The design of the MESA study has been described in detail previously (20). In brief, a total of 6,814 participants aged 45–84 years and free of CVD at baseline visit (exam 1, 2000–2002) were initially recruited from six field centers. Follow-up examinations were conducted during 2002–2003 (exam 2), 2004–2005 (exam 3), 2005–2007 (exam 4), and 2010–2012 (exam 5). Per protocol, CAC was measured for all participants at study entry and repeated during the follow-up (20). In our analyses, we only included participants with at least three CT measurements, finally enrolling a total of 3,616 individuals. The studies involving human participants were reviewed and approved by the Ethics Committee of MESA. The patients/participants provided their written informed consent to participate in this study.

Methodology of CT scans and CAC measurements have been well-described in previous reports (21, 22). Briefly, either electron-beam or multi-detector CT (three sites) was used to scan these participants (11, 23). Two consecutive measurements were carried out and the results were read by a trained professor at a centralized reading center (Los Angeles Biomedical Research Institute, Torrance, CA, USA) (11). CAC scores were determined as phantom-adjusted Agatston scores by averaging the results from the two scans. Since the CAC scores were non-normally distributed data, they were converted to logarithmic values [log(CAC + 1)] before analysis (8, 17, 24).

Information on demographic characteristics, lifestyle factors, and past histories was obtained through standardized questionnaires or measured by health interviewers. Blood pressures were measured using an oscillometric method. Body mass index (BMI) was calculated by weight (kg) divided by square of height (m²). History of hypertension was defined as systolic blood pressure (BP) ≥140 mmHg and diastolic BP ≥90 mmHg, self-reported hypertension, or using anti-hypertensive medications. History of diabetes mellitus was defined as self-reported diabetes, a fasting serum glucose ≥126 mg/dl or use of anti-diabetic medications. Fasting serum glucose, creatinine, total cholesterol, high-density (HDL-C) or low-density lipoprotein cholesterol (LDL-C), and triglycerides were measured as previously described (20). Daily consumptions of different nutrients were quantified by data from a 120-item food frequency questionnaire during the previous year. Other available variables of interests used in the current study included age, sex, level of education, marital status, smoking status, alcoholic use, family income, physical activity, total energy consumption, dietary calcium, phosphate, and vitamin D.

Clinical events were ascertained by telephone interviews every 9–12 months, with all events adjudicated by an independent MESA committee (20). Participants were followed from baseline exam 1 until they experienced events of interest, were lost to follow-up, or until 31 December 2015. The main outcome in this study is hard CVD, including a composite of myocardial infarction (MI), resuscitated cardiac arrest, stroke (not transient ischemic attack), CHD death, and stroke death, which is defined and has been approved by the MESA Steering Committee. All CHDs (including MI, resuscitated cardiac arrest, definite angina, probable angina, and CHD death), chronic heart failure (CHF), and stroke were examined as secondary outcomes.

We used a GBTM to identify latent groups in participants’ CAC trajectories with the user-written program “traj” for STATA (25). Repeated CAC measurements were fitted as a mixture of several latent trajectories in a censor normal model with a polynomial function of age (e.g., linear, quadratic). The adequacy of the final model was evaluated using a low Bayesian information criterion (BIC) value and a probability of belonging higher than 0.70 (26, 27). Details of the selection process are presented in the Supplementary Materials.

Descriptive statistics are presented as means ± standard deviations (SD) for continuous variables and proportions for categorical variables according to the exposure trajectories.

Cox proportional hazards regression was used to examine the association of identified CAC trajectories with the studied outcomes, taking the low-stable group as reference. We used three models with increasing degrees of adjustment to assess the associated risks: (1) crude model (non-adjusted); (2) simple-adjusted model (adjusted for age, sex, race, alcoholic use, level of education, marital status, family income); and (3) full-adjusted model (adjusted for age, sex, race, alcoholic use, level of education, marital status, family income, body mass index, history of hypertension, diabetes, smoking status, physical activity, total energy consumption, dietary calcium, vitamin D and phosphate, LDL-C, HDL-C, total cholesterol, triglycerides, glucose, creatinine, systolic blood pressure, diastolic blood pressure, anti-hypertensive medication, anti-diabetic medication, and lipid-lowering medication). Results were reported as hazard ratios (HR) and 95% confidence intervals (95% CI). The proportional hazards assumption was checked by plotting the log[-log(survival)] versus log (survival time). We did not find evidence suggesting potential violation of these results.

Missing values on all sociodemographic covariates were handled by the Markov Chain Monte Carlo multiple imputation method and the results from 10 multiple imputation cycles were combined to draw a final output. The variables entered in the imputation method were potential confounders with missing values. The proportion of missing values was presented in Supplementary Table S1.

To mitigate potential bias, we repeated the main analysis after excluding participants with missing data on baseline covariates and performed a series of subgroup analyses. The association between different CAC trajectories and hard CVD was examined by subgroups of age (≤65 or >65 years), sex (male or female), race (Caucasian or non-Caucasian), hypertension (yes or no), and diabetes (yes or no).

