Abstract 4366401: Artificial Intelligence-Based Coronary Artery Calcium (AI-CAC) Score Empowers the “Power of Zero”: An AI-CVD Study within the Multi-Ethnic Study of Atherosclerosis (MESA)

Background: The “Power of Zero” refers to the excellent negative predictive power of a coronary artery calcium (CAC) score of zero. Because CAC is currently reported using the Agatston score and this scoring method is based on a fixed radiologic density threshold (>130 HU) and minimum size (contiguous voxels >=1 mm 2 in area), small or semi-calcified plaques are sometimes missed. We hypothesized AI-based coronary artery calcium (AI-CAC) scoring that does not require these fixed thresholds provides a superior predictive performance in individuals with an Agatston score of zero. Methods: CAC scans and follow-up data (median [IQR]: 14.2 [13.6 – 14.8] years) from 3260 participants with an Agatston score of zero in the Multi-Ethnic Study of Atherosclerosis (MESA) baseline examination were analyzed (ages 57.9 ± 9.2 years, 62.8% female, 32.8% White, 30.9% Black, 24.4% Hispanic, 11.9% Chinese). AI-CAC scores were measured using AI-CVD automated coronary artery segmentation (HeartLung Technologies, Houston, TX). Voxels within the coronary segmentations were calibrated and standardized based on MESA’s phantom used during scanning. Voxels were additionally weighted based on the intensity of their in-plane neighbors. We examined the prognostic separation of AI-CAC=zero score versus Agatston=zero score with the cumulative incidence for myocardial infarction (MI), hard coronary heart disease (CHD) events, and all CHD events. Significant differences in proportions were calculated using the log-rank test or McNemar’s test for overlapping groups. Results: AI-CAC provided a continuous range of values in individuals with Agatston=zero, revealing a previously unrecognized spectrum of semi-calcified plaque burden. 1,789 (54.6%) participants had an AI-CAC score of zero, and 1.485 (45.4%) had a non-zero score. In incident CHD cases, the mean ± SD for AI-CAC was 5.9 ± 11.6 and median (IQR) was 2.27 (0-5.5). There were significantly fewer MI, hard CHD and all-CHD events in AI-CAC=zero vs. Agatston=zero over 5, 10 and 15 years of follow up. Within the Agatston=zero population the AI-CAC>0 showed significantly higher event rates over AI-CAC=0. (See table below) Conclusion: AI-CAC scores significantly improve on the negative predictive power of the Agatston CAC score of zero in MESA possibly by identifying small, semi-calcified or “soft” plaques. Future studies are warranted to corroborate these findings in other longitudinal cohorts.