Scientific Papers

Nonlinear connection between remnant cholesterol and stroke risk: evidence from the China health and retirement longitudinal study | Lipids in Health and Disease

Data source and study population

The CHARLS is a population-based, continuous longitudinal cohort study conducted nationally to assess social, health status, and economic factors [13]. The CHARLS provided the data for the current research. With 10257 homes and 17708 subjects participating in the baseline survey, the CHARLS sample was obtained from 450 communities in 150 districts, 28 provinces, and multi-stage probability sampling [13]. The sample was taken from the general retired population in China between June 2011 and March 2012 [13]. A longitudinal, face-to-face, computer-assisted personal interview was employed biennially to interact with the participants in the CHARLS [13]. Following the baseline survey from 2011 to 2012, survivors underwent three additional follow-up assessments from 2013 to 2014 (Wave 2), 2015 to 2016 (Wave 3), and 2017 to 2018 (Wave 4) [13]. The CHARLS website ( has detailed information about the data [13]. Follow-up of fewer than two years, age < 45 years at baseline, lack of information on RC at baseline, history of stroke, lack of information on stroke at baseline, or treatment for stroke at baseline were exclusion criteria for the 17,708 patients enrolled in the study at the beginning. In the end, a total of 10067 subjects were included in the follow-up investigation (Fig. 1).

Fig. 1
figure 1

Data collection

The CHARLS personnel at Peking University taught interviewers to use the computer-assisted personal interview method to conduct interviews in respondents’ homes. Demographics, health conditions and functioning, chronic diseases diagnosed by a doctor, lifestyle, and health-related habits (drinking, physical activity, and smoking) were all included in the core CHARLS questionnaire [13]. In addition, to measure the health performance and functioning of the interviewees, the interviewers had the equipment to measure their blood pressure, height, and weight. Participants were invited to the township hospital or the nearby Centers for Disease Control and Prevention (CDC) office, where a skilled nurse took an 8 ml fasting blood sample. Complete blood count (CBC) testing was done between one and two hours after the sample was taken. While the CBC was done, the remaining blood was divided into plasma and blood cells and transported in a -20°C environment. For CDC analysis, all blood samples were shipped back to Beijing [13]. Diabetes mellitus was characterized by having fasting plasma glucose (FPG) levels ≥ 7.0 mmol/L, hemoglobin A1c (HbA1c) levels ≥ 6.5% [14], or a self-reported history of diabetes. Heart diseases were defined as myocardial infarction, coronary heart disease, congestive heart failure, or other heart diseases [13]. Hypertension was defined as blood pressure greater than 140/90 mmHg (average of 3 measurements) or a history of hypertension [13]. Physical activity was categorized as vigorous or moderate activity for more than 30 min at least once a week [13]. Participants who had previously smoked were defined as ever smokers, and those who still smoked were defined as current smokers [13]. Participants who reported that they had consumed alcoholic beverages in the past were defined as ever drinkers, and those who had consumed any alcoholic beverages within the last year were defined as current drinkers [13]. Participants with less than two years of follow-up (i.e., participants who did not receive any follow-up) were considered as lost to follow-up. Participants who were not followed up to Wave 4 (2017–2018) and did not reach the endpoint but were followed up for more than two years were defined as censorship.

Covariates assessments

Based on a combination of prior research and clinical expertise, covariates were chosen [13, 15, 16]. The subsequent factors were used as covariates: (i) categorical variables: diabetes mellitus, sex, hypertension, smoking status, drinking status, physical activity, and lipid therapy; (ii) continuous variables: age, serum creatinine (Scr), C-reactive protein (CRP), and serum cystatin C.

Assessment of remnant cholesterol level

The following is a description of the precise measurement process for RC: RC = total cholesterol (TC) (mmol/L)—low-density lipoprotein cholesterol (LDL-C)—high-density lipoprotein cholesterol (HDL-C) [17, 18].

Diagnosis of stroke

The outcome variable was incident stroke throughout the follow-up period. As noted before, the following standard query was used to collect information on incident stroke: “Have you been informed by a doctor that you have a stroke diagnosis?” or “Because of your stroke, are you presently receiving any of the following treatments: taking Western medicine, physical therapy, taking Chinese traditional medicine, acupuncture therapy, and occupational therapy” [13]. If the individual provided a positive response at follow-up, the respondent was classified as having a first stroke diagnosis, and the self-reported time was recorded as the onset of the stroke. The time of the event was determined by subtracting the time of the baseline survey from the time of stroke onset.

