Atherosclerosis Imaging Quantitative Computed Tomography (AI‐QCT) to guide referral to invasive coronary angiography in the randomized controlled CONSERVE trial

Abstract Aims We compared diagnostic performance, costs, and association with major adverse cardiovascular events (MACE) of clinical coronary computed tomography angiography (CCTA) interpretation versus semiautomated approach that use artificial intelligence and machine learning for atherosclerosis imaging‐quantitative computed tomography (AI‐QCT) for patients being referred for nonemergent invasive coronary angiography (ICA). Methods CCTA data from individuals enrolled into the randomized controlled Computed Tomographic Angiography for Selective Cardiac Catheterization trial for an American College of Cardiology (ACC)/American Heart Association (AHA) guideline indication for ICA were analyzed. Site interpretation of CCTAs were compared to those analyzed by a cloud‐based software (Cleerly, Inc.) that performs AI‐QCT for stenosis determination, coronary vascular measurements and quantification and characterization of atherosclerotic plaque. CCTA interpretation and AI‐QCT guided findings were related to MACE at 1‐year follow‐up. Results Seven hundred forty‐seven stable patients (60 ± 12.2 years, 49% women) were included. Using AI‐QCT, 9% of patients had no CAD compared with 34% for clinical CCTA interpretation. Application of AI‐QCT to identify obstructive coronary stenosis at the ≥50% and ≥70% threshold would have reduced ICA by 87% and 95%, respectively. Clinical outcomes for patients without AI‐QCT‐identified obstructive stenosis was excellent; for 78% of patients with maximum stenosis < 50%, no cardiovascular death or acute myocardial infarction occurred. When applying an AI‐QCT referral management approach to avoid ICA in patients with <50% or <70% stenosis, overall costs were reduced by 26% and 34%, respectively. Conclusions In stable patients referred for ACC/AHA guideline‐indicated nonemergent ICA, application of artificial intelligence and machine learning for AI‐QCT can significantly reduce ICA rates and costs with no change in 1‐year MACE.


| INTRODUCTION
Invasive coronary angiography (ICA) allows for evaluation of stable symptomatic patients with suspected coronary artery disease (CAD) to guide decisions of coronary revascularization. 1,2 While current American College of Cardiology (ACC)/American Heart Association (AHA) guidelines outline appropriate selection of patients for elective ICA, in real-world practice, most individuals who undergo non-emergent ICA do not have actionable CAD. 3,4 For these patients, ICA has been shown add to unnecessary health care system costs and increase the risk for potential procedural complications. 5,6 In the 2021 Updated ACC/AHA Chest Pain guideline, coronary computed tomography angiography (CCTA) has been elevated to a class IA indication to serve as a first line test for identification and exclusion for obstructive CAD with a high sensitivity of 95%−99%. [7][8][9] Evaluation of CCTA in stable symptomatic patients referred for nonemergent ICA has been done previously in the Coronary Computed Tomographic Angiography for Selective Cardiac Catheterization (CONSERVE) randomized controlled trial (RCT), which observed a selective referral strategy that incorporates a CCTA-first approach before catheterization was associated with a 77% reduction in ICA. 10 This deferral of ICA was associated with reduced rates of coronary revascularization and downstream costs, with no differences in 12-month rates of major adverse cardiovascular events (MACE) as compared to a direct ICA referral strategy.
In this analysis of the CONSERVE RCT, we hypothesized that application of atherosclerosis imaging and quantitative cardiac computed tomography (AI-QCT) would allow for better determination of patients with and without obstructive CAD who may benefit from ICA, and that this approach would be associated with reduced ICA and lower costs without added risk of MACE. AI-QCT was performed using a commercially available software platform (Cleerly Labs, Cleerly, Inc.) that performs atherosclerosis imaging quantitative CCTA (AI-QCT) analysis 15-17 using a series of validated convolutional neural network models for quantitative image quality assessment, coronary segmentation and labeling, vascular morphology measurements, and atherosclerotic plaque characterization. 15 Hundred percent of studies were analyzable by AI-QCT and included in the study results. A case example with invasive angiography correlation is shown in Figure 1. Prior validation of AI-QCT has been reported in 2 multicenter trials. 15,17 Study analysis was performed in-kind for this investigator-initiated study.

