In this paper, we present a method for quantifying coronary artery calcification as measured by CT that is independent of a threshold and instead calibrated to the phantom where the attenuation levels represent known densities. We adjust for noise in an automated way using spatial information in the image. The SWCS is shown to be highly related to both risk of CHD events and traditional CHD risk factors. In fact, it retains its strong relationship to traditional risk factors even when the AS equals zero.
Historically, interest in CAC has largely focused on its ability as a predictor of clinical CHD events, and the AS is an effective method of quantifying CAC for these purposes. However, our approach is motivated by studies where the primary interest in CAC is its ability to measure atherosclerotic burden across all levels of disease. In these cases, the AS is less suitable. Specifically, a continuous score for CAC would be more tractable in situations where the extent of atherosclerosis itself is of primary interest.
The primary goal of our analysis was to validate the effectiveness of the SWCS at quantifying CAC. Unfortunately, the “true” burden of CAC is unknown in the MESA patients. Instead, we examined the association of the SWCS with CHD events and known risk factors (via a linear predictor).
In validating the new score, we had three objectives: 1. Ensure that the new score does not lose any information about subclinical disease and risk of CHD events that the AS provides. 2. Determine that a positive score given to participants with AS = 0 is not merely noise, but rather measures lower levels of true subclinical disease. 3. Determine that the new score is at least as reproducible as the AS. We next address how we assessed these three points.
For Point 1, we compared the relationship of the AS and events to that of the SWCS and events. We observed statistically significant and strong relationships between the SWCS and CHD events and risk factors. For the subset of participants with AS > 0, the relationships between the LP and risk of events and CAC scores was similar for both CAC scores, suggesting that the SWCS is replicating the information in the AS in those participants with AS > 0. Furthermore, a simple visual examination of the plot (Figure1) of each non-zero AS against its corresponding SWCS suggests very strong correlation between the two scores. For all participants, and participants with AS > 0, the R2 values for the SWCS and AS are comparable. We emphasize that since the goal of our analysis is validation of the SWCS, the focus should be on R2 as a descriptive statistic comparing the SWCS and AS rather than as a measure of predictive accuracy. Nonetheless, the relatively low R2 values make sense, as the linear regression model only includes CAC, site, weight, and height. In particular, when we consider that CAC has been demonstrated to provide additional predictive information in addition to the traditional cardiovascular risk factors [8, 10, 15, 16, 28, 29] (which were used to construct the linear predictor), it is unsurprising that the R2 values are not high.
Examining these relationships in participants with AS = 0 addresses Point 2. For all CHD events, a doubling of SWCS was associated with a hazard ratio of 1.11 (95% CI, 0.79 to 1.55). For CHD risk, we observed a significant association between the SWCS and the LP; each unit increase in the log(SWCS + 1) was associated with an increase in the mean LP of 0.18 (95% CI, 0.14 to 0.22). The Kaplan-Meier curves for the first two quartiles of the SWCS show a distinct separation. Furthermore, the 50–62.5 and 62.5-75.0 percentile curves for the SWCS appear to be better separated than those of the AS. This suggests that the SWCS algorithm was in fact detecting additional useful information, and that the continuous scale represents a meaningful ordering of CAC suitable for measuring atherosclerotic burden and the attendant risk of CHD events.
To address Point 3, we assessed the reproducibility of the SWCS in comparison with the AS. Using the percent difference as a measure of reproducibility, we found that the median percent difference was 16.87% and 18.29% for the SWCS and AS, respectively. The intraclass correlations were 0.988 and 0.989 for the SWCS and AS, respectively. Our results suggest that the SWCS is at least as reproducible as the Agatston score.
There are two main limitations for the SWCS. One limitation was the lack of a true gold standard for validating the SWCS. The MESA study is composed of a large population of asymptomatic individuals and thus cardiac catheterization (with either angiography or intravascular ultrasound) was not a practical validation method, as it is restricted to studies of symptomatic populations or in treatment trials [18, 30–32]. Of course, histologic validation [1, 2] was not an option either. We instead used prediction of events and risk of events as a surrogate for a true gold standard, which in itself was limited by the small number of events in those participants with AS = 0. Furthermore, the emergence of models combining CAC measures with traditional risk summaries such as the Framingham risk [28, 29] suggests potential for using alternative, potentially richer models for validation. Regardless of the validation used, it would be useful to perform further validation using an independent but comparable dataset. In particular, separate in vitro experimental studies to validate SWCS would be highly desirable. Another limitation involves the applicability of the results outside of the MESA. The SWCS was designed to be a cost-effective approach to rescoring the MESA CT images in a less conservative manner than the original approach to reduce the number of false negatives (when the CAC score is zero but subclinical disease is in fact present). To use the approach presented here in other large studies where comparable calibration phantoms were not scanned or tracing the coronary arteries was not done is not possible.