Predictive value of cholesterol, high-density lipoprotein, and glucose index-obesity-derived index for stroke: a cohort study Authors: Yueyi He, Chang Liu, Puhua Cao, Bixing Yan, Ling Zhao, Xiaolong Xie, Wei Fang Published: 2026 ABSTRACT BackgroundIn clinical practice, integrating biochemical markers with anthropometric parameters has become a prevalent approach to optimizing disease forecasting. While the cholesterol-high-density lipoprotein-glucose index (CHG) has emerged as a novel metric for quantifying metabolic stress, its prognostic utility for stroke risk—especially when synergized with obesity-related measures—remains to be fully elucidated. This research systematically evaluates the predictive efficacy of CHG-based obesity indices for stroke, utilizing the traditional triglyceride-glucose (TyG) index as a benchmark for comparison.MethodsData for this prospective investigation were derived from the China Health and Retirement Longitudinal Study (CHARLS), encompassing 7,905 individuals aged 45 years who were stroke-free at study inception. A set of seven composite indicators—integrating the CHG index with various obesity metrics (specifically CHG-BMI, CHG-WC, CHG-WHtR, CHG-WWI, CHG-BRI, CHG-ABSI, and CHG-CVAI)—underwent rigorous evaluation.. The discriminative power was quantified via the area under the receiver operating characteristic curve (AUC) and net reclassification improvement. Statistical associations with incident stroke were modeled using multivariate logistic regression, longitudinal trajectory clustering, and restricted cubic splines.ResultsDuring follow-up, stroke was reported in 739/7905 (9.3%) participants. Among all CHG-derived indices, CHG-CVAI and CHG-WC demonstrated the best stroke prediction with an identical AUC value of 0.664 and outperformed TyG-related indices. After full adjustment for confounders, stroke risk for subjects in the highest quartile (Q4) of CHG-CVAI and CHG-WC was 2.44 times (OR = 2.44, 95% CI: 1.66–3.60) and 2.06 times (OR = 2.06, 95% CI: 1.43–2.95) higher compared to those in the Q1 quartile, respectively. Longitudinal analysis demonstrated that stroke risk significantly increased among individuals with high cumulative exposure levels or sustained high levels over time.ConclusionsIn the middle-aged and older Chinese population, CHG-CVAI and CHG-WC are significantly and linearly associated with an elevated risk of incident stroke. The predictive accuracy of these markers surpasses that of TyG and its related indicators. Given these findings, routine surveillance of CHG-CVAI or CHG-WC levels is recommended to facilitate the early screening of susceptible individuals, enabling more effective targeted prevention strategies in healthcare and community settings.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at 10.1186/s12944-026-02980-9. Background A stroke is defined as a critical cerebrovascular event triggered by the abrupt rupture or obstruction of a cranial blood vessel, leading to the cessation of oxygenated blood flow to cerebral tissues. This condition frequently results in permanent neurological damage and is linked to substantial rates of mortality and long-term disability [1]. Data from the 2021 Global Burden of Disease Study indicate an annual global death toll of approximately 7.3 million, with a vast majority of survivors suffering from persistent functional deficits [2]. Beyond the profound deterioration of individual quality of life, stroke imposes a significant socioeconomic and caregiving strain on families, healthcare systems, and society [3]. In the context of China, where stroke represents the primary driver of death and permanent impairment, the rising prevalence necessitates the urgent creation of novel preventive measures [4]. Consequently, exploring modifiable risk factors and establishing streamlined, effective risk assessment instruments are vital for modernizing primary prevention paradigms [5, 6]. Insulin resistance (IR) is a widely established independent risk factor for stroke [7, 8]. Beyond conventional assessment methods, several alternative IR indicators based on routine tests have been extensively studied and applied. These include the triglyceride-glucose index (TyG) and the Atherosclerotic Potential Index, which are widely used in risk assessment because of their simplicity [9, 10]. Mansoori et al. proposed that the cholesterol-high-density lipoprotein-glucose index (CHG) was an useful tool for assessing an individual’s IR and metabolic burden. The cumulative evaluation of fasting blood glucose (FBG), high-density lipoprotein cholesterol (HDL-C), and total cholesterol (TC), also known as the CHG, has demonstrated superior predictive ability for type 2 diabetes risk compared to TyG [11]. CHG is also independently associated with multiple disease risks [12–17]. It exhibits a stronger link to cardiovascular disease (CVD) risk than TyG, thereby suggesting its unique and superior predictive value [18]. Additionally, the CHG is independently linked to escalating blood pressure and the development of incident hypertension, reinforcing its role in comprehensive vascular health assessment [19]. Obesity represents another major stroke risk factor [20]. In addition to conventional anthropometric indices such as body mass index (BMI) and waist circumference (WC), several novel measures—including Chinese visceral adiposity index (CVAI), anthropometric body shape index (ABSI), body roundness index (BRI), weight-adjusted waist index (WWI), and waist-to-height ratio (WHtR)—have been increasingly used to accurately capture body fat distribution and the corresponding health implications [21–25]. Recent research trends indicate that combining blood markers for glucose and lipid metabolism with anthropometric indicators reflect body fat distribution more accurately and can more effectively assess disease risk [26–28]. However, none of the published studies have systematically investigated the predictive value of the CHG for stroke risk when combined with various classic and novel obesity indicators. Therefore, this study used a prospective cohort data from the China Health and Retirement Longitudinal Study (CHARLS) to evaluate the predictive performances of a series of CHG-obesity-derived indicators for stroke events in middle-aged and older adults, and compared them with TyG and its derived indicators. The study hypothesized that combining CHG with obesity measures would effectively capture synergistic risk information related to metabolic abnormalities and body fat distribution, thereby enhancing the prediction of stroke risk. The novelty of this study lies in the fact that it is the first to systematically construct and compare the predictive capabilities of CHG-obesity-derived indicators for stroke in a nationwide prospective cohort. It also aimed to examine their cumulative exposure levels and longitudinal trajectories, thereby providing new evidence for dynamic risk assessment. Methods Study design and participants The foundation of this nationwide prospective cohort investigation was derived from the open-access database of the China Health and Retirement Longitudinal Study (CHARLS). This extensive longitudinal initiative aims to monitor the health and socioeconomic status of the middle-aged and elderly population in China. While the initial baseline assessment was finalized during 2011–2012, subsequent follow-up surveys were systematically executed in 2013, 2015, 2018, and 2020. To ensure national representation, a multistage stratified probability sampling approach was employed, recruiting households from 450 urban and rural communities across 28 provinces. Data collection for eligible participants involved standardized face-to-face interviews, anthropometric measurements, and the acquisition of venous blood samples. To investigate the association between CHG and obesity-derived indicators with new stroke events, a participant