An alternative method for quantifying coronary artery calcification: the multi-ethnic study of atherosclerosis (MESA)

The MESA was established to study the prevalence, progression, and risks of subclinical cardiovascular disease. Details of the study design have been previously published [22]. A cohort of 6814 men and women, aged 45–84, without known clinical cardiovascular disease was recruited between July 2000 and September 2002 from six urban communities (Baltimore City and Baltimore County, Maryland; Chicago, Illinois; Forsyth County, North Carolina; Los Angeles County, California; New York, New York; and St. Paul, Minnesota). Each of the six centers recruited a population-based sample while oversampling blacks, Chinese, and Hispanics based on self-reported ethnicity at time of enrollment. Institutional review boards at each site approved the study, and all participants gave written informed consent. Further documentation regarding data collection and protocols can be found at http://​www.​mesa-nhlbi.​org[23].

Information about risk factors and demographics was obtained at enrollment and during the baseline examination. Participants were given questionnaires to obtain information about tobacco usage, passive smoke exposure, alcohol use, medical conditions, medical care access, family history of CVD, reproductive history, and medication use and history. Physical activity was measured using a questionnaire from the Cross-Cultural Activity Participation Study; diet was measured using a modified version of the Insulin Resistance Atheroslecrosis study instrument; anger and anxiety were measured using questionnaires administered on the Spielberger trait anger and anxiety scales; depression was measured on the Center for Epidemiologic Studies Depression scale. Height, weight, and waist and hip circumferences were measured. Blood pressure was measured three times using an automated oscillometric sphygmomanometer (Dinamap Pro 1000; Critikon, Tampa, Florida). Blood samples were taken after a 12-hour fast, and were analyzed by a central laboratory at the University of Vermont for many different measurements including total cholesterol, lipids and lipoproteins (including HDL and LDL), insulin resistance, plasma glucose, triglycerides, high-sensitivity C-reactive protein, and creatinine measurements. Diabetes was defined as use of insulin or oral hypoglycemic agents, or fasting glucose of 126 mg/dL or greater. Body-mass index was defined as weight in kilograms divided by the square of height in meters. Documentation further detailing the questionnaire and measurement protocols can be found at http://​www.​mesa-nhlbi.​org[23] and Bild et al. [22].

We followed all participants for cardiovascular events for an average of 6.0 years. Interviewers called each participant at intervals of 9–12 months to obtain information about new CVD conditions, CVD interventions, hospitalizations, treatments, changes in life habits, and death. Detailed descriptions of MESA events definitions and follow-up procedure have been previously published [11, 15]. For our analyses, we separately considered any CHD (definite or probable myocardial infarction, definite CHD death, or definite or probable angina) and hard CHD (definite or probable myocardial infarction or definite CHD death) events. We note that probable angina cases were only classified as a CHD event if accompanied with a revascularization procedure.

A detailed report of the CT scanning and reading protocol has been previously reported [24]. At the baseline examination, each participant was given two consecutive unenhanced chest CT scans. The Chicago, Los Angeles, and New York sites used electron-beam CT systems (Imatron C150; GE Medical Systems, Milwaukee, Wisconsin), which scan at a slice thickness of 3.0 mm. The Baltimore, Forsyth County, and St. Paul sites used multi-detector CT systems (Lightspeed QXi, Lightspeed Plus; GE Medical Systems, Milwaukee, Wisconsin; Volume Zoom; Siemens, Erlangen, Germany), which scan at a slice thickness of 2.5 mm. To allow for calibration to compensate for inter-scan variability, radiographic phantoms with known densities of calcium hydroxyapatite (0, 50, 100, and 200 mg/mL) were placed beneath the thorax of each participant for each scan. Participants over 100 kg and at sites using multi-detector CT systems were scanned using slightly different settings.

Each scan was read by one of two physician readers at the CT reading center (Los Angeles Biomedical Research Institute at Harbor-UCLA, Torrance, California). As each participant received two consecutive scans, the scans were randomized to lessen the possibility of analysts consecutively reading scans from the same participant. Reading was done using an interactive system to identify the phantoms, determine the artery trajectories and classify candidate calcified plaques. For all techniques presented in this paper, the search for calcified lesions was restricted to be within an 8 mm radius of the trajectories defined by the readers. Further details can be found in previously published reports [24].

After arterial trajectories were determined and a phantom-based adjustment was applied, candidate calcified plaques were identified by the software with the criteria that each plaque be composed of at least 4 contiguous voxels with an attenuation level of 130HU or greater. The analysts then reviewed each candidate plaque and accepted or rejected its classification as calcified plaque.