| METHODS
Coronary segments with a diameter ≥ 2 mm were included in the analysis using a modified 18-segment SCCT model. 10,16 Each segment was evaluated for the presence or absence of coronary atherosclerosis, defined as any tissue structure > 1 mm 3 within the coronary artery wall that was differentiated from the surrounding epicardial tissue, epicardial fat or the vessel lumen itself. The following CAD features were evaluated: • Stenosis: Utilizing a normal proximal reference vessel crosssectional slice, the start and the end of the lesion were identified, and from the cross-sectional slice that demonstrated the greatest absolute narrowing, % diameter stenosis severity was automatically calculated. Obstructive stenosis was defined at ≥50% and ≥70% stenosis thresholds. All vessels with 0% stenosis were defined as having no CAD.
• Atherosclerosis: Atherosclerosis characterization was performed by AI-QCT for every coronary artery and its branches. Plaque volumes (PVs) (mm 3 ) were calculated for each coronary lesion and then summated to compute the total PV at the patient level.
Plaque with a minimum volume of ≥3mm 3 was included for analysis. This provided data for analysis on both the per-lesion and per-patient level. PV was further categorized using Hounsfield unit (HU) ranges with noncalcified plaque (NCP) defined as HU between −30 and +350; low density-NCP (LD-NCP) defined as plaques < 30 HU; and calcified plaque (CP) defined as >350 HU. 17 All statistical analyses were performed using SAS version 9.4 (SAS).
Continuous data are reported as mean ± standard deviation, and categorical variables are presented as absolute numbers with corresponding percentages. The rates of stenosis, 0%, 1%−24%, 25%−49%, ≥50% and ≥70% were compared individually between AI-QCT and Level II/III site readers on a per patient and per vessel basis. The per-patient differences were evaluated using McNemar's test of the paired data. The per-vessel rates were compared using the logistic Generalized Estimating

| Clinical characteristics of the study population
Demographic and clinical characteristics of the study cohort (60 ± 12 years, 49% women) are listed in Table 1. There was a high prevalence of CAD risk factors, including: 57% hypertension, 33% hyperlipidemia and 30% smokers. 88% of patients experienced symptoms suggestive of CAD, with the majority (70%) having typical or atypical angina.

| Comparison of an AI-QCT approach to clinical CCTA interpretation
Application of AI-QCT identified 87% and 95% patients without stenosis ≥ 50% and ≥70%, respectively, who would be eligible for ICA deferral ( Table 2). For intermediate stenoses 50-69%, AI-QCT Example of a patient with AI-QCT analysis demonstrating severe obstructive (>70%) luminal stenosis in the proximal LAD with invasive cath correlation. AI-QCT total plaque volume, calcified and noncalcified plaque is also shown. AI-QCT, atherosclerosis imaging and quantitative cardiac computed tomography; LAD, left anterior descending.

T A B L E 1 Baseline demographics and clinical characteristics.
Variable (% or mean ± SD)

| Cost-analysis
Results of an AI-QCT-based strategy for referral management of only patients with high-grade stenosis to ICA are listed in Table 4. At a ≥ 50% and ≥70% stenosis threshold, application of AI-QCT would have resulted in 87% and 95% patients, respectively, avoiding unnecessary ICA at a 26% and 34% cost-savings, respectively.

| DISCUSSION
In this present study, we evaluated for the first time an AI-QCT strategy to guide judicious referral to nonemergent ICA for patients with an ACC/ AHA Guideline indication and determined that adoption of an AI-QCT approach could reduce unnecessary ICA by 87%−95% based upon stenosis severity thresholds. The rates of safe ICA deferral from AI-QCT were significantly higher than those based upon Level II/III reader interpretation of CCTA. Further, the AI-QCT approach was safe, with no patient experiencing MACE during the length of the follow-up period who had been quantified as having non-severe stenosis by AI-QCT.
Finally, an AI-QCT approach was cost-efficient compared to standard of care Level II/III CCTA interpretation, with a 26-34% reduction in costs by AI-QCT-based ICA deferral.
To our knowledge, these present study results represent the first to evaluate within a multicenter RCT the clinic-economic feasibility of   20 This so-called "diagnostic-therapeutic cascade," if broken, may reduce unnecessary PCI for patients who will not benefit from its performance. 21 In the original CONSERVE trial, PCI rates were reduced by~50% and, based upon the current study findings, could be further reduced by application of an AI-QCT strategy.

| LIMITATIONS
The present study is not without limitations. The current analyses were performed in post hoc fashion from an international, multicenter, RCT.

| CONCLUSIONS
Application of AI to typically acquired CCTA is a clinically effective, safe and cost savings approach to guide referral management of patients being considered for ICA.