In addition, univariate multinomial logistic regression models were used to identify potential risk factors for the identified trajectories. A change of the effect estimate >1% and a p-value <0.05 were further involved in the multivariate multinomial logistic regression models. To further evaluate the characteristics of traditional cardiovascular risk factors for the identified CAC trajectories, we estimated the dynamic change of systolic blood pressure, diastolic blood pressure, blood glucose, HDL-C, LDL-C, and BMI over time within each trajectory group. Second-order fractional polynomial models were fitted using the mfp and fracpred commands in STATA.

All data were analyzed using STATA version 15.1 (StataCorp/SE, College Station, TX, USA). All statistical tests were two-sided and the significance level was set at 0.05.

We enrolled a total of 3,616 participants in this study, among whom 78.29% (2,831/3,616) had three CT measurements and 21.71% (785/3,616) had four. The mean age of study participants at enrollment was 60.55 years (SD 9.54), 47.76% were men, and 39.30% were Caucasian. We compared the BIC values and average posterior probabilities (AVPP) in models with 1–6 classes with linear or quadratic functions. The final models were selected based on a low BIC and high probabilities of belonging. Finally, four distinct trajectories in CAC were identified (Figure 1): 24.17% (n = 874) of the participants were assigned in the low-stable class (class 1); 27.60% (n = 998) in the low-increasing class (class 2); 30.56% (n = 1,105) in the moderate-increasing class (class 3), and 17.67% (n = 639) in the elevated-increasing class (class 4). The mean AVPPs were 0.82 (SD 0.20), 0.78 (SD 0.22), 0.88 (SD 0.14), and 0.90 (SD 0.15) for those assigned in classes 1–4, respectively. The distribution of baseline and latest CAC scores across groups is depicted in the violin plot (Figure 2). Baseline characteristics, including sociodemographic information, health behaviors, and biomarkers according to different CAC trajectory groups, are summarized in Table 1.

In general, participants who maintained a low-stable CAC score were more likely to be female and non-Caucasian, less likely to drink or smoke, and had a lower proportion of hypertension and diabetes. They were less educated but had better metabolically favorable profiles (lower BMI, lower fasting glucose, lower blood pressure, lower LDL-C, and higher HDL-C).

During a mean follow-up of 13.58 years (SD 2.25), hard CVD occurred in 291 of the study participants. Overall, there was a graded positive association between CAC trajectories and cardiovascular risks. The event rates of hard CVD per 1,000 person-years by the exposed trajectories (from class 1 to class 4) were 2.23 (95% CI 1.53–3.25), 4.60 (95% CI 3.60–5.89), 7.67 (95% CI 6.38–9.21), and 10.37 (95% CI 8.41–12.80). Compared to participants assigned to class 1 (low-stable), the non-adjusted HRs of hard CVD for class 2 (low-increasing), class 3 (moderate-increasing), and class 4 (elevated-increasing) were 2.08 (95% CI 1.32–3.26), 3.48 (95% CI 2.29–5.29), and 4.74 (95% CI 3.07–7.29), respectively. The risk estimates persisted after simple adjustments with age, sex, race, use of alcohol, level of education, marital status, and family income. Further inclusion of the cardiovascular risk factors in the full-adjusted model did not notably alter the HR of hard CVD with CAC trajectories (HR 2.10, 95% CI 1.33–3.01 for class 2; HR 3.17, 95% CI 2.07–4.87 for class 3; HR 4.30, 95% CI 2.73–6.78 for class 4). Exclusion of participants with missing data on baseline covariates yielded similar results (Supplementary Table S2). Moreover, the graded associations between CAC trajectories and hard CVD were consistently observed in subgroups of age, sex, and race, with the presence or absence of hypertension or diabetes (Supplementary Table S2).

Similarly, a monotonically increased hazard risk for CHD was found for participants assigned to classes 2, 3, and 4. The full-adjusted HRs were 2.91 (95% CI 1.59–5.31), 5.38 (95% CI 3.06–9.46), and 11.80 (95% CI 6.68–20.87), respectively. With regard to CHF and stroke, the associations were weaker. The increased hazard for CHF in the full-adjusted model was only significantly observed in the elevated-increasing class (class 4), while for stroke, the highest hazard and event rate were noted in the moderate-increasing category. Details of the event rates and hazards adjusted for different covariates are summarized in Table 2. The cumulative hazards for the studied outcomes across different trajectories are plotted in Figure 3.

To identify the potential risk factors for the observed trajectories, we first examined the available sociodemographic factors in the univariate multinomial logistic regression model (Supplementary Table S3) and further tested significant ones in the multivariate model. As described in Table 3, younger male individuals, with a history of hypertension or diabetes, with lower HDL-C levels had a higher risk of being in the low-increasing category, while smoking, Caucasian race, and higher BMI and LDL-C level further predisposed participants to the moderate-increasing and elevated-increasing categories. We next characterized the dynamic change patterns of traditional cardiovascular risk factors within different CAC trajectories. Overall, patients in the moderate- and elevated-increasing trajectories had higher systolic and diastolic blood pressure, higher fasting glucose, lower HDL-C levels, and higher BMI. However, unexpectedly, participants in the elevated-increasing trajectory had lower LDL-C levels than those in the other trajectory groups (Figure 4).