screening process was established (Fig. 1). Using 2011 as baseline, individuals younger than 45 years old, those with missing core exposure indicators, and those with a confirmed stroke diagnosis at baseline were excluded. Subsequently, to examine cumulative exposure and identify longitudinal trajectories of each indicator, individuals who had experienced the outcome by the 2015 follow-up or missing exposure variable data in 2015 were further excluded. The final cohort was comprised of participants with complete exposure information and had completed at least one follow-up visit.Fig. 1Study flowchart Study flowchart Definition and calculation of exposure The core exposure indicator in this study is a composite index derived from CHG and various obesity metrics. All indicators were calculated based on the biochemical parameters measured from fasting venous blood and anthropometric data using the following formulas:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$CHG\hspace{0.17em}=\hspace{0.17em}Ln [TC (mg/dL) \times FBG (mg/dL) / 2 \times HDL (mg/dL)]$$\end{document} Obesity Indicators: BMI = Weight (kg)/Height2 (m2). WC = Circumference of the waist at the umbilical level in a natural standing position (cm). WHtR = WC (cm)/Height (cm). WWI = WC (cm)/√Weight (kg). BRI = [WC (cm)/(2 × π × √Height (m)/(2 × π)] – 10. ABSI = WC (cm)/[BMI^(2/3) (kg/m2) × Height^(1/2) (m)]. CVAI = [WC (cm)/(Height (m) × √Height (m))] + [TG (mg/dL)/1.7] + [FBG (mg/dL)/5.6]—[HDL-C (mg/dL)/1.0]—10.0 Note: TG denotes triglycerides. By multiplying the CHG index with each of the aforementioned obesity indicators, we constructed seven composite derived exposure indices: CHG-BMI, CHG-WC, CHG-WHtR, CHG-BRI, CHG-WWI, CHG-CVAI, and CHG-ABSI. Definition and determination of the primary outcome In this study, new-onset stroke served as the primary outcome measure. The definition entailed the first instance of a physician-confirmed stroke diagnosis reported by participants during any follow-up interview held in 2013, 2015, 2018, or 2020. To ensure consistency in event capturing, the CHARLS project utilized standardized assessment tools, specifically incorporating the question: “Have you ever been diagnosed with a stroke by a doctor?” If a participant answered “yes” at any follow-up, that visit date was recorded as the event date. For participants reporting stroke at multiple follow-ups after baseline, the first report date was used. For participants reporting that stroke had not occurred throughout follow-up, the observation ended on the date of their last visit or the study's conclusion in 2020, whichever came first. Collection and definition of covariates To mitigate the influence of potential confounding variables, a comprehensive array of multi-domain covariates was extracted from the CHARLS baseline data and integrated into multivariate adjustment models. These covariates primarily consisted of the following five major categories: (1) Demographic characteristics such as age, gender, place of residence, educational attainment, and marital status; (2) Lifestyle and behavioral factors, including smoking status and drinking status; (3) Medical History and Health Status, which was based on the participants' self-reported physician diagnoses, and included the presence of hypertension, diabetes, heart disease, liver disease, kidney disease, and cancer history; (4) Physical measurement indicators, including height, weight, WC, mean systolic blood pressure (SBP), and mean diastolic blood pressure (DBP); (5) Laboratory tests, including FBG, TC, TG, HDL-C, low-density lipoprotein cholesterol (LDL-C), C-reactive protein (CRP), uric acid (UA), glycated hemoglobin (HbA1c), blood urea nitrogen (BUN), creatinine (Cr), and cystatin C (Cys C). To assess renal function, the estimated glomerular filtration rate (eGFR) was determined via the CKD-EPI 2009 equation, utilizing serum creatinine concentrations as the primary input. Statistical analysis All computational analyses were conducted in the Python environment. Significance levels were set at P < 0.05 (two-tailed). Descriptive statistics for continuous variables are presented in the form of mean ± standard deviation, whereas categorical variables are reported as counts (percentages). For the evaluation of intergroup variations, independent t-tests, Mann–Whitney U tests, or X2 tests were selected in accordance with the specific data distribution characteristics. To handle missing data in the covariates, an iterative optimized prediction mean matching method was used for multiple imputations. The iterative process continued until the rate of change in the imputed values fell below a pre-set threshold (0.001) or reached a maximum of 10 iterations. The final imputation result was based on the converged nearest-neighbor matching values. The proportion of missing data in the covariates permitted for imputation should be below 30%. To balance the model’s predictive performance with epidemiological reliability, this study used a strategy combining prior knowledge with data-driven approaches to select covariates. This strategy not only enhanced the model’s interpretability, but also optimized its predictive performance. The screening process included the following three steps: (1) Conducting univariate logistic regression analysis of all the candidate covariates and retaining variables significantly associated with stroke outcomes (P < 0.05); (2) Applying both the Boruta algorithm and Recursive Feature Elimination (RFE) to rank and select the importance of the initial variables and selecting the top 15 variables. Then, variables identified as important by both methods were retained. Furthermore, age, gender, smoking, alcohol consumption, and heart disease were included as established stroke risk factors. The INTERSTROKE study explicitly identified these risk factors as the primary determinants of the global burden of stroke [29]; (3) Multicollinearity was checked in the final variable set. If the absolute value of the Pearson correlation coefficient between any two variables exceeded 0.6, variables with relatively weaker clinical significance were removed to ensure model stability. To ensure that the covariates selected by the Boruta and RFE algorithms were statistically reliable, the predictive performance of the underlying models on which they rely were evaluated. To evaluate and compare the predictive performance of each exposure indicator, a “baseline” logistic regression model containing only the final selected covariates was constructed. Subsequently, the CHG index and each derived indicator were sequentially added individually to the baseline model, thereby forming a series of new models. Subsequently, receiver operating characteristic (ROC) curves were generated for each indicator, and the corresponding area under the curve (AUC) values served as discrimination metrics. Pairwise comparisons of AUC values across indicators were performed using the DeLong test. Additionally, net reclassification improvement (NRI) and integrated discriminant improvement were computed to assess the added value of new indicators in terms of risk reclassification capability and overall discriminative ability. To enhance the comparability of various metrics, Z-score standardization was performed for CHG, TyG, and their derived metrics prior to performance prediction. To evaluate the relationship between the best-performing CHG-derived metric and the incidence of stroke, a multivariate logistic regression model was utilized in a progressive manner. The analytical framework consisted of three hierarchical stages: first, an unadjusted crude model (Model 1); second, a model adjusted for age, gender, and behavioral habits (smoking and drinking) (Model 2); and finally, a fully adjusted model (Model 3) that accounted for all relevant covariates prioritized during feature selection. Because of the binary nature of stroke events in this