Abstract
Background
In hospitalized patients undergoing coronary computed tomography angiography (CCTA), variability in stenosis interpretation may contribute to missed obstructive coronary artery disease (CAD) or unnecessary invasive coronary angiography.
Project Summary
We retrospectively evaluated inpatient CCTA interpretation workflows at a tertiary academic center between 2023 and 2025. Among 177 hospitalized patients undergoing CCTA, 139 had both artificial intelligence (AI)–assisted and physician interpretations available for comparison. Twenty patients underwent same-admission invasive coronary angiography, with 19 included in final invasive angiographic analysis. AI-assisted interpretation demonstrated higher agreement with invasive coronary angiography than physician interpretation within the exploratory invasive angiography subgroup (78.6% vs 55.0%). Discordance patterns were primarily driven by false-negative physician interpretations and AI-assisted stenosis overestimation. Significant CAD was identified in 73.7% of the invasive angiography subgroup, with 52.6% ultimately undergoing revascularization.
Take-Home Message
Hybrid AI-physician interpretation workflows may support more consistent inpatient CAD evaluation while preserving physician-guided clinical decision making.
Key words: cardiovascular disease, coronary angiography, imaging
Visual Summary
Coronary computed tomography angiography (CCTA) has become an increasingly important noninvasive modality for evaluating suspected coronary artery disease (CAD), particularly among patients presenting with chest pain syndromes and intermediate-risk clinical profiles.1 In hospitalized patients, interpretation of CCTA findings directly influences downstream diagnostic pathways, including invasive coronary angiography (ICA), revascularization decisions, and discharge planning.
Despite advances in image acquisition and reporting standardization, clinically meaningful variation in CCTA interpretation persists. Differences in stenosis grading and plaque characterization may contribute to missed high-risk disease or unnecessary invasive testing. These concerns are particularly relevant in inpatient populations, as elevated clinical acuity and competing comorbidities frequently complicate diagnostic decision making.
Artificial intelligence (AI)–assisted CCTA analysis has emerged as a potential strategy to support consistency in cardiovascular imaging workflows.2 Prior studies have demonstrated promising applications for AI-assisted plaque characterization and stenosis quantification, although the role of hybrid AI-physician interpretation pathways in real-world inpatient practice remains incompletely understood. Beyond diagnostic performance alone, uncertainty remains regarding how AI-assisted CCTA interpretation may influence real-world inpatient workflow standardization, invasive angiography utilization, and downstream clinical decision making. These considerations are increasingly relevant as health care systems emphasize value-based care, resource stewardship, and reduction of unwarranted diagnostic variation across cardiovascular imaging pathways.
Case Summary Prompting Project Launch
Increasing utilization of inpatient CCTA at our institution revealed notable variability in plaque detection and stenosis characterization between AI-assisted analysis and physician interpretation. In several patients undergoing ICA, discordant interpretations appeared to influence downstream management decisions, including referral for invasive testing and revascularization strategy selection. These observations prompted a quality improvement initiative focused on evaluating diagnostic variation and identifying opportunities to improve consistency in inpatient CAD evaluation workflows.
Project Rationale
Reducing unwarranted diagnostic variation represents an important quality improvement target in cardiovascular care. Missed obstructive CAD may delay revascularization and expose patients to adverse cardiovascular outcomes, whereas overestimation of disease severity may contribute to unnecessary invasive angiography, prolonged hospitalization, increased procedural risk, and excess health care utilization.
Because inpatient CCTA increasingly serves as a gatekeeper for invasive testing, improving consistency in interpretation has important implications for patient safety, resource stewardship, and value-based care delivery. Hybrid AI-physician interpretation pathways may provide complementary strengths by combining AI-assisted plaque detection with physician-guided clinical judgment.
Project Description
We conducted a retrospective evaluation of inpatient CCTA interpretation workflows in hospitalized adults undergoing clinically indicated CCTA at a tertiary academic medical center between 2023 and 2025. Routine physician interpretation was performed by board-certified cardiologists with dedicated cardiovascular computed tomography imaging expertise according to institutional workflow. AI-assisted analysis was performed using commercially available coronary plaque and stenosis analysis software integrated into the imaging workflow. Coronary stenosis severity was categorized using standardized Society of Cardiovascular Computed Tomography–guided stenosis assessment categories.3
A clinically selected high-risk subgroup undergoing ICA during the same hospitalization was used to evaluate diagnostic agreement and downstream revascularization outcomes. Significant CAD on ICA was defined as ≥70% stenosis in a major epicardial coronary artery or ≥50% left main stenosis.
Primary quality improvement measures included diagnostic concordance with ICA and characterization of discordance patterns between AI-assisted and physician interpretation. Secondary measures included downstream revascularization outcomes, including coronary artery bypass grafting and percutaneous coronary intervention.
Categorical variables were summarized descriptively using counts and percentages. Given the modest size of the invasive angiography subgroup, findings were considered exploratory and hypothesis generating (Figure 1).
Figure 1.
Study Design and Patient Flow
Flow diagram demonstrating cohort selection and study design for hospitalized patients undergoing clinically indicated coronary computed tomography angiography (CTTA) between 2023 and 2025. Among 177 hospitalized patients undergoing CCTA, 140 had analyzable artificial intelligence (AI)–assisted imaging data available, and 139 had both AI-assisted and physician interpretations available for comparison. Twenty patients underwent same-admission invasive coronary angiography, with 19 included in the final invasive angiography validation subgroup after exclusion of incomplete angiographic data.
Project Deliverables
Among 177 hospitalized patients undergoing CCTA, 140 had analyzable AI-assisted imaging data available, and 139 had both AI-assisted and physician interpretations available for direct comparison. The mean age of the cohort was 59.5 years, and 53.1% were women.
Twenty patients underwent same-admission ICA, with 19 