Through the analysis of data from the China Health and Retirement Longitudinal Study (CHARLS), this study aims to clarify the relationship between metabolic and vascular dysfunction in stroke risk prediction. Specifically, we seek to investigate how the combined TyG-ePWV index can improve stroke risk prediction. Using advanced statistical methodologies, including multivariable Cox proportional hazards models and Kaplan-Meier survival analysis, this research will provide novel insights into the independent and synergistic contributions of TyG and ePWV in stroke risk stratification. Ultimately, these findings may contribute to the development of more effective strategies for stroke prevention.
This study utilized data from the CHARLS, a nationally representative survey that tracks individuals aged 45 and older over time. Conducted by the National School of Development at Peking University, CHARLS comprises five survey waves carried out between 2011 and 2020. For our analysis, we focused on participants from the 2011 baseline cohort (n = 17,708) and followed them across four subsequent waves in 2013, 2015, 2018, and 2020. Data collection involved structured interviews and standardized questionnaires administered by trained personnel, capturing information on socio-demographic characteristics, Lifestyle habits, and health-related parameters. Participants were excluded if they lacked baseline blood test results, did not participate in follow-ups, were below 45 years of age, or had a pre-existing stroke diagnosis at baseline. Ultimately, a total of 8,444 individuals were included in the final analytical sample (Fig. 1). The CHARLS study protocol received ethical clearance from the Biomedical Ethics Review Board of Peking University (IRB00001052-11,015), and all participants provided written informed consent. The study procedures complied with the ethical standards set forth in the Declaration of Helsinki.
Venous blood samples were collected by nursing professionals, initially stored at -20 °C, and subsequently transferred to a deep freezer at -80 °C following the necessary analyses. Fasting plasma glucose (FPG) and triglyceride (TG) concentrations were determined using an enzyme colorimetric method in the laboratory. TyG index was calculated using the formula: ln (TG [mg/dL] × FBG [mg/dL]/2). The ePWV was derived using mean blood pressure (MBP) and age through the following equation: ePWV = 9.587-0.402 × age + 4.560 × 10 × age – 2.621 × 10 × age × MBP + 3.176 × 10 × age × MBP – 1.832 × 10 × MBP. MBP was determined using diastolic blood pressure (DBP) and systolic blood pressure (SBP) based on the formula: MBP = DBP + 0.4 × (SBP – DBP). To evaluate the risk of stroke during follow-up, ePWV and TyG values were categorized into two groups based on the median, allowing for analysis as a categorical variable. TyG-ePWV index was calculated by multiplying TyG by ePWV: TyG-ePWV = TyG (mg/dL) × ePWV (m/s)/10.
The cumulative TyG-ePWV (CumTyG-ePWV) was derived using the formula: Cum TyG-ePWV = [(TyG-ePWV + TyG-ePWV)/2] × (2015 — 2012). We determined the optimal cut-off values for baseline TyG-ePWV and CumTyG-ePWV to classify participants into high- and low-exposure groups (Fig. 2, Figure S1). To evaluate longitudinal patterns of TyG-ePWV exposure, K-means clustering analysis was employed using TyG-ePWV values measured at wave 1 (2011) and wave 3 (2015) as input variables. Each participant was characterized by a two-dimensional vector: [TyG-ePWV_wave1, TyG-ePWV_wave3], allowing the algorithm to identify distinct groups based on their baseline levels and trajectory patterns over the 4-year period. The optimal number of clusters was identified as 3 based on the elbow method and the silhouette coefficient approach (Fig. 2, Figure S1).
The main outcome of this study was the occurrence of stroke. In line with previous studies, stroke events were identified using a standardized question:”Did your doctor tell you that you were diagnosed with a stroke?” The timing of stroke onset was established based on the period between the most recent interview date and the interview when the stroke was first reported.
The study considered several covariates, including gender, baseline age, and marital status, which was categorized as either ‘married’ or ‘other.’ Educational attainment was classified into three levels — ‘Less than upper secondary education,’ ‘Upper secondary,’ and ‘Tertiary education’ — based on the number of years of schooling. Residential status was defined as either ‘urban’ or ‘rural.’ Lifestyle factors such as current smoking, alcohol consumption, and physical activity were recorded as binary variables (‘yes’ or ‘no’). Health-related factors encompassed loneliness, self-reported diagnoses of hypertension, diabetes, heart disease, dyslipidemia, and chronic kidney disease, along with the use of antihypertensive or antidiabetic medications. Depression was evaluated using the CESD-10 scale in accordance with prior studies. Additionally, blood test parameters included high-sensitivity C-reactive protein (CRP) levels.
Baseline characteristics across the four combination groups of TyG and ePWV were examined using appropriate statistical methods based on data distribution. These included analysis of variance (ANOVA) for normally distributed continuous variables, chi-square tests for categorical variables, and the Kruskal-Wallis rank-sum test for non-normally distributed data. Continuous variables were summarized as means with standard deviations, while categorical variables were expressed as percentages for each group. To address missing data, multiple imputation was conducted using the mice package to enhance the robustness and validity of the analyses.
Kaplan-Meier survival curves were utilized to estimate and compare the cumulative incidence of stroke across the groups throughout the follow-up period. The risk of stroke was assessed using multivariable Cox proportional hazards models, which estimated hazard ratios (HRs) and 95% confidence intervals (CIs) for TyG, ePWV, and TyG-ePWV. Both unadjusted and adjusted models were employed, with the latter accounting for all covariates. In the Cox proportional hazards models, time zero was defined as the baseline interview date. Follow-up time for each participant was calculated as the duration from the baseline interview date to either the date of stroke occurrence or the date of censoring. For participants who experienced stroke during follow-up, the event time was determined based on the interview date when stroke was first reported. The proportional hazards assumption was assessed using Schoenfeld residuals, and the cox.zph() function was applied to evaluate both individual covariates and the overall model. The analysis confirmed no significant violations of the proportional hazards assumption, as all P-values exceeded 0.05, indicating the model’s validity.
For multiplicative interactions, a product term combining TyG and ePWV indicators was included in the Cox model. For additive interactions, three metrics were calculated using the delta method: the interaction contrast ratio (ICR), also referred to as the relative excess risk due to interaction (RERI); the attributable proportion (AP); and the synergy index (SI). A cross-lagged panel model was employed to examine the bidirectional relationship between TyG and ePWV over time. In this model, we adjusted for time-varying covariates, including age, comorbid conditions (hypertension, diabetes, and heart disease), medication use. This adjustment aimed to account for potential confounding effects due to the dynamic nature of these covariates throughout the study period. To evaluate the mediating effects of TyG and ePWV, we dichotomized baseline TyG and ePWV at their median values. Regression models were then used to assess the direct and indirect effects of high TyG and high ePWV on stroke risk.
The predictive ability of different indices was assessed through receiver operating characteristic (ROC) curve analysis. The added predictive value between groups was evaluated using the Net Reclassification Index (NRI) and the Integrated Discriminant Improvement Index (IDI). To determine the contribution of TyG, ePWV, and their combined measure (TyG-ePWV) in predicting stroke risk, their relative importance was analyzed alongside conventional risk factors by calculating the R values from Cox proportional hazards models. Furthermore, the explainable log-likelihood for each risk factor was computed to ensure the robustness and consistency of the findings.
To ensure the robustness of our results and to explore potential differences, we performed several sensitivity analyses: (1) For the Cox regression analysis, the TyG, ePWV, and TyG-ePWV indices were reclassified into new groupings. (2) To ensure a thorough and reliable assessment of the associations, analyses were performed using datasets with complete covariate information. (3) Stratified analyses were conducted based on sex (male and female), age categories (< 60 and ≥ 60 years), and the presence of diabetes, hypertension, heart disease, and loneliness. (4) A 3-knot restricted cubic spline (RCS) model was applied to investigate the dose-response relationship and assess the linear association of stroke risk with TyG, ePWV, and TyG-ePWV. (5) E-values were computed to assess the potential influence of unmeasured confounders on the observed associations within this observational study.
All statistical analyses were conducted using R software (version 4.2.2). To account for multiple comparisons, the Bonferroni correction was applied.
