Study population
Tehran Lipid and Glucose Study (TLGS) is a prospective population-based cohort study. In 1999–2002 the first phase of the study recruited 15,005 men and women aged three years and more, living in Tehran, district No.13, a representative sample of an urban Iranian population. The participants of TLGS have been followed up for 20 years, approximately every three years, and so far, the data have been collected across six subsequent phases. Details regarding the methods and design of TLGS have been reported previously [13,14,15]. In brief, trained social workers invited participants to the TLGS unit and took written informed consent from them. Demographic and lifestyle information was obtained using self-reported standard questionnaires. Then trained physicians interviewed participants to get past medical history, smoking habits, and physical exam. Systolic and diastolic blood pressure (mm Hg) was taken as the average of two measurements in the sitting position after five-minute rest using a standard mercury sphygmomanometer. Anthropometric measurements were taken according to the standard protocols with shoes removed and the participants wearing light clothing. A blood sample is drawn from all study participants after a 12–14 h overnight fast. A second blood sample is taken two hours after glucose ingestion according to the standard protocol. Biochemical measurements including fasting plasma glucose (FPG), two-hour post-glucose load (2-hPG), and all blood lipid analyses were performed at the TLGS research laboratory on the day of blood collection using Selectra 2auto-analyzer (Vital Scientific, Spankeren, Netherlands). Diabetes was defined by fasting glucose ≥ 126 mg/dL or the use of diabetes medication. The Ethics Committee of the Research Institute for Endocrine Sciences, Shahid Beheshti University of Medical Sciences, approved all protocols of TLGS.
For the current study, all participants of TLGS aged 40–79 years who attended the baseline assessment and at least one additional follow-up re-examination between the second (2002–2005) and fourth cycles (2009–2011) were included (n = 4268). After excluding participants with prevalent CVD (n = 331), those with missing data on the CVD risk score at baseline (n = 161), and those with missing data on the CVD risk score during all examination cycles one to four (n = 77), 3699 participants were included in the trajectory analysis. We also followed those attending the fourth examination cycle and did not have prevalent CVD and missing data in this cycle (n = 2619) up to March of 2018. This sub-sample was used in survival analysis to link the CVD risk trajectory groups (defined using the original sample) to incident hard CVD, including non-fatal myocardial infarction, fatal coronary heart disease, and fatal or non-fatal stroke.
Assessment of cardiovascular risk
The Pooled Risk Equations recommended by ACC/AHA for non-Hispanic white men and women were used to calculate the 10-year risk of hard CVD [5]. These equations included covariates of age, total cholesterol, HDL cholesterol, treated or untreated systolic blood pressure, history of diabetes (Y/N), and current smoking status (Y/N); the validity of these equations was previously evaluated in the TLGS [16]. The risk score components were drawn from questionnaires and clinical examination data at six examination cycles: 1999–2001, 2002–2005, 2006–2008, 2009–2011, 2012–2014, and 2015–2018.
Outcomes
The primary outcome for the prospective analysis was hard CVDs, including non-fatal myocardial infarction, fatal coronary heart disease, and fatal or non-fatal stroke [5, 16]. The TLGS participants are followed up for any medical event, including CVDs and death during the previous year, by telephone calls annually. An outcome committee consisting of an internist, endocrinologist, cardiologist, and epidemiologist adjudicates all events. Deaths are confirmed through death certificate records. The cause of death is determined based on the death certificate and detailed review of medical records and all information provided by attending physicians, medical examiners, and/or family members. For the prospective analysis, participants were followed from the fourth examination cycle (2009–2011) until March 2018.
Statistical analysis
Data are shown as mean ± standard deviation for continuous variables or as number (%) for categorical variables. To identify distinct baseline to 10-year CVD risk trajectories, we used group-based trajectory models in a Stata plugin program (Stata Proc Traj) [17]. This method models the dependent variable (ACC/AHA risk score) as a function of time. It identifies individuals' clusters following a similar underlying trajectory on the dependent variable over time within a population, based on a maximum likelihood method [17, 18]. We applied a censored normal model [19] to identify distinct trajectories of the CVD risk score. We fitted different models to determine the "best" model, treating CVD risk score as the dependent variable and time at follow-up as the independent variable. We developed different models by varying numbers of groups, ranging from two to five groups, and shapes (linear, quadratic, and cubic). We then compared them using Bayesian Information Criteria (BIC) and a sufficient proportion of participants in each subgroup [18]. The results of this process are summarized in Additional file 1: Table S1. To ensure the adequacy of the selected model, we assessed four models that fit diagnostic criteria as suggested by Nagin [20]: (1) average posterior probability of assignment for each group j (AvePPj) equal to 0.7 or greater for all groups; (2) the odds of correct classification (OCCj) equal to 5 or higher for all groups; (3) similarity between the proportion of a sample assigned to a specific group and the group probabilities estimated from the model; and (4) narrow CIs of the estimated proportion. Using this approach, we identified three distinct trajectories for CVD risk score. The selected CVD risk trajectory groups were labeled according to their CVD risk at baseline and examination cycle 4 to show the trajectory of the risks during the time. The median of the trajectory groups' risk was 3%, 17%, and 38%, which are compatible with the ACC/AHA risk categories of low < 7.5%, medium ≥ 7.5 to < 20, and high ≥ 20.
After identifying CVD risk score trajectory groups, we evaluated the associations of trajectory subgroup membership (as a categorical exposure) with incident hard CVD after the fourth examination cycle using Cox proportional hazards regression model. Given that total cholesterol, HDL cholesterol, systolic blood pressure, history of diabetes, and current smoking status are included in the ACC/AHA pooled cohort risk algorithm, we did not consider them in the multivariate model. Since age is a strong non-modifiable risk factor and increases during follow-up, it was adjusted in model 2—education level and family history of premature CVD as non-modifiable factors adjusted in model 3. Moreover, BMI, lipid-lowering drug use, and physical activity have not been included in the ACC/AHA pooled cohort model. However, we did not adjust them because their intermediate variables (i.e., blood pressure, cholesterol, and diabetes) are still in the original model. This adjustment results in underestimating HRs for risk trajectories. Statistical significance was considered using a two-sided P < 0.05. All analyses were performed using Stata software version 14 (STATA Corp., TX, US).
Sensitivity analysis
As a sensitivity analysis, to assess the impact of increasing age on the CVD risk estimates and the overall shape of the trajectories, we repeated trajectory analysis by calculating new risk scores using risk factor values at each examination cycle but the age at the first exam.
Besides, we assessed the trend of each risk factor included in the ACC/AHA risk score containing systolic blood pressure (SBP), total cholesterol, HDL, fasting blood sugar (FBS), smoking, as well as CVD risk scores by the trajectory groups identified in the primary analysis using a generalized estimating equation (GEE) analysis.