To calculate the AS, each accepted lesion was assigned a score by multiplying the lesion volume by a coefficient based on its maximum HU (coefficient of 1 if maximum = 130–199, 2 if 200–299, 3 if 300–399, 4 if > = 400). The AS is the sum of the scores across all accepted lesions. As each participant received two scans, for purposes of this study the average of the two Agatston scores are used for each participant. Use of the average of two scans is intended to reduce noise and to be consistent with methods from other MESA papers using the AS [11, 15]. Further details on the Agatston score can be found in previously published reports [17, 24]. Participants were notified of their scores on the scale of no coronary calcification, less than average, average, and greater than average.

In developing the SWCS, an important property was for the SWCS to quantify CAC in a manner comparable to the AS for those with AS > 0, while providing meaningful non-zero scores for those with AS = 0. Calculation of the SWCS started with the set of voxels identified by the reader as representing the coronary arteries. First, we assigned a weight to each voxel using a weighting function with parameters derived from the scan’s phantom. The goal of the first step was to calibrate and weight each voxel according to the phantom so that scores across the images were comparable. Second, each voxel was then assigned a score depending on the weight assigned to it and its neighbors. The goal of the second step was to use the surrounding information of each voxel to obtain a more accurate value, and to reduce the impact of noise by upweighting those voxels with neighboring voxels that had high attenuation levels and downweighting those whose neighbors had low attenuation levels. The detailed algorithm appears in the Additional file 1: Appendix with illustrations of the weighting scheme.

We used Cox proportional hazards regression models similar to those used by Detrano et al. [11] to estimate hazard ratios for hard CHD and any CHD events as they relate to the SWCS and AS in all participants and the subset with AS > 0. The models were all adjusted for race, age, sex, smoking, diabetes, cholesterol, blood pressure, lipid-lowering medication use, and antihypertensive medication use. The AS and SWCS were transformed by taking the base-2 logarithm after adding 1 to the score (log₂[score + 1]).

Kaplan-Meier survival curves for all CHD events were estimated and plotted for both the SWCS and the AS. The AS K-M curves were stratified into the following percentiles: 0-50%, 50–62.5%, 62.5-75%, 75–87.5%, and 87.5-100%. The SWCS K-M curves were stratified similarly, but with additional 0-25% and 25-50% stratifications to further evaluate any additional ordering the SWCS may provide for participants with levels of CAC that are undetectable by the AS.

To summarize the association between the CAC scores and traditionally recognized CHD risk factors, we first summarized the risk factors in a linear predictor (LP) based on MESA data. The LP for each participant was the sum of their covariate values multiplied by the fitted coefficients from a Cox proportional hazards model. The Cox proportional hazards model was fitted using all CHD events as the outcome and baseline age, systolic blood pressure, diastolic blood pressure, total cholesterol, HDL, diabetes, smoking, and sex as covariates. The covariates were chosen based on the Framingham risk variables [25, 26]. We used linear regression models to estimate the relationship of the LP with the SWCS and AS in all participants and the subset with AS > 0. The models were all adjusted for weight, height, and site to accommodate different scanners across sites and different settings used for participants over 100 kg scanned using multi-detector CT scanners. The SWCS and AS were transformed by taking the natural logarithm after adding 1 to the score (log[score + 1]).

We used methods similar to those used by Callister et al. [19] to assess the reproducibility of the scores. We calculated the absolute difference between the scores of each scan relative to their mean: x 1 − x 2 / x 1 + x 2 / 2 × 100 where x ₁ is the score from the first scan and x ₂ is the score from the second scan. This was used as the primary measure of reproducibility and referred to as the percent difference for both the SWCS and AS. We also used the intraclass correlation coefficients to assess the reproducibility of the SWCS and the AS. Analyses were all done in R version 2.11.1 [27].

We compared demographics and risk factors between those with AS = 0 and AS > 0 (Table 1). The two groups were significantly different with respect to age, sex, hypertension, and treated diabetes, but not for status as a current smoker. We also bifurcated those with AS = 0 into those above and below the median based on their SWCS, comparing demographics and risk factors between these two subgroups. The two subgroups were significantly different with respect to sex, hypertension, treated diabetes, and Framingham 10-year risk, but not for age and current smoking status. As expected, the SWCS of participants with AS = 0 tend to be lower than those with AS > 0, with medians of 0.78 (range, 0.00-284.48) versus 58.99 (range, 0.22-4262.70), respectively. We note that while some participants had very small SWCS, all participants had positive SWCS. A visual representation of the relationship of the AS and SWCS among participants with AS > 0 is shown in the Figure1. As suggested by the scatterplot, there is a high correlation of 0.99 and the linear relationship can be summarized by the equation y = 0.52 + 0.67 x.

