Hagopian and colleagues developed a deep-learning model to quantify coronary artery calcium on computed tomography scans across Veterans Affairs hospitals. This AI-derived model successfully predicted cardiovascular outcomes and identified patients who could benefit from lipid-lowering therapy. Evidence Rating Level: 2 (Good)
Study Rundown
Coronary artery calcium (CAC) quantification using electrocardiogram (ECG)-gated computed tomography (CT) scans is well-established for determining cardiovascular risk. However, CAC is not routinely quantified on non-gated CT scans obtained for noncardiac reasons. Hagopian and colleagues developed AI-CAC, a deep-learning model that automatically segments and scores CAC on non-contrast, non-gated chest CTs. The model was developed with expert segmentations and was tested against clinical electrocardiogram (ECG)-gated CAC scores in patients with paired scans. Main outcomes included threshold classification accuracy and agreement with reference scores. In 795 paired scans, AI-CAC achieved 89.4% accuracy for distinguishing zero from non-zero CAC and 87.3% accuracy for differentiating scores below or above 100. AI-CAC also demonstrated good agreement (kappa=0.72) with ECG-gated CAC. Additionally, four cardiologists qualitatively reviewed a random sample of patients with AI-CAC scores greater than 400 and confirmed that 99.2% would benefit from lipid-lowering therapy. This study demonstrated the potential for using AI and CT scans to screen for CAC on a larger scale.
In-Depth [retrospective cohort]
AI-CAC was developed within the United States Veterans Affairs system using non-contrast, non-gated CT scans from 98 medical centers. Coronary artery calcium (CAC) on 548 scans was manually segmented by a cardiac CT specialist. 795 patients with non-gated CTs were paired with ECG-gated CAC scans for independent testing. Additionally, a separate cohort of 8,052 low-dose lung cancer screening CTs was used for opportunistic screening. Performance metrics included accuracy, F1 score, agreement across CAC categories, and time-to-event analyses for mortality and first myocardial infarction, stroke, or death. AI-CAC demonstrated 89.4% accuracy (F1=0.93) for zero versus nonzero CAC and 87.3% accuracy (F1=0.89) for scores below 100 versus at least 100. For score groups of 0, 1-100, 101-400, and greater than 400, AI-CAC and ECG-gated CAC scores showed good agreement (test-paired: kappa=0.72). AI-CAC was also predictive of 10-year all-cause mortality (CAC 0 vs. >400: 25.4% vs. 60.2%, Cox hazard ratio 3.49; P<0.005), and the composite of initial stroke, MI, or death (CAC 0 vs. >400: 33.5% vs. 63.8%, Cox hazard ratio 3.00; P<0.005) on the Test-Paired dataset. Among the opportunistic screening cohort, AI-CAC also accurately identified patients who would benefit from lipid-lowering therapy. This study was limited by the predominantly male veteran population and lack of prospective evidence that implementation of AI-CAC improves treatment or outcomes. Overall, this study demonstrated that an AI algorithm has the potential to expand CAC screening using chest CT scans ordered for non-cardiac reasons.