study and the relatively fixed follow-up duration, logistic regression was selected as the primary statistical method to ensure consistency in the analytical framework and comparability of results. All assessments of predictive performance (ROC curves, IDI, NRI) and effect estimates were based on this model. Although Cox regression is commonly used for time-to-event analysis, logistic regression is a more direct and methodologically consistent choice under the current study design. To capture clinically meaningful dynamic patterns in indicators from longitudinal data, unsupervised classification of the optimal derivative indicator trajectories was performed using the computationally efficient and intuitive K-means clustering algorithm based on measurements from baseline and 2015. The number of clusters was determined using the elbow method combined with clinical interpretability. This resulted in the identification of subgroups with distinct dynamic evolution characteristics and comparing outcome risk differences across these subgroups. Since this study is comprised of repeated measurement data from only two time points, the choice of K-means over group-based trajectory modeling (GBTM) was primarily based on considerations of methodological suitability. GBTM and its extensions, such as latent category growth models, typically require at least three time points to reliably estimate trajectory shapes, such as linear or quadratic terms [30, 31]. When only two waves of data are available, the trajectories are closer to patterns of change or represent classifications rather than continuous functional forms. In this context, the K-means algorithm can achieve the classification objective in a transparent and reproducible manner and avoid potential issues of overfitting or unidentifiability associated with parametric models. Furthermore, the K-means algorithm does not require any prior assumptions about data distribution. It is computationally efficient and produces intuitive, clear classification results, facilitating clinical interpretation and subsequent regression analysis [32]. This method has been applied in several two-wave longitudinal studies based on CHARLS data [26, 27, 33] and therefore provides a basis for comparability within the field for the present study. A restricted cubic spline (RCS) model was utilized to detect possible nonlinearities in the dose–response relationship between predictive indicators and outcome risks. To minimize the impact of outliers on the analysis results, continuous variables within the lower and upper 5% ranges were trimmed. Therefore, observations below the 5th percentile and above the 95th percentile were replaced with the 5th and 95th percentiles, respectively. Moreover, subgroup analyses were conducted with parameters such as age (< 60 years, ≥ 60 years), gender, marital status, education level, smoking status, drinking status, presence of hypertension, presence of diabetes, and Cardiovascular-Kidney-Metabolic (CKM) stage. Interaction terms were included to assess effect heterogeneity. Four sensitivity tests were carried out to verify the stability of the primary results. These included: (1) performing the logistic regression on raw data limited to complete cases prior to any multiple imputation; (2) accounting for temporal differences in follow-up by applying Cox regression to analyze stroke as time-to-event data; (3) excluding individuals who experienced the outcome during the fifth follow-up period to test the consistency of the associations; (4) After controlling for covariates that overlap with the study indicators, an incremental predictive value analysis was conducted once again to ensure that the predictive value was not unduly adjusted or diminished. Results Baseline characteristics This study included 7,905 participants, including 739 individuals (9.3%) in the stroke group and 7,166 individuals (90.7%) in the non-stroke group. Individuals in the stroke cohort were generally older and presented with markedly higher baseline values for blood pressure (SBP and DBP), glycemic markers (FBG and HbA1c), and anthropometric indices (weight, BMI, and WC). Furthermore, the stroke group exhibited significantly elevated concentrations of TC, TG, LDL-C, and CRP, whereas HDL-C levels and eGFR estimates were significantly reduced compared to the non-stroke group. Moreover, the proportions of hypertension, diabetes, cardiac disease, and stage 4 CKD were significantly higher among individuals in the stroke cohort compared to those without stroke. However, both groups did not show significant differences in gender, residence, history of liver disease, or history of cancer. The quartile distributions of CHG-derived indices, CHG-CVAI and CHG-WC, showed a significant increase in stroke proportion with rising index levels(Table 1). In the cumulative exposure analysis, 4,544 participants with two measurement dates were included. Their baseline characteristics aligned with the overall population trends, with the stroke group exhibiting significantly higher cumulative CHG-CVAI and CHG-WC levels than the non-stroke group (Table S1).Table 1Baseline characteristics of the study population, stratified by stroke statusCharacteristicsOverall N = 7905Non-stroke N = 7905Stroke N = 739PAge(year), M (Q₁, Q₃)59.00 (52.00, 65.00)58.00 (52.00, 65.00)61.00 (55.00, 67.00) <.001BUN(mg/dL), M (Q₁, Q₃)15.10 (12.55, 18.21)15.10 (12.55, 18.21)15.21 (12.62, 17.97)0.737Cr(mg/dL), M (Q₁, Q₃)0.76 (0.66, 0.88)0.76 (0.64, 0.88)0.77 (0.67, 0.89)0.009Cys C(mg/dL), M (Q₁, Q₃)0.98 (0.86, 1.12)0.98 (0.86, 1.12)1.00 (0.87, 1.16)0.001HbA1c(mg/dL), M (Q₁, Q₃)5.20 (4.90, 5.40)5.10 (4.90, 5.40)5.20 (4.90, 5.60) <.001DBP(mmHg), M (Q₁, Q₃)74.50 (66.50, 83.00)74.00 (66.50, 82.50)78.00 (70.00, 87.00) <.001SBP(mmHg), M (Q₁, Q₃)126.00 (113.50, 141.00)125.50 (113.50, 140.00)134.00 (119.50, 150.50) <.001TC(mg/dL), M (Q₁, Q₃)191.75 (168.56, 216.11)190.98 (168.17, 215.72)195.23 (172.04, 220.75)0.002TG(mg/dL), M (Q₁, Q₃)104.43 (74.34, 150.45)103.54 (73.46, 148.68)115.05 (84.96, 164.61) <.001HDL-C(mg/dL), M (Q₁, Q₃)49.48 (40.59, 60.31)49.87 (40.98, 60.70)47.17 (39.43, 56.44) <.001LDL-C(mg/dL), M (Q₁, Q₃)115.59 (94.72, 138.40)115.21 (94.33, 137.63)119.85 (98.78, 141.88) <.001CRP(mg/dL), M (Q₁, Q₃)1.02 (0.55, 2.14)0.98 (0.54, 2.09)1.33 (0.68, 2.64) <.001UA(mg/dL), M (Q₁, Q₃)4.29 (3.56, 5.13)4.28 (3.56, 5.11)4.35 (3.56, 5.36)0.054FBG(mg/dL), M (Q₁, Q₃)102.24 (94.50, 112.32)102.06 (94.50, 111.96)104.40 (96.03, 116.73) <.001BMI 2011(kg/m2), M (Q₁, Q₃)23.14 (20.86, 25.74)23.06 (20.77, 25.62)24.02 (21.69, 26.80) <.001Height(kg), M (Q₁, Q₃)1.57 (1.52, 1.64)1.57 (1.52, 1.64)1.58 (1.52, 1.64)0.534Weight(kg), M (Q₁, Q₃)57.70 (50.80, 65.70)57.50 (50.70, 65.40)60.00 (53.00, 68.70) <.001WC(cm), M (Q₁, Q₃)84.80 (78.00, 92.00)84.20 (77.90, 91.60)88.60 (81.00, 96.00) <.001eGFR, M (Q₁, Q₃)94.46 (83.89, 102.00)94.66 (84.28, 102.28)92.48 (81.41, 99.17) <.001CHG, M (Q₁, Q₃)5.28 (5.03, 5.56)5.27 (5.03, 5.55)5.39 (5.15, 5.65) <.001TyG, M (Q₁, Q₃)8.59 (8.22, 9.02)8.57 (8.21, 9.01)8.74 (8.38, 9.13) <.001WHtR, M (Q₁, Q₃)0.54 (0.49, 0.59)0.53 (0.49, 0.58)0.56 (0.51, 0.60) <.001VAI, M (Q₁, Q₃)3.41 (2.04, 5.90)3.31 (2.00, 5.82)4.04 (2.49, 6.78) <.001ABSI, M (Q₁, Q₃)0.08 (0.08, 0.09)0.08 (0.08, 0.09)0.08 (0.08, 0.09) <.001BRI, M (Q₁, Q₃)4.05 (3.23, 5.10)4.00 (3.20, 5.04)4.54 (3.62, 5.52) <.001CVAI, M (Q₁, Q₃)94.96 (68.80, 125.28)92.98 (67.53, 122.77)112.43 (83.65, 140.29) <.001WWI, M (Q₁, Q₃)11.13 (10.59, 11.72)11.11 (10.57, 11.70)11.31 (10.80, 11.91) <.001Liver Disease, n(%)0.114 No7632 (96.55)6926 (96.65)706 (95.53) Yes273 (3.45)240 (3.35)33 (4.47)Kidney Disease, n(%)0.049 No7454 (94.29)6769 (94.46)685 (92.69) Yes451 (5.71)397 (5.54)54 (7.31)Heart Disease, n(%) <.001 No6988 (88.40)6391 (89.19)597 (80.78) Yes917 (11.60)775 (10.81)142 (19.22)Cancer, n(%)0.619 No7839 (99.17)7105 (99.15)734 (99.32) Yes66 (0.83)61 (0.85)5 (0.68)Diabetes, n(%) <.001 No7421 (93.88)6748 (94.17)673 (91.07) Yes484 (6.12)418 (5.83)66 (8.93)Hypertension, n(%) <.001 No5875 (74.32)5469 (76.32)406 (54.94) Yes2030 (25.68)1697 (23.68)333 (45.06)Sex, n(%)0.273 Female4259 (53.88)3875 (54.07)384 (51.96) Male3646 (46.12)3291 (45.93)355 (48.04)Residence, n(%)0.991 City2772 (35.07)2513 (35.07)259 (35.05) Village5133 (64.93)4653 (64.93)480 (64.95)Education, n(%)0.020 Below primary3996 (50.55)3602 (50.27)394 (53.32) Primary schools1705 (21.57)1531 (21.36)174 (23.55) Middle school1454 (18.39)1346 (18.78)108 (14.61) High school and above750 (9.49)687 (9.59)63 (8.53)Marital, n(%)0.075 Other972 (12.30)866 (12.08)106 (14.34) Married6933 (87.70)6300 (87.92)633 (85.66)Smoking, n(%)0.004 Never4855 (61.42)4433 (61.86)422 (57.10) Former680 (8.60)595 (8.30)85 (11.50) Now2370 (29.98)2138 (29.84)232 (31.39)Drinking, n(%)0.004 Never4644 (58.75)4235 (59.10)409 (55.35) Former646 (8.17)563 (7.86)83 (11.23) Now2615 (33.08)2368 (33.04)247 (33.42)CKM, n(%) <.001 0 stage543 (6.87)525 (7.33)18 (2.44) 1 stage1007 (12.74)958 (13.37)49 (6.63) 2 stage3465 (43.83)3189 (44.50)276 (37.35) 3 stage1980 (25.05)1726 (24.09)254 (34.37) 4 stage910 (11.51)768 (10.72)142 (19.22)CHG-CVAI quartile, n(%) <.001 Q11976 (25.00)1876 (26.18)100 (13.53) Q21976 (25.00)1831 (25.55)145 (19.62) Q31976 (25.00)1774 (24.76)202 (27.33) Q41977 (25.01)1685 (23.51)292 (39.51)CHG-WC quartile, n(%) <.001 Q11976 (25.00)1871 (26.11)105 (14.21) Q21976 (25.00)1829 (25.52)147 (19.89) Q31976 (25.00)1766 (24.64)210 (28.42) Q41977 (25.01)1700 (23.72)277 (37.48)Values are expressed as mean (standard deviation), median (quartile range), or n (%)TG Triglycerides, HDL-C High-Density Lipoprotein Cholesterol, LDL-C Low-Density Lipoprotein Cholesterol, hs-CRP High-Sensitivity C-Reactive Protein, FBG Fasting Blood Glucose, AIP Atherogenic Index of Plasma, CVAI Chinese Visceral Adiposity Index, ABSI A Body Shape Index, TyG Triglyceride-Glucose Index, UA Uric Acid, eGDR Estimated Glucose Disposal Rate, CKM Cardiovascular-Kidney-Metabolic, Age Age, BUN Blood Urea Nitrogen, Cr Creatinine, Cys C Cystatin C, HbA1c Hemoglobin A1c, DBP Diastolic Blood Pressure, SBP Systolic Blood Pressure, TC Total Cholesterol, CRP C-Reactive Protein, BMI Body Mass Index, WC Waist Circumference, CHG Cholesterol, high-density lipoprotein, and glucose index, WHtR Waist-to-Height Ratio, VAI Visceral Adiposity Index, BRI Body Roundness Index, WWI Weight-Waist Index Table 1Baseline characteristics of the study population, stratified by stroke statusCharacteristicsOverall N = 7905Non-stroke N = 7905Stroke N = 739PAge(year), M (Q₁, Q₃)59.00 (52.00, 65.00)58.00 (52.00, 65.00)61.00 (55.00, 67.00) <.001BUN(mg/dL), M (Q₁, Q₃)15.10 (12.55, 18.21)15.10 (12.55, 18.21)15.21 (12.62, 17.97)0.737Cr(mg/dL), M (Q₁, Q₃)0.76 (0.66, 0.88)0.76 (0.64, 0.88)0.77 (0.67, 0.89)0.009Cys C(mg/dL), M (Q₁, Q₃)0.98 (0.86, 1.12)0.98 (0.86, 1.12)1.00 (0.87, 1.16)0.001HbA1c(mg/dL), M (Q₁, Q₃)5.20 (4.90, 5.40)5.10 (4.90, 5.40)5.20 (4.90, 5.60) <.001DBP(mmHg), M (Q₁, Q₃)74.50 (66.50, 83.00)74.00 (66.50, 82.50)78.00 (70.00, 87.00) <.001SBP(mmHg), M (Q₁, Q₃)126.00 (113.50, 141.00)125.50 (113.50, 140.00)134.00 (119.50, 150.50) <.001TC(mg/dL), M (Q₁, Q₃)191.75 (168.56, 216.11)190.98 (168.17, 215.72)195.23 (172.04, 220.75)0.002TG(mg/dL), M (Q₁, Q₃)104.43 (74.34, 150.45)103.54 (73.46, 148.68)115.05 (84.96, 164.61) <.001HDL-C(mg/dL), M (Q₁, Q₃)49.48 (40.59, 60.31)49.87 (40.98, 60.70)47.17 (39.43, 56.44) <.001LDL-C(mg/dL), M (Q₁, Q₃)115.59 (94.72, 138.40)115.21 (94.33, 137.63)119.85 (98.78, 141.88) <.001CRP(mg/dL), M (Q₁, Q₃)1.02 (0.55, 2.14)0.98 (0.54, 2.09)1.33 (0.68, 2.64) <.001UA(mg/dL), M (Q₁, Q₃)4.29 (3.56, 5.13)4.28 (3.56, 5.11)4.35 (3.56, 5.36)0.054FBG(mg/dL), M (Q₁, Q₃)102.24 (94.50, 112.32)102.06 (94.50, 111.96)104.40 (96.03, 116.73) <.001BMI 2011(kg/m2), M (Q₁, Q₃)23.14 (20.86, 25.74)23.06 (20.77, 25.62)24.02 (21.69, 26.80) <.001Height(kg), M (Q₁, Q₃)1.57 (1.52, 1.64)1.57 (1.52, 1.64)1.58 (1.52, 1.64)0.534Weight(kg), M (Q₁, Q₃)57.70 (50.80, 65.70)57.50 (50.70, 65.40)60.00 (53.00, 68.70) <.001WC(cm), M (Q₁, Q₃)84.80 (78.00, 92.00)84.20 (77.90, 91.60)88.60 (81.00, 96.00) <.001eGFR, M (Q₁, Q₃)94.46 (83.89, 102.00)94.66 (84.28, 102.28)92.48 (81.41, 99.17) <.001CHG, M (Q₁, Q₃)5.28 (5.03, 5.56)5.27 (5.03, 5.55)5.39 (5.15, 5.65) <.001TyG, M (Q₁, Q₃)8.59 (8.22, 9.02)8.57 (8.21, 9.01)8.74 (8.38, 9.13) <.001WHtR, M (Q₁, Q₃)0.54 (0.49, 0.59)0.53 (0.49, 0.58)0.56 (0.51, 0.60) <.001VAI, M (Q₁, Q₃)3.41 (2.04, 5.90)3.31 (2.00, 5.82)4.04 (2.49, 6.78) <.001ABSI, M (Q₁, Q₃)0.08 (0.08, 0.09)0.08 (0.08, 0.09)0.08 (0.08, 0.09) <.001BRI, M (Q₁, Q₃)4.05 (3.23, 5.10)4.00 (3.20, 5.04)4.54 (3.62, 5.52) <.001CVAI, M (Q₁, Q₃)94.96 (68.80, 125.28)92.98 (67.53, 122.77)112.43 (83.65, 140.29) <.001WWI, M (Q₁, Q₃)11.13 (10.59, 11.72)11.11 (10.57, 11.70)11.31 (10.80, 11.91) <.001Liver Disease, n(%)0.114 No7632 (96.55)6926 (96.65)706 (95.53) Yes273 (3.45)240 (3.35)33 (4.47)Kidney Disease, n(%)0.049 No7454 (94.29)6769 (94.46)685 (92.69) Yes451 (5.71)397 (5.54)54 (7.31)Heart Disease, n(%) <.001 No6988 (88.40)6391 (89.19)597 (80.78) Yes917 (11.60)775 (10.81)142 (19.22)Cancer, n(%)0.619 No7839 (99.17)7105 (99.15)734 (99.32) Yes66 (0.83)61 (0.85)5 (0.68)Diabetes, n(%) <.001 No7421 (93.88)6748 (94.17)673 (91.07) Yes484 (6.12)418 (5.83)66 (8.93)Hypertension, n(%) <.001 No5875 (74.32)5469 (76.32)406 (54.94) Yes2030 (25.68)1697 (23.68)333 (45.06)Sex, n(%)0.273 Female4259 (53.88)3875 (54.07)384 (51.96) Male3646 (46.12)3291 (45.93)355 (48.04)Residence, n(%)0.991 City2772 (35.07)2513 (35.07)259 (35.05) Village5133 (64.93)4653 (64.93)480 (64.95)Education, n(%)0.020 Below primary3996 (50.55)3602 (50.27)394 (53.32) Primary schools1705 (21.57)1531 (21.36)174 (23.55) Middle school1454 (18.39)1346 (18.78)108 (14.61) High school and above750 (9.49)687 (9.59)63 (8.53)Marital, n(%)0.075 Other972 (12.30)866 (12.08)106 (14.34) Married6933 (87.70)6300 (87.92)633 (85.66)Smoking, n(%)0.004 Never4855 (61.42)4433 (61.86)422 (57.10) Former680 (8.60)595 (8.30)85 (11.50) Now2370 (29.98)2138 (29.84)232 (31.39)Drinking, n(%)0.004 Never4644 (58.75)4235 (59.10)409 (55.35) Former646 (8.17)563 (7.86)83 (11.23) Now2615 (33.08)2368 (33.04)247 (33.42)CKM, n(%) <.001 0 stage543 (6.87)525 (7.33)18 (2.44) 1 stage1007 (12.74)958 (13.37)49 (6.63) 2 stage3465 (43.83)3189 (44.50)276 (37.35) 3 stage1980 (25.05)1726 (24.09)254 (34.37) 4 stage910 (11.51)768 (10.72)142 (19.22)CHG-CVAI quartile, n(%) <.001 Q11976 (25.00)1876 (26.18)100 (13.53) Q21976 (25.00)1831 (25.55)145 (19.62) Q31976 (25.00)1774 (24.76)202 (27.33) Q41977 (25.01)1685 (23.51)292 (39.51)CHG-WC quartile, n(%) <.001 Q11976 (25.00)1871 (26.11)105 (14.21) Q21976 (25.00)1829 (25.52)147 (19.89) Q31976 (25.00)1766 (24.64)210 (28.42) Q41977 (25.01)1700 (23.72)277 (37.48)Values are expressed as mean (standard deviation), median (quartile range), or n (%)TG Triglycerides, HDL-C High-Density Lipoprotein Cholesterol, LDL-C Low-Density Lipoprotein Cholesterol, hs-CRP High-Sensitivity C-Reactive Protein, FBG Fasting Blood Glucose, AIP Atherogenic Index of Plasma, CVAI Chinese Visceral Adiposity Index, ABSI A Body Shape Index, TyG Triglyceride-Glucose Index, UA Uric Acid, eGDR Estimated Glucose Disposal Rate, CKM Cardiovascular-Kidney-Metabolic, Age Age, BUN Blood Urea Nitrogen, Cr Creatinine, Cys C Cystatin C, HbA1c Hemoglobin A1c, DBP Diastolic Blood Pressure, SBP Systolic Blood Pressure, TC Total Cholesterol, CRP C-Reactive Protein, BMI Body Mass Index, WC Waist Circumference, CHG Cholesterol, high-density lipoprotein, and glucose index, WHtR Waist-to-Height Ratio, VAI Visceral Adiposity Index, BRI Body Roundness Index, WWI Weight-Waist Index Baseline characteristics of the study population, stratified by stroke status Values are expressed as mean (standard deviation), median (quartile range), or n (%) TG Triglycerides, HDL-C High-Density Lipoprotein Cholesterol, LDL-C Low-Density Lipoprotein Cholesterol, hs-CRP High-Sensitivity C-Reactive Protein, FBG Fasting Blood