CT image data was available in 6568 of the 6814 participants. Of these, 6253 (91.8%) had image data available for both CT scans and 315 (4.6%) for one scan. The distribution of the SWCS was quite skewed; thus we present the smoothed density of the SWCS plus a constant of 1, on the natural log scale in Figure2. For the 6568 participants who received at least one scan, we show the smoothed density for all participants, participants with AS > 0, and participants with AS = 0.

There were 6508 participants with available measures of all the variables used for the events analysis. The average follow-up time was 6.0 years, during which 291 participants experienced a CHD event of which 171 were hard CHD events. Among those with AS = 0, 34 participants experienced a CHD event, of which 22 were hard CHD events.

An increase in the SWCS was associated with an increase in risk of CHD events (Table 2). A statistically significant relationship was observed when the model was applied to all participants, and when restricted to participants with AS > 0. Furthermore, the relationships were similar to those seen when the SWCS was replaced by the AS in the same models.

A twofold increase in the SWCS or the AS was associated with a 23% (95% CI, 16 to 30) or 19% (95% CI, 14 to 26) increase in risk of hard CHD events, respectively. Furthermore, among participants with AS > 0, a twofold increase in the SWCS or AS was associated with an 18% (95% CI, 9 to 27) or 16% (95% CI, 8 to 25) increase in risk of hard CHD events, respectively.

When we restricted the model to only participants with AS = 0, a twofold increase in the SWCS was associated with a hazard ratio of 1.11 (95% CI, 0.79 to 1.51) for hard CHD events.

The cumulative event rate and number of all CHD events among participants grouped by different percentile stratifications of the scores is shown in Figure3. An ideal risk score would have Kaplan-Meier curves that are well separated and ordered across the quantile-based strata. For the 75–87.5 and 87.5-100 percentile stratifications, the SWCS and AS appear to do equally well at differentiating the quantiles in terms of event risk. The 50–62.5 and 62.5-75 percentile stratifications of the SWCS separate appropriately over time, while those of the AS appear in the incorrect order throughout most of follow-up. That is, that the higher quantile group has lower risk. Finally, the 0–25 and 25–50 percentile stratifications of the SWCS appear to demonstrate some separation, suggesting the SWCS provides a degree of meaningful ordering of CAC in terms of risk in the lower quantiles as well.

Among the 6510 participants with complete data for the risk factors, an increase in the SWCS was associated with an increase in the LP (Table 3). For all participants, on the log(CAC + 1) scale, each unit increase in the SWCS and AS was associated with an increase in the mean of the LP of 0.23 (95% CI, 0.22 to 0.24) and 0.18 (95% CI, 0.17 to 0.19), respectively. For participants with AS > 0, on the log(CAC + 1) scale, each unit increase in the SWCS and AS was associated with an increase in the mean of the LP of 0.18 (95% CI, 0.16 to 0.20) and 0.16 (95% CI, 0.15 to 0.18), respectively. For participants with AS = 0, each unit increase in the log of the SWCS plus one was associated with an increase in the mean of the LP of 0.18 (95% CI, 0.14 to 0.22).

For all participants, and participants with AS > 0, the R² values for the SWCS and AS are comparable. For participants with AS = 0, the R² value is 0.12. A scatterplot of the LP against the log(SWCS + 1) for all participants is shown in Figure4. The dot colors represent participants with AS = 0 or AS > 0. Smoothed curves through the points for all participants, and the subsets of participants with AS = 0 and participants with AS > 0 are also shown in different colors. The vertical lines at the bottom and top of the plot are visual aids indicating the distributions of log(SWCS + 1) for participants with AS = 0 or AS > 0.

After excluding scans judged by the scan readers to have unacceptable levels of noise, there were 3102 participants who had at least one scan with AS > 0. The median percent difference across participants was 16.87% and 18.29% for the SWCS and AS, respectively (Table 4). The median of the percent difference in AS minus the percent difference in SWCS was 2.02% (95% CI, 1.66 to 2.40; P < 0.001).

When we performed the same analysis on the subset of 2849 participants with AS > 0 for both scans, the SWCS still showed better reproducibility compared to the AS. We also performed the same analysis further restricted to participants with AS < 50 (n = 892) for both scans and AS < 100 (n = 1307) for both scans. Among these subsets of participants with low but non-zero AS, SWCS still showed better reproducibility.

For the participants for whom we had CT image data for both scans (n = 6253), the intraclass correlation coefficients for the SWCS and AS were 0.988 and 0.989, respectively. For those with AS > 0 for both scans (n = 2849), the intraclass correlation coefficients for the SWCS and AS were both 0.986. For those with AS = 0 for both scans (n = 3151), the intraclass correlation coefficient for the SWCS was 0.406.