adjudicated medical records or registry linkage, which may introduce some outcome misclassification. CKM stage 3 was adapted to CHARLS variables and could not include albuminuria, coronary artery calcium, cardiac biomarkers, imaging-defined subclinical CVD, or PREVENT-based risk. Residual selection bias may remain in repeated-measure analyses despite inverse probability weighting and multiple imputation. External validation in cohorts with adjudicated outcomes and more repeated biomarker waves would help clarify its role in future risk-characterization frameworks. Conclusion Among CHARLS participants with adapted CKM syndrome stages 0–3, baseline TyG-VLR, repeated-measure TyG-VLR burden, and higher two-time-point TyG-VLR patterns were associated with incident CVD, particularly stroke. These findings support TyG-VLR as an epidemiological marker of combined metabolic, lipid, and adiposity burden. Incremental prediction beyond the reference clinical model was modest, and TyG-VLR showed broadly comparable discrimination to closely related indices. Further validation is warranted before routine clinical application. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1