Missing data processing

It is difficult to completely avoid missing data in observational research, which is a common issue [19]. There were 8 (0.08%), 1502 (14.92%), 91 (0.90%), 54 (0.54%), 11 (0.11%), 5996 (59.56%), 169 (1.68%), 12 (0.12%), 13 (0.13%), 80 (0.79%), and 2464 (24.48%) individuals with missing data for gender, BMI, diabetes mellitus, hypertension, drinking status, physical activity, smoking status, FPG, Scr, HbA1c, and Cystatin C, respectively. Missing clinical variables were addressed via chained equations using multiple imputations for modeling purposes. Age, sex, BMI, diabetes mellitus, hypertension, drinking status, physical activity, smoking status, lipid therapy, FPG, Scr, HbA1c, CRP, triglyceride (TG), and Cystatin C were all incorporated in the imputation model, which employed linear regression and ten iterations. The assumption of missing at random was used in missing data analysis processes [19].

Statistical analysis

All participants were divided into four groups based on RC quartiles (Q1 ≤ 0.310 mmol/L; 0.310 mmol/L < Q2 ≤ 0.520 mmol/L; 0.520 mmol/L < Q3 ≤ 0.84 mmol/L; Q4 > 0.84 mmol/L). The continuous baseline data were expressed as the mean ± the standard deviation (SD) (normally distributed data) and medians (quartile) (nonnormally distributed data). The current study used the expression in numbers (percentages) for categorical data. Comparisons were made using either ANOVA (nonnormally distributed data), the χ2 test (categorical data), or the Kruskal–Wallis test (skewed data). Incidence rates were expressed in person-years and cumulative incidence.

Using Cox proportional hazards models, the current study calculated the hazard ratios (HR) and 95% confidence intervals (CI) for stroke events. In addition, three Cox proportional hazards models were performed. The current study referenced prior research and clinical expertise in selecting covariates. The inclusion of these covariates in the model led to a substantial change of 10% or more in the HR. In Model 1, no covariates were adjusted. In Model 2, BMI, gender, age, drinking status, physical activity, heart diseases, diabetes mellitus, hypertension, and smoking status were controlled. BMI, gender, age, drinking status, physical activity, heart diseases, diabetes mellitus, hypertension, smoking status, lipid therapy, CRP, Scr, and cystatin C were used as adjustment variables of Model 3. Controlling variables were adjusted based on clinical knowledge and published reports [13, 15, 16]. Based on the results of the collinearity screening, no covariates were excluded from the Cox proportional hazards regression models according to the screening results for collinearity (Table S1).

The current study performed numerous sensitivity analyses to evaluate the reliability of the results. The current study converted RC into a categorical variable depending on the quartiles of the RC. By determining P for the trend, the results of RC as a continuous variable were examined, and the possibility of nonlinearity between RC and incident stroke was investigated. Hypertension and obesity are strong risk factors for stroke [20, 21]. Given their potential to confound the association between RC and incident stroke, we adopted an exclusionary approach in our analysis. Specifically, individuals with BMI ≥ 24 kg/m2 and hypertension were systematically excluded from our study cohort. This methodological decision was undertaken with the primary aim of elucidating and corroborating the relationship between RC and stroke while ensuring the stability and robustness of our findings. Furthermore, to minimize potential sources of bias and enhance the internal validity of our investigation, we extended the exclusion criteria to encompass non-hypertensive individuals with BMI < 24 kg/m2. These exclusions were conducted to foster a more homogenous study population and minimize the influence of confounding variables, thereby allowing for a more precise examination of the relationship between RC and stroke.

Moreover, the current study tested the nonlinear relationship between RC and stroke risk using the Cox proportional hazards regression model with cubic spline functions and the smooth curve fitting. If nonlinearity was observed, the current study’s initial step involved utilizing a recursive algorithm to calculate the inflection point. The algorithm was initiated arbitrarily, and filtering and smoothing techniques were employed to identify the inflection point. Subsequently, a two-piece Cox proportional hazards regression model was constructed on either side of the inflection point. The log-likelihood ratio was employed to determine the most appropriate model for elucidating the association between RC and the risk of stroke.

All the results were reported under the STROBE statement [22]. All analyses were conducted using R statistical software tools and Empower Stats. The threshold for statistical significance (two-sided) was established at P values less than 0.05.

Source link