Opportunistic Detection of Coronary Artery Calcium on Noncardiac Chest Computed Tomography: An Emerging Tool for Cardiovascular Disease Prevention: A Scientific Statement From the American Heart Association.
Oct 16, 2025DOI: 10.1161/cir.0000000000001382Journal ArticleAHA
Written by: Randi ForakerLaurence SperlingLisa BratzkeMatthew BudoffMichelle LeppertAlexander C RazaviFatima RodriguezMichael D ShapiroSeamus WheltonNathan D WongEugene YangAmerican Heart Association Prevention Science Committee of the Council on Epidemiology and Prevention and Council on Cardiovascular and Stroke Nursing; Council on Clinical Cardiology; and Council on Cardiovascular Radiology and Intervention
Find these authors on Roon
AR
Dr. Alexander Razavi
Preventive Cardiology · Emory University Hospital
MS
Dr. Michael Shapiro
Preventive Cardiology · Wake Forest School of Medicine
Rationale
Coronary artery calcium (CAC) is a direct measure of calcified subclinical atherosclerotic plaque, strongly associated with atherosclerotic cardiovascular disease (ASCVD) risk.1 When quantified on electrocardiogram (ECG)-gated non-contrast cardiac computed tomography (CT), CAC has consistently demonstrated improvement in ASCVD risk stratification beyond traditional risk factor-based approaches2,3 and provides more personalized information to guide clinician-patient discussions regarding the initiation and intensity of preventive therapies.4-7 As such, consideration of measurement of CAC is recommended by national and international guidelines for individuals with a borderline or intermediate 10-year ASCVD risk as a tool for risk reclassification.8-11 Despite these guidelines, there has been a low adoption of CAC scans among the 17.6 million individuals in the United States (US) who may be eligible, due to lack of awareness among clinicians. Low clinician awareness is multifaceted and may include a lack of knowledge regarding how CAC scores may change management of ASCVD risk, the risk-benefit of radiation, potential for incidental findings, and out of pocket costs resulting from lack of uniform insurance coverage.12,13The low adoption of formal CAC scoring is a part of a broader underutilization of ASCVD risk assessment tools, including risk calculators (e.g. estimation of 10- and 30-year risk). For example, survey data suggests that less than one-half of physicians in the US perform routine ASCVD risk estimation despite the integral role of risk calculators in guidelines for identifying individuals who may benefit from statin and antihypertensive therapies.14 Such low uptake of guideline-endorsed risk assessment tools may have direct implications for preventive cardiology care and has identified a need to pursue additional strategies for the identification of high-risk individuals prior to an ASCVD event.Opportunistic detection of CAC (ODCAC), which is easily detected on non-ECG-gated chest CT scans performed for medical reasons other than to measure CAC (Figure 1), has the potential to bridge this gap and alert physicians of patient risk, independent of guideline-recommended risk calculator use. Approximately 19 million noncardiac chest CT scans are performed annually in the US, considerably more than the 1 million dedicated CAC scans per year.16 Therefore, quantification and reporting of ODCAC represents a significant opportunity to improve ASCVD risk stratification and appropriate allocation of preventive medication among a much larger and more diverse population of patients. Furthermore, using data that has already been collected alleviates potential barriers to CAC scoring such as additional radiation exposure, out-of-pocket cost, and concerns regarding detection of incidental findings (Figure 2). The potential utility of opportunistic detection of subclinical atherosclerosis by non-ECG gated chest CT extends beyond CAC to identify and report evidence of calcification in the aortic and mitral valve, as well as the thoracic aorta.Several societies have provided guidance on the reporting of CAC on non-cardiac chest CT,17,18 which can be measured using the Agatston method or categorized through qualitative visual assessment (mild, moderate, severe), both of which are associated with ASCVD risk.18,19 Recently, quantification of CAC has become more automated and there are several Food and Drug Administration (FDA)-approved artificial intelligence (AI) algorithms for automated CAC scoring.20 AI algorithms are currently in use at several large healthcare systems across the US, as they expedite the quantification of ODCAC for CT readers with the goal of increasing the likelihood of ODCAC reporting. Importantly, reporting of ODCAC using an AI-based approach has been shown to lead to an increase in appropriate allocation of preventive therapies compared to standard care.21The rapid expansion of AI technology to enable quantification of ODCAC has brought forward the need for implementation of systems-based approaches to ensure appropriate interpretation, reporting, and clinical action related to findings. Notably, there are foreseen challenges managing an anticipated significant increase in the number of individuals identified with subclinical atherosclerosis. To reduce the burden of a mismatch between disease detection and healthcare delivery and to maximize preventive care, systems-based efforts will be needed to: 1) develop systematic reporting mechanisms and identify those with ODCAC not yet receiving appropriate preventive therapies; 2) improve precise utilization of preventive therapies according to CAC burden; 3) identify appropriate treatment algorithms and clinical care teams (e.g. radiology, preventive cardiology, trained primary care clinicians) for longitudinal follow-up care; and 4) avoid unnecessary downstream testing.
Technical Approaches to CAC Assessment
AI evaluation of CT scans may have important advantages over traditional reader-based techniques. Given that AI evaluation is not affected by human factors such as visual fatigue, cognitive workload, distractions, work environment, experience level, and stress, it can reduce reader workload, improve intra and inter-observer variability, and be much faster, allowing for routine assessment of ODCAC on all non-cardiac chest CT. However, there are important considerations in the measurement of CAC from non-cardiac chest CT scans that include 1) potential differences in CAC measurement that can occur from non-gated as compared with traditional gated CT scans, 2) differences in AI algorithms for measurement of CAC and the extent to which they have been externally validated, as well as 3) differences in how CAC has been reported from these strategies (e.g., CAC values versus CAC categories).While formal CAC scans performed on multidetector CT scanners are ECG gated to minimize motion artifact,22 non-cardiac chest CT scans are more likely to suffer from motion artifact. There are times when opportunistic CAC assessment may not be feasible, due to severe coronary motion or artifacts present on the scan, and radiologists need to indicate when image quality does not allow CAC assessment. Quantification of CAC on contrast studies is more problematic, as the contrast densities often overlap with densities of calcific plaque. In this context, it is prudent to report visual estimates of CAC as ordinal assessment of CAC still predicts risk.23,24 Studies have previously shown very good agreement between gated and non-gated scans in assessment of CAC, and most misclassification occurs for CAC scores close to reference categories. In a sub study of the Multiethnic Study of Atherosclerosis where 2,750 participants received gated and non-cardiac chest CT scans on the same day and using the ECG gated scans as the gold standard where subjects were classified into CAC categories of 0, 1-100, 101-400, and >400, the non-cardiac chest CT scans correctly classified 86% of participants into the same CAC tiers.25 In this study, only 7% of the non-cardiac chest CT scans overestimated CAC, and significant CAC (>400) was interpreted as mild or intermediate in only 1% of cases with a weighted kappa and 95% confidence interval of 0.86 (0.84, 0.89). In an earlier systematic review and meta-analysis of five studies comprising 1,316 individuals for validation, there was a pooled kappa for agreement of 0.89 (0.83, 0.95) for increasing CAC categories, and non-cardiac chest CT showed a false-negative rate of 9% and underestimated high CAC scores in 19% of individuals.26 In a more recent meta-analysis of 4,000 patients from 16 studies, there was a pooled correlation coefficient of 0.94 (0.89, 0.97) with a weighted kappa for CAC severity of 0.85 (0.79, 0.91).27
Artificial Intelligence Algorithms for Detection of CAC
An early study examining AI ODCAC measurement using a machine-learning k-nearest neighbor approach involved analysis of 1,749 non-contrast low-dose non-cardiac chest CT scans and showed a reliability of k=0.85 compared to a manually assessed Agatston category with a mean difference in Agatston score of 2.5.28 More recently, deep learning (DL) techniques have been used, including convolutional neural network (CNN) models where 97% of calcifications were detected with 84% accuracy in risk category assessment compared to ECG gated CAC.29 Other CNN-derived algorithms have shown over 90% sensitivity and specificity with >80% of cases classified to the correct CAC risk category.30An important consideration is ensuring that these DL models can be generalized across scanner types and hospital systems. An evaluation of chest CT scans from multiple scanner vendors and hospitals showed a strong correlation between AI-assisted and manual CAC (r=0.89, p<0.001) with 80% concordance in risk category.31 Studies demonstrate high concordance between DL techniques for CAC quantification on non-cardiac chest CT and the gold-standard Agatston scoring system, with intraclass correlation coefficient between 0.94 to 0.99 and kappa consistently above 0.7.30,32,33
Reporting Methods for CAC
Both the Society of Cardiothoracic Computed Tomography (SCCT) and the Society of Thoracic Radiology provide a class I recommendation for physicians to provide at least a qualitative interpretation of CAC from non-cardiac chest CT. Several methods for reporting CAC from non-cardiac chest CT have been proposed including 1) simple visual interpretation, 2) total visual score and length-based methods, 3) the Weston score, 4) segment-specific extent-based scores, and 5) vessel-specific extent-based scores.34 In particular, the SCCT recommends the use of the Coronary Artery Calcium Data and Reporting System (CAC-DRS) for standardized reporting, which includes visual burden of CAC (mild, moderate, severe) and the number of vessels with CAC.35 Visual scores of mild, moderate, and severe correspond to Agatston CAC scores of 1-99, 100-299, and ≥300, respectively.A simple visual score does not involve measurement of the Agatston score and at its most basic level is simply noted as the presence or absence of CAC and not even quantitated on an ordinal scale. Being subjective, interobserver agreement for CAC severity is between 0.80 and 0.85, although studies have shown a simple evaluation to be equally as prognostic as presence based on the Agatston score.34 Since formal CAC scans are performed with standardized 120 kV and 2.5 mm slice thickness, varying kV settings and slice thicknesses used on non-cardiac CT scans can also create discrepancies with lower kV scans more prone to false positive and thicker slices may underestimate CAC. Total visual score and length-based methods are more accurate as they measure the length of the calcification resulting in improved inter- and intra-observer agreement with a very high negative and positive predictive value, although this method is more time consuming.The Weston Score34 quantifies the extent of “blooming” of CAC, correlates well with the Agatston score (r=0.81), and has been shown to be predictive of CVD events. However, this score is subjective and challenging for physicians not trained to do such assessments.34 Segment-specific extent-based scores involve assessment of the number of coronary segments involved with CAC in the entire coronary tree, but this method is more time consuming than simpler methods. The vessel-specific extent-based score has been greater utilized and reports on the number of vessels involved with CAC, and the severity of each vessel (mild, moderate, severe calcification).35 It is also the most time-efficient method with good interobserver agreement, as well as excellent sensitivity and specificity, and has been demonstrated to predict CVD events.34More recent AI techniques offer Agatston scoring of CAC with one DL model showing nearly identical Agatston score (535.3 vs. 542.0), mass (128.9 vs. 129.5), and volume measurements (454.2 vs. 460.6) obtained at the total and vessel-specific levels compared to manual methods.36 Such techniques reduce burden on the CT reader and increase consistency of reporting compared to formal ECG CAC scoring.15,37-39 A summary of aspects related to traditional scoring (“naked eye”) versus AI-based quantification methods is shown in Table 1.
Landscape of Current Data
Non-ECG-gated chest CTs are performed at a volume that far exceeds gated chest CT scans, with important implications for increasing the detection of CAC across diverse populations who may not seek preventive care and who may not be able to afford the out-of-pocket costs associated with dedicated scans. Patients undergoing non-gated scans for non-cardiac chest CT tend to be more often from racial and ethnic minority groups and have greater social need as compared to those undergoing ECG gated chest CT.40 Hence, the detection of CAC using non-cardiac chest CT represents an important opportunity for intervention and may be a potential bridge to address health disparities.The prevalence of incidental CAC is dependent on the population studied, many of whom are enriched in ASCVD risk factors. The prognostic value of incidental CAC for ASCVD risk predication is very similar to that of CAC identified on dedicated cardiac CT and is similar for both men and women.39,41 Such findings have been demonstrated across a range of populations, including those undergoing low-dose radiation CT scans for lung cancer screening, whose risk of death from ASCVD exceeds that of lung cancer.42-44 Individuals undergoing low-dose CT for lung cancer screening represent one population where ODCAC is most readily applicable, and therefore could serve as a population to assess best practices for implementation on standardized reading and reporting. Further quantification of ODCAC through DL algorithms enhances risk prediction. In one study at a large academic health system, among 5,678 adults free of ASCVD at baseline, ODCAC quantified through a DL algorithm (≥100 Agatston units) was associated with an increased risk of all-cause death and adverse cardiovascular outcomes beyond traditional risk factors.39 Furthermore, less than 20% of these high-risk individuals were already on statin therapy. In a separate cohort of more than 8,000 patients undergoing lung cancer screening CT, more than one in three had CAC >400 of whom 30% were not on lipid-lowering therapy.41 Overall, these results suggest significant opportunities to improve preventive care. Beyond prevention of atherosclerotic outcomes, AI-CAC cardiac chamber volumetry has been developed and adds to risk prediction beyond ASCVD, including total ASCVD events over the Agatston score,45 heart failure (beyond NT-pro-BNP, risk factors, and CAC),46 and atrial fibrillation [beyond Cohorts for Heart and Aging Research in Genomic Epidemiology for Atrial Fibrillation CHARGE-AF)],47 potentially providing opportunities for early intervention to prevent these conditions as well.Knowledge of CAC facilitates shared-decision making, is a powerful motivator for lifestyle modification,48 and improves precision in the allocation of statin17 and non-statin49 lipid-lowering therapies, aspirin,50 intensive blood pressure lowering,51 and semaglutide weight-loss therapy.52 The Early Identification of Subclinical Atherosclerosis by Noninvasive Imaging Research (EISNER) randomized trial showed that asymptomatic patients who received CAC scoring information had superior cardiovascular risk factor control compared to those who did not receive CAC scoring and no increase in downstream medical testing and procedures.53 Similarly, the Coronary Calcium (CorCal) study randomized 601 mostly non-Hispanic White participants to a CAC versus risk score-based strategy for initiating statin therapy for primary prevention.54 This study found that a CAC-guided approach resulted in a 1-year adherence to statin recommendation of 62% compared to 42% for the pooled cohort equation (PCE)-based strategy. Additionally, the CorCal study found CAC scoring to be a cost-effective adjunct to PCE for guiding primary prevention, as it reduced the need for laboratory tests and office visits for monitoring.55Among individuals without clinical ASCVD with a CAC score between 1 to 399 Agatston Units, the Coronary Artery Calcium Score: Use to Guide Management of Hereditary Coronary Artery Disease (CAUGHT-CAD) trial demonstrated that an intervention including knowledge of CAC score, lifestyle modification, and statin therapy led to lower plaque progression on CCTA (15.4 versus 24.9 mm3) and lower LDL-cholesterol (-51 mg/dL versus -2 mg/dL) compared to patients blinded to knowledge of their CAC score and standard care.56In a quality improvement study performed at Stanford Health Care, Incidental Coronary Calcification Quality Improvement Project (NOTIFY-1), statin-naïve patients were randomized to notification of ODCAC versus usual care. Identification of ODCAC was performed among 2,113 individuals who had undergone non-gated, non-contrast CT over 5 years, and 20% of individuals had ODCAC. After trial exclusions, 173 individuals were included. Notification patients received their personalized CAC images and guideline recommendations, as did their primary care clinicians. Among statin eligible individuals, the notification arm had a statin prescription rate of 51% versus 7% for the usual care group at 6 months, with benefit seen across racial and ethnic groups as well as among women.18 These data support the feasibility of incorporating notification of CAC into routine clinical workflows and the potential impact of notifying patients and their clinicians of the presence of ODCAC on non-cardiac chest CT scans.Additionally, the Risk Or Benefit IN Screening for CArdiovascular disease (ROBINSCA) trial randomized over 40,000 subjects to CAC scanning, Systematic COronary Risk Evaluation (SCORE) risk assessment, or usual care and will provide the most definitive evidence to date regarding the value of CAC scanning and resulting interventions from it on improving cardiovascular outcomes.57 The CAC scan intervention within the Danish Cardiovascular Screening trial (DANCAVAS) trial likely contributed to the higher initiation of statin and aspirin therapy that was responsible for improvements in median survival observed among the subgroup of men ages 65 to 69 years old.58
Implementation Considerations for CAC Scoring
The use of CAC can encourage conversations related to behavior change and emphasize the role of early intervention (i.e., medication) in disease prevention.59 A current National Lipid Association Scientific Statement recommends referral for incidental findings of non-cardiac chest CT.17 However, numerous barriers to implementation of ODCAC reporting exist including geography (i.e., medically underserved areas). Berzingi et al. reported on the utility of mobile scanning units for medically underserved areas where clinicians added CAC scoring to low-dose CT lung cancer screenings.60 Fifty-four percent of the 526 screened had identified CAC, however, less than 8% of those patients with identified risk had CAC scores reported and were referred for follow-up; even fewer had appropriate therapy initiated (i.e., statins, aspirin). While these findings demonstrated that mobile CT units may improve access challenges faced by individuals who do not live in the vicinity of a medical center, they also highlight the need to approach systematic barriers holistically when designing implementation of a system of referral for CAC scores.Despite certain barriers to ODCAC reporting, there are several promising solutions and areas for improvement (Table 2).61 Grant and colleagues trialed an intervention that involved both clinician education coupled with template and reporting changes within the electronic health record (EHR) that could serve as a roadmap for others interested in improving rates of referral and follow up. A convenience sample of patients undergoing chest CT for lung cancer screening and CT for reasons other than lung cancer screening was collected. Implementation of the interventions resulted in a significant increase in the reporting of mild (7% vs. 37%), moderate (0% vs. 80%) to severe (0.5% to 93%) CAC during the study.62EHR functionality and clinical workflows for ODCAC and associated referrals are health-system dependent. However, most EHRs can enable referrals for or notifications of CAC findings from radiology to primary care or preventive cardiology. While pragmatic ODCAC screening and referrals offer promise, it is important to avoid unnecessary downstream testing leading to higher healthcare costs. It will also be critical to collect data on treatment outcomes of patients identified in clinical workflows that are implemented to enhance ODCAC. A proposed clinical workflow for ODCAC assessment is shown in Figure 3.Upon identification of incidental CAC, clinicians should engage in a risk discussion with patients regarding the significance and prognostic implications of such findings.63 First, clinicians should present images of CAC to the patient as a visual aid to communicate risk as prior evidence suggests that patient visualization of atherosclerosis may enhance understanding and adherence to therapy.56 In addition to communicating the presence of calcified atherosclerosis, clinicians should discuss qualitative burden (mild, moderate, severe) with patients, which may also help with comparing an individual’s risk with someone of similar age and sex. Finally, after patients have understood the ASCVD risk associated with CAC, clinicians should engage in discussions involving evidence-based approaches for comprehensive risk reduction according to the most recent guidelines.64 Such approaches may include but are not limited to improving lifestyle habits, discussion of low-density lipoprotein cholesterol and blood pressure goals, diabetes prevention and treatment, obesity risk stratification, and antithrombotic therapy.49,50,52,65-67
Future Directions and Considerations for Automated Intelligence-based CAC Detection
The detection of CAC is increasingly recognized as a valuable factor in assessing ASCVD risk and informing treatment decisions. As advances in technology continue to evolve, the role of ODCAC quantification in risk stratification and treatment decision-making is expanding. By offering clarity on which patients might benefit most from intensive preventive measures, CAC scoring can help identify candidates for targeted pharmacologic interventions beyond lifestyle modifications. Recent developments in risk assessment frameworks, such as the AHA’s Predicting Risk of cardiovascular disease EVENTs (PREVENT) equations, may benefit from incorporating CAC (including ODCAC from non-cardiac chest CT) as an additional stratifying variable and tool for risk reclassification. CAC could be utilized within these models to add specificity, identifying those with a high burden of subclinical disease and aligning therapy to individual risk more precisely, particularly among those with equivocal traditional risk factors. Establishing formal recommendations on how CAC scores should be integrated with extant risk assessments (Table 3) and influence treatment initiation and titration will be important.The presence and extent of CAC have been shown to correlate with future cardiovascular events, thus providing clinicians with an actionable metric to guide patient management. This approach aligns with the broader objective of precision medicine in cardiovascular care, moving toward a more personalized model of prevention based on an individual’s specific risk profile.68 Further, staging of atherosclerosis based on extent of CAC (with specific clinical recommendations for each stage) has been proposed for the purpose of providing for more actionable interventions.69AI has the potential to automate CAC detection on routine and ODCAC scans, rapidly identifying calcified coronary lesions and calculating scores with minimal human intervention. Similar processes may be achieved for AI-based interpretation of coronary artery anatomy, plaque burden and composition, in addition to stenosis severity from non-coronary CT angiography studies. Beyond the coronary arteries, AI algorithms for ODCAC may be helpful for the detection of aortic valve calcium, mitral valve calcium, and thoracic aorta calcium (TAC), which may improve risk prediction of aortic stenosis, atrial fibrillation, and heart failure, respectively.70 For example, while CAC and TAC are closely correlated, they reflect different biological processes (CAC: intimal atherosclerosis, TAC: medial atherosclerosis) and emerging evidence suggests that the presence of severe TAC may be associated with an increased risk of heart failure.71 Further, AI-CAC cardiac chamber volumetry (measuring left and right atrial and ventricular volumes) has been developed and adds to the prediction of total ASCVD events over the Agatston score, along with further plaque characterization including the number of plaques, the number of vessels with plaques, plaque density, and location not reported by the Agatston score, which improves prediction of coronary heart disease in those with CAC 1-99 defined by the Agatston score.45The pragmatic detection of CAC through opportunistic use of chest CT, particularly when enhanced by AI technology, holds the potential to favorably impact ASCVD risk assessment and management. AI technologies may be able to be scaled effectively across healthcare systems, making it feasible to assess ASCVD risk in large populations efficiently and could increase access to CAC screening, making it feasible in community settings, underserved regions, and resource-limited environments where manual interpretation by highly specialized readers may not be available. For AI-based CAC detection to reach its full potential, ongoing research and validation will be necessary to establish accuracy, reliability, and integration within clinical workflows. Studies to assess the clinical utility of AI-derived CAC scores in various populations will further inform best practices, including thresholds for treatment initiation or modification. Additionally, collaboration between AI and clinical experts will be needed in developing user-friendly tools that integrate into clinical practice.
Conclusions
Looking to the future, the incorporation of CAC scoring into routine risk assessment and leveraging AI to broaden access to CAC data hold the promise of more precise, personalized cardiovascular disease prevention. These advances may enable clinicians to make better-informed decisions, tailoring interventions to individuals’ specific risk profiles and improving outcomes at both the individual and population levels, and AI tools may be deployed if there exists adequate return on investment as has been demonstrated with ODCAC. There is a driving need to enable ODCAC in current clinical workflows. If ODCAC becomes a cornerstone of practice, we can improve risk prediction, enhance cardiovascular disease prevention, and impact a much larger and diverse range of patients currently only reached by traditional CAC scanning.
MS
Dr. Michael Shapiro
Cardiology · Preventive CardiologyWake Forest University Health Sciencesmichaels