Glucose, AIP Atherogenic Index of Plasma, CVAI Chinese Visceral Adiposity Index, ABSI A Body Shape Index, TyG Triglyceride-Glucose Index, UA Uric Acid, eGDR Estimated Glucose Disposal Rate, CKM Cardiovascular-Kidney-Metabolic, Age Age, BUN Blood Urea Nitrogen, Cr Creatinine, Cys C Cystatin C, HbA1c Hemoglobin A1c, DBP Diastolic Blood Pressure, SBP Systolic Blood Pressure, TC Total Cholesterol, CRP C-Reactive Protein, BMI Body Mass Index, WC Waist Circumference, CHG Cholesterol, high-density lipoprotein, and glucose index, WHtR Waist-to-Height Ratio, VAI Visceral Adiposity Index, BRI Body Roundness Index, WWI Weight-Waist Index Feature selection O identify and integrate relevant covariates into the predictive framework, a systematic variable screening protocol was executed. The refinement process began with a univariate analysis, where a significance threshold of P < 0.05 (refer to Table S2) was utilized as the primary criterion for variable retention. The initial variable set was selected based on the intersection of the top 15 variables identified by the Boruta algorithm and RFE (Fig. 2), combined with prior domain-specific knowledge. AUC value for the tree-based model using Boruta to distinguish stroke events was 0.661 (95% CI: 0.642–0.680) after tenfold cross-validation; AUC value for the logistic regression model using RFE was 0.664 (95% CI: 0.644–0.684). These metrics indicate that the base models possess good generalization ability, and the feature importance rankings are stable and reliable (Figure S1). After further excluding collinear variables (Figure S2), the final model incorporated the following 12 variables: Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, and eGFR.Fig. 2Feature selection process. A Variable importance ranking using the Boruta algorithm. B Falling process of the Boruta algorithm. C A ridge plot illustrates the feature importance of RFE Feature selection process. A Variable importance ranking using the Boruta algorithm. B Falling process of the Boruta algorithm. C A ridge plot illustrates the feature importance of RFE Incremental predictive value of CHG obesity-derived index and TyG-related index The incremental predictive value of seven CHG-based obesity markers (including CHG-BMI, CHG-WC, CHG-WHtR, CHG-BRI, CHG-ABSI, CHG-WWI, and CHG-CVAI) was rigorously examined in the current study. As illustrated in Fig. 3A, each evaluated indicator notably augmented the forecasting capacity of the initial reference model. This enhanced predictive accuracy was supported by strong statistical evidence, as most P-values achieved a significance level of < 0.001. Among them, CHG-CVAI and CHG-WC demonstrated optimal predictive performance, both with AUC values of 0.664. The 95% confidence interval ranges for CHG-CVAI and CHG-WC were 0.658–0.668 and 0.657–0.668, respectively. The NRI for CHG-WC and CHG-CVAI were 0.139 (95% CI: 0.072, 0.210) and 0.125 (95% CI: 0.048, 0.186), respectively. Furthermore, the two best-performing predictors, CHG-CVAI and CHG-WC, were compared with TyG and its multiple derivatives. Both CHG-CVAI and CHG-WC demonstrated superior predictive value compared to TyG-related indicators (Fig. 3B). The entire incremental prediction metric results are shown in Table S3. Furthermore, regardless of whether covariates were unadjusted or adjusted, CHG-CVAI consistently yielded the highest OR, followed by CHG-WC. The details are presented in Table S4.Fig. 3Incremental Predictive Value of the CHG Obesity-Derived Index and Comparison with the TyG Obesity-Derived Index. A ROC curves show the incremental predictive values of CHG and its obesity-derived indicators. B ROC curves show the incremental predictive values of the two best-performing CHG obesity-derived indicators combined with the TyG obesity-derived indicator Incremental Predictive Value of the CHG Obesity-Derived Index and Comparison with the TyG Obesity-Derived Index. A ROC curves show the incremental predictive values of CHG and its obesity-derived indicators. B ROC curves show the incremental predictive values of the two best-performing CHG obesity-derived indicators combined with the TyG obesity-derived indicator Association of CHG-CVAI and CHG-WC with stroke As shown in Table 2, both CHG-CVAI and CHG-WC were positively associated with stroke risk. In the unadjusted model, each unit increase in CHG-CVAI (OR = 1.45, 95% CI: 1.35–1.56) and CHG-WC (OR = 1.43, 95% CI: 1.30–1.59) was associated with a 45% and 43% increase in stroke risk, respectively. After quartile stratification, both indicators demonstrated a clear dose–response relationship. After full adjustment for confounders (Model 3), the highest quartile (Q4) of CHG-CVAI showed a 2.44-fold higher stroke risk (95% CI: 1.66–3.60) compared to the lowest quartile (Q1), with a continuous variable OR of 1.30 (95% CI: 1.13–1.49). The association strength for CHG-WC was slightly lower, with the Q4 group exhibiting a 2.06-fold higher risk than the Q1 group (95% CI: 1.43–2.95) and a continuous variable OR of 1.21 (95% CI: 1.01–1.44). Overall, CHG-CVAI demonstrated a stronger association with stroke risk.Table 2Association of CHG-CVAI and CHG-WC with strokeVariablesModel1Model2Model3OR (95%CI)POR (95%CI)POR (95%CI)PCHG-CVAI Quartile11.00 (Reference)1.00 (Reference)1.00 (Reference) Quartile21.49 (1.14 ~ 1.93)0.0031.50 (1.15 ~ 1.96)0.0031.43 (1.08 ~ 1.89)0.014 Quartile32.14 (1.67 ~ 2.74) <.0012.16 (1.68 ~ 2.79) <.0011.90 (1.39 ~ 2.59) <.001 Quartile43.25 (2.57 ~ 4.12) <.0013.14 (2.47 ~ 4.00) <.0012.44 (1.66 ~ 3.60) <.001 Continuous1.45 (1.35 ~ 1.56) <.0011.46 (1.36 ~ 1.57) <.0011.30 (1.13 ~ 1.49) <.001CHG-WC Quartile11.00 (Reference)1.00 (Reference)1.00 (Reference) Quartile21.43 (1.11 ~ 1.85)0.0061.48 (1.14 ~ 1.91)0.0031.37 (1.05 ~ 1.80)0.023 Quartile32.12 (1.66 ~ 2.70) <.0012.19 (1.71 ~ 2.80) <.0011.81 (1.35 ~ 2.42) <.001 Quartile42.90 (2.30 ~ 3.67) <.0012.98 (2.35 ~ 3.78) <.0012.06 (1.43 ~ 2.95) <.001 Continuous1.43 (1.30 ~ 1.59) <.0011.40 (1.26 ~ 1.55) <.0011.21 (1.01 ~ 1.44)0.033OR Odds Ratio, CI Confidence IntervalModel1: CrudeModel2: Adjust:Age, Sex, Smoking, DrinkingModel3: Adjust:Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, eGFR Table 2Association of CHG-CVAI and CHG-WC with strokeVariablesModel1Model2Model3OR (95%CI)POR (95%CI)POR (95%CI)PCHG-CVAI Quartile11.00 (Reference)1.00 (Reference)1.00 (Reference) Quartile21.49 (1.14 ~ 1.93)0.0031.50 (1.15 ~ 1.96)0.0031.43 (1.08 ~ 1.89)0.014 Quartile32.14 (1.67 ~ 2.74) <.0012.16 (1.68 ~ 2.79) <.0011.90 (1.39 ~ 2.59) <.001 Quartile43.25 (2.57 ~ 4.12) <.0013.14 (2.47 ~ 4.00) <.0012.44 (1.66 ~ 3.60) <.001 Continuous1.45 (1.35 ~ 1.56) <.0011.46 (1.36 ~ 1.57) <.0011.30 (1.13 ~ 1.49) <.001CHG-WC Quartile11.00 (Reference)1.00 (Reference)1.00 (Reference) Quartile21.43 (1.11 ~ 1.85)0.0061.48 (1.14 ~ 1.91)0.0031.37 (1.05 ~ 1.80)0.023 Quartile32.12 (1.66 ~ 2.70) <.0012.19 (1.71 ~ 2.80) <.0011.81 (1.35 ~ 2.42) <.001 Quartile42.90 (2.30 ~ 3.67) <.0012.98 (2.35 ~ 3.78) <.0012.06 (1.43 ~ 2.95) <.001 Continuous1.43 (1.30 ~ 1.59) <.0011.40 (1.26 ~ 1.55) <.0011.21 (1.01 ~ 1.44)0.033OR Odds Ratio, CI Confidence IntervalModel1: CrudeModel2: Adjust:Age, Sex, Smoking, DrinkingModel3: Adjust:Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, eGFR Association of CHG-CVAI and CHG-WC with stroke OR Odds Ratio, CI Confidence Interval Model1: Crude Model2: Adjust:Age, Sex, Smoking, Drinking Model3: Adjust:Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, eGFR Cumulative exposure and longitudinal trajectory analysis To explore how temporal variations in indicators influence outcomes,cumulative exposure and trajectory models were applied to the dataset spanning from baseline to the two 2015 measurements (Table 3). Regarding the examination of associations between CHG-CVAI, CHG-WC and stroke, the variance inflation factors for the fully adjusted models were less than 5 in both the baseline analysis and the cumulative exposure and longitudinal trajectory analyses (Table S5). The cumulative exposure regression analysis results indicate that cumulative CHG-CVAI (cuCHG-CVAI) and cumulative CHG-WC (cuCHG-WC) were independently associated with stroke risk, with ORs of 1.35 (95% CI: 1.11–1.65) and 1.21 (95% CI: 1.01–1.44), respectively, in the fully adjusted model. Using K-means clustering, three clinically significant longitudinal trajectory patterns were identified (Fig. 4). For both the CHG-CVAI and CHG-WC trajectories, cluster 3 exhibited higher systolic blood pressure, higher HbA1c, higher C-reactive protein, and higher body weight compared with cluster 1. Furthermore, the prevalence of heart disease and the proportion of smokers and drinkers were significantly elevated in cluster 3 than in cluster 1. These characteristics are consistent with the known phenotype of high-risk populations for stroke (Table S6, Table S7). The longitudinal trajectory logistic regression results showed that individuals in the cluster 2 and cluster 3 trajectory groups had a significantly higher risk of stroke than those in the cluster 1 trajectory group. Specifically, the CHG-CVAI cluster 3 group exhibited the highest risk (OR = 1.98, 95% CI: 1.29–3.04), whereas OR for the CHG-WC cluster 3 group was 1.75 (95% CI: 1.26–2.42). This finding indicates that sustained high levels of CHG-derived indices over time serve as strong predictors of future stroke events.Table 3Association between cumulative exposure to CHG-CVAI and CHG-WC and their longitudinal trajectories with strokeVariablesModel1Model2Model3OR (95%CI)POR (95%CI)POR (95%CI)PcuCHG CVAI Cluster 11.00 (Reference)1.00 (Reference)1.00 (Reference) Cluster 21.86 (1.40 ~ 2.46) <.0011.82 (1.37 ~ 2.42) <.0011.58 (1.15 ~ 2.17)0.005 Cluster 32.89 (2.16 ~ 3.87) <.0012.71 (2.01 ~ 3.65) <.0011.98 (1.29 ~ 3.04)0.002 Continuous1.44 (1.30 ~ 1.60) <.0011.44 (1.30 ~ 1.60) <.0011.35 (1.11 ~ 1.65)0.003cuCHG WC Cluster 11.00 (Reference)1.00 (Reference)1.00 (Reference) Cluster 21.58 (1.20 ~ 2.09)0.0011.62 (1.22 ~ 2.15) <.0011.43 (1.07 ~ 1.91)0.014 Cluster 32.40 (1.78 ~ 3.23) <.0012.41 (1.78 ~ 3.26) <.0011.75 (1.26 ~ 2.42) <.001 Continuous1.43 (1.30 ~ 1.59) <.0011.40 (1.26 ~ 1.55) <.0011.21 (1.01 ~ 1.44)0.033OR Odds Ratio, CI Confidence IntervalModel1: CrudeModel2: Adjust:Age, Sex, Smoking, DrinkingModel3: Adjust:Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, eGFRFig. 4Clustering Analysis of Longitudinal Trajectories for CHG-CVAI and CHG-WC. A Scatter plot of the cluster analysis. B Line chart of the cluster analysis. C Density map of CHG-CVAI and CHG-WC in 2012. D Density map of CHG-CVAI and CHG-WC in 2015 Table 3Association between cumulative exposure to CHG-CVAI and CHG-WC and their longitudinal trajectories with strokeVariablesModel1Model2Model3OR (95%CI)POR (95%CI)POR (95%CI)PcuCHG CVAI Cluster 11.00 (Reference)1.00 (Reference)1.00 (Reference) Cluster 21.86 (1.40 ~ 2.46) <.0011.82 (1.37 ~ 2.42) <.0011.58 (1.15 ~ 2.17)0.005 Cluster 32.89 (2.16 ~ 3.87) <.0012.71 (2.01 ~ 3.65) <.0011.98 (1.29 ~ 3.04)0.002 Continuous1.44 (1.30 ~ 1.60) <.0011.44 (1.30 ~ 1.60) <.0011.35 (1.11 ~ 1.65)0.003cuCHG WC Cluster 11.00 (Reference)1.00 (Reference)1.00 (Reference) Cluster 21.58 (1.20 ~ 2.09)0.0011.62 (1.22 ~ 2.15) <.0011.43 (1.07 ~ 1.91)0.014 Cluster 32.40 (1.78 ~ 3.23) <.0012.41 (1.78 ~ 3.26) <.0011.75 (1.26 ~ 2.42) <.001 Continuous1.43 (1.30 ~ 1.59) <.0011.40 (1.26 ~ 1.55) <.0011.21 (1.01 ~ 1.44)0.033OR Odds Ratio, CI Confidence IntervalModel1: CrudeModel2: Adjust:Age, Sex, Smoking, DrinkingModel3: Adjust:Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, eGFR Association between cumulative exposure to CHG-CVAI and CHG-WC and their longitudinal trajectories with stroke OR Odds Ratio, CI Confidence Interval Model1: Crude Model2: Adjust:Age, Sex, Smoking, Drinking Model3: Adjust:Age, Sex, Smoking, Drinking, Heart Disease, Weight, SBP, CRP, HDL-C, Cr, HbA1c, eGFR Clustering Analysis of Longitudinal Trajectories for CHG-CVAI and CHG-WC. A Scatter plot of the cluster analysis. B Line chart of the cluster analysis. C Density map of CHG-CVAI and CHG-WC in 2012. D Density map of CHG-CVAI and CHG-WC in 2015 Dose–response relationship RCS analysis demonstrated (Fig. 5) that, after fully adjusting for covariates, both baseline estimates of CHG-CVAI and CHG-WC and their cumulative exposure estimates exhibited a linear dose–response relationship with stroke risk (P-value for nonlinearity > 0.05).Fig. 5Dose–response relationship of stroke risk with baseline/cumulative CHG-CVAI and CHG-WC using restricted cubic splines Dose–response relationship of stroke risk with baseline/cumulative CHG-CVAI and CHG-WC using restricted cubic splines Subgroup analysis Subgroup analysis indicated (Fig. 6) a consistent positive association between CHG-CVAI and CHG-WC with stroke risk across the majority of subgroups. However, interaction analysis revealed significant heterogeneity regarding the effect of CHG-CVAI across different age groups (P-interaction = 0.024), with a more pronounced association identified in individuals < 60 years old (OR = 1.49) compared to those > = 60 years old (OR = 1.30). Hypertensive status also functioned as a significant effect modifier (P-interaction < 0.01). Both CHG-CVAI and CHG-WC demonstrated significant predictive value in non-hypertensive individuals (OR 1.69 and 1.55, respectively), whereas neither association reached statistical significance in those diagnosed with hypertension. Furthermore, neither CHG-CVAI nor CHG-WC show significant association with stroke risk in diabetic patients. Moreover, meaningful interactions were not observed in other subgroups such as gender, education level, smoking/drinking status, or different CKM stages (P-interaction > 0.05), thereby indicating relatively consistent predictive effects across these populations.Fig. 6Subgroup Analysis. A CHG-CVAI; B CHG-WC Subgroup Analysis. A CHG-CVAI; B CHG-WC Sensitivity analysis For the purpose of verifying the findings' reliability, four categories of sensitivity evaluations were conducted. Initially, rerunning the logistic regression analysis on the raw dataset prior to multiple imputation yielded findings demonstrating high concordance with the main analysis (Table S8). For the next phase of the sensitivity check, the associations were re-evaluated through a Cox proportional hazards model. By considering stroke incidence as time-to-event data, the model still showed significantly higher hazard ratios (HR) for the highest quartile groups of CHG-CVAI and CHG-WC (2.20 and 1.88, respectively), thereby supporting the reliability of the primary analysis (Table S9). Furthermore, conducting the analysis after removing subjects who presented with stroke during the 2020 follow-up period showed stable or even slightly strengthened associations (Table S10). This result confirms that the findings were not influenced by late events. Lastly, to prevent issues related to over-adjustment, HDL-C and body weight, which are constituents of the CHG-related indicators, were not included in the model. The resulting estimates exhibited high consistency with those of the main analysis. This indicates that the association between the CHG-derived indicators and stroke was not driven by multicollinearity or over-adjustment, and that their predictive value was robust (Table S11). Discussion His study found that among several obesity-derived indicators based on CHG, CHG-CVAI and CHG-WC demonstrated optimal predictive values for the incidence of stroke within mid-aged and senior populations. These two composite indicators exhibited superior prognostic performance compared to conventional CHG and TyG-related indicators. Moreover, their positive association with stroke risk remained robust even following rigorous adjustment for various confounding factors and exhibited a pronounced dose–response relationship. Furthermore, longitudinal analysis results showed that individuals with higher cumulative exposure levels or trajectory patterns characterized by persistent elevated levels of the indicators exhibited significantly increased stroke risk. Collectively, these findings indicate that combining CHG—a blood-based indicator reflecting glucose and lipid metabolism—with anthropometric measures like CVAI and WC, which precisely reflect central obesity and visceral fat distribution, establishes a more efficient, convenient, and comprehensive risk assessment tool. This approach provides new evidence-based support for primary stroke prevention. In recent years, composite metabolic indicators have garnered significant attention because of their ability to integrate information from multiple pathophysiological pathways [34–37]. TyG is a reliable alternative biomarker for IR and has exhibited substantial ties to CVD as well as stroke incidence [38, 39]. This study also validated how TyG and linked obesity-based indicators performed predictively in the same cohort. Confirming the observations documented by Wang et al. [27], TyG-WC and TyG-CVAI stood out for their exceptional performance among TyG obesity-derived indices. However, the newly introduced CHG series of indices exhibited comparable or even slightly superior predictive efficacy. Mansoori et al. [11] defined the CHG using three components: FBG, TC, and HDL-C and demonstrates superior discriminatory ability to TyG in diagnosing type 2 diabetes. Mo et al. [18] first validated CHG in CVD risk prediction and reported a strong linear association of CHG with CVD risk and higher hazard ratios (HR) compared to TyG. This study focused specifically on stroke as an outcome and developed novel obesity-derived indices. Among these, CHG-CVAI and CHG-WC exhibited the highest AUC value (0.664) across all derived indices, surpassing both TyG and its derivatives. Compared with TyG, simultaneous integration of TC and HDL-C, two traditional lipid parameters directly associated with atherosclerosis is a distinctive feature of CHG. The TC/HDL-C ratio itself is a core component of the atherosclerosis index and is closely linked with inflammation and endothelial dysfunction; an elevated TC/HDL-C ratio reflects an increased pro-atherosclerotic lipid burden and impaired cholesterol reverse transport [40]. Therefore, the findings indicate that although CHG and TyG have similar predictive performance when used alone, combining CHG with central obesity indicators that reflect body fat distribution yields a notable improvement in prediction performance for the resulting hybrid indices. Specifically, CHG-WC and CHG-CVAI demonstrate this enhanced predictive ability. These composite indices also demonstrate superior predictive performance compared to corresponding TyG-derived metrics. This indicates a significant synergistic effect between the pathophysiological information included in CHG and the metabolic and morphological abnormalities represented by obesity. It suggests that CHG has greater potential for integrating multidimensional risk information, enabling more effective identification of high-risk individuals concurrently exhibiting abnormalities in glucose and lipid metabolism alongside visceral fat accumulation. Furthermore, although the CHG-derived indicators show only a limited improvement in AUC compared with TyG-related indicators, they perform better in terms of NRI. This suggests that, in clinical practice, they can more effectively reclassify certain individuals in the moderate-risk group as either high-risk or low-risk, thereby guiding adjustments to intervention strategies. Consequently, this improvement holds potential clinical significance, especially in primary care screening settings where resources are limited. This study found that CHG-CVAI and CHG-WC demonstrate comparable predictive capabilities, a phenomenon warranting further investigation. It suggests that composite index design does not simply involve linear stacking of parameters to enhance efficacy, but rather involves biological synergy and complementarity of the selected components. As a visceral fat measurement specifically developed for Chinese individuals, CVAI integrates age, WC, BMI, TG, and HDL-C [41]. Consequently, CHG-CVAI functions as an extensively synthesized index incorporating WC, TG, HDL-C, age, and BMI components from CVAI. Furthermore, TG and HDL-C appear redundantly in both components, effectively amplifying the weight of these two critical lipid abnormality signals. This design endows CHG-CVAI with high specificity for reflecting visceral fat accumulation, associated atherogenic dyslipidemia (increased TG along with reduced HDL-C), and abnormalities in glycemic control. In contrast, CHG-WC features a simpler composition of the CHG index and WC. Previous studies have shown that WC is the most classic, direct, and easily measurable indicator for assessing central obesity, exhibiting strong correlation with visceral fat area [42, 43]. The CHG-WC combination directly reflects the synergistic effect of abnormal blood metabolic markers and abnormal body fat distribution. The similar performances of these two composite indices may be because of the following two reasons: First, WC is a core component of CVAI, and both measures are highly consistent in capturing central obesity as a key risk dimension. Second, although CVAI incorporates more variables, CHG already includes TC and HDL-C, resulting in partial information overlap. When CHG and CVAI are combined, TC and HDL-C are counted twice. This may not proportionally increase predictive value but could instead dilute the contribution of other unique factors, such as FBG. In contrast, the CHG-WC combination is simpler and amplifies the interaction between “metabolic abnormalities” and “abdominal obesity”—two independent, strong risk factors. This suggests that when constructing composite indicators, identifying core elements with independent predictive value and strong pathophysiological linkage for combination may be more effective than simply increasing the number of variables. Moreover, concise and efficient combinations may be more conducive to clinical translation and application [44, 45]. This study examined not only baseline values but also identified trajectories through cumulative exposure calculations and K-means clustering. Individuals demonstrating long-term high CHG-CVAI or high CHG-WC trajectories exhibited nearly double the stroke risk of those with persistently low-level trajectories. This finding aligns with recent studies [46–48]. This suggests that stroke occurs because of the cumulative effects of prolonged, persistent metabolic and adipose tissue-related damage [49–51]. Single measurements are influenced by short-term fluctuations or merely represent a “snapshot” at a specific time point, whereas longitudinal monitoring more reliably identifies individuals chronically exposed to high-risk metabolic environments [52]. Therefore, incorporating CHG-WC or CHG-CVAI as routine monitoring indicators in health management—observing their trend changes—may provide greater utility than relying exclusively on static measurements taken at one specific moment for the proactive detection of susceptible cohorts and prompt clinical action. In subgroup analyses, stroke risk was positively associated with CHG-CVAI and CHG-WC, especially pronounced in non-hypertensive individuals but significantly attenuated or even insignificant in patients with established hypertension. This is because hypertension is characterized as a paramount and direct autonomous contributor to stroke occurrence [53]. In those with established hypertension, the vascular system has already endured sustained high-pressure loads, leading to arterial stiffening, endothelial damage, and vascular remodeling, resulting in a significantly elevated baseline stroke risk [54]. At this stage, the additional risk represented by composite metabolic-obesity indicators may be diluted or masked relative to the significant risk posed by hypertension itself. Furthermore, patients diagnosed with hypertension are more likely to receive regular medical follow-up and interventions, including lifestyle guidance and antihypertensive medication. Some of these interventions also demonstrate vascular protection, anti-inflammatory effects, and metabolic improvements in addition to reducing blood pressure [55–57]. Consequently, the hypertension cohort may represent a partially medically managed population where metabolic abnormalities like blood glucose and lipids are already controlled or modified. This could weaken the baseline association between metabolic-obesity indicators and long-term stroke outcomes. Therefore, clinicians should rely more on blood pressure control itself and comprehensive risk assessment models rather than on any single metabolic indicator alone in those with a diagnosis of hypertension. Within this investigation, the AUC value for stroke risk prediction using the CHG-CVAI and CHG-WC was approximately 0.664, which is considered moderate. Although this degree of discrimination does not reach the levels of certain integrated medical tools (exemplified by the Framingham Stroke Risk Score), it is worth noting that this model was constructed solely based on routine physical examination indicators and did not incorporate strong predictors such as blood pressure or atrial fibrillation. Furthermore, the improvement in AUC was achieved on top of traditional risk factors. Furthermore, both the NRI and IDI displayed meaningful statistical gains. This finding implies that the model established in the present research possesses independent incremental predictive value. In terms of clinical applicability, calculation of CHG-CVAI and CHG-WC requires only routine physical examination data and does not incur additional testing costs. Therefore, it offers good accessibility and scalability. Although various comprehensive risk assessment tools are currently available, many of these require collection of multiple variables that are sometimes difficult to implement at the primary care level. CHG-derived indices can serve as screening tools. Once high-risk individuals have been identified, they can be managed through stratification in conjunction with more comprehensive assessment tools. Strengths and limitations Several distinct benefits are offered by this investigation. First, the research design was innovative. It systematically constructed and compared the predictive value of the CHG and a derivative index combining seven obesity indicators for stroke risk using the nationwide prospective cohort CHARLS for the first time. It also directly compared these indices with the established TyG derivative indicator and provided new evidence for selecting metabolic composite indicators. Methodologically, this study used a multidimensional analytical framework encompassing multivariable-adjusted models, RCS analysis, longitudinal cumulative exposure, and trajectory clustering analysis. The incremental predictive value was calculated for each index and the conclusions were validated through multiple sensitivity analyses and demonstrated robust reliability. Furthermore, computation of CHG-CVAI and CHG-WC is simple and requires only routine physical examination data at low cost. Therefore, these indices are highly suitable for primary care settings. This study emphasizes their longitudinal monitoring value in identifying risk evolution trajectories, enabling “dynamic management” for prevention. Subgroup analyses further demonstrate high predictive efficacy in non-hypertensive populations, thereby facilitating “upstream” precision screening. These advantages collectively ensure the reliability of our findings and highlight their direct potential for clinical translation and public health applications. Certain constraints should be acknowledged in the present work. First, the nature of this research is observational. Despite extensive adjustment for confounding factors and multiple sensitivity analyses, the possibility of residual confounding persists, nor can causality be definitively confirmed. Second, stroke diagnosis relied on participant self-reporting. Although standardized questionnaires were used and confirmed during follow-up, recall bias or misclassification cannot be avoided, thereby potentially affecting the precision of outcome measurements [58]. Third, the data did not distinguish between stroke subtypes such as ischemic and hemorrhagic strokes, which may involve different risk factors. Fourth, since the investigated cohort consisted of middle-aged and older Chinese adults, the external validity regarding younger cohorts or other ethnic groups may be compromised. Fifth, cumulative exposure and trajectories were defined using only baseline and 2015 data points. Incorporating more frequent follow-up data at shorter intervals in future studies could provide a more refined analysis of risk. Sixth, the K-means clustering method has inherent limitations compared to GBTM. Since this study only utilized data from two waves of follow-up, the K-means clustering method represents a viable option for identifying dynamic patterns in CHG-derived indicators despite data constraints. However, future studies with more waves of follow-up data at a higher frequency should adopt GBTM or latent class growth mixture models to investigate their evolutionary trajectories and complex associations with stroke risk in greater detail. Seventh, because of limitations with the availability of the CHARLS database, this study was unable to include more detailed lipid markers, such as lipoprotein particle concentrations (LDL-p, HDL-p) or apolipoprotein B (ApoB). Recent studies have reported that LDL-p is more closely associated with atherosclerosis and cerebrovascular events than LDL-C [59, 60]; moreover, these markers are more accurate in reflecting the burden of atherogenic lipoproteins. CHG-derived indices may not fully capture lipid-related stroke risk. Therefore, future studies should validate these indices using databases containing the aforementioned markers and optimize the predictive performances of the composite. Finally, this study was the first to demonstrate optimal predictive performances for CHG-CVAI and CHG-WC based on baseline ROC curves. Subsequently, an in-depth longitudinal analysis of these two indicators was conducted. However, indicators with strong baseline predictive power may not accurately reflect the etiological contribution to outcomes through their dynamic changes. Because of space constraints and the need to focus on the analysis, this study did not conduct a complete longitudinal analysis of all seven indicators individually. This may have resulted in the omission of some indicators with significant etiological value. Conclusions The present study, based on CHARLS prospective cohort data, establishes that CHG-CVAI and CHG-WC stand out among all CHG-based obesity indices for their superior ability to forecast stroke risk in middle-aged and older adults. Both metrics exhibit a positive linear relationship with stroke occurrence, outperforming established TyG-related indicators in predictive accuracy. Therefore, maintaining and tracking ideal CHG-CVAI or CHG-WC levels in both clinical and community settings is effective for identifying those at elevated risk, thereby enabling timely and targeted interventions. These findings support the incorporation of these simple, low-cost composite indices into routine health assessments to strengthen primary stroke prevention. In the future research studies, these indicators must be integrated into risk stratification models to evaluate their role in guiding personalized prevention strategies, thereby helping to reduce the overall burden of stroke in an aging society. Supplementary Information Supplementary Material 1. Supplementary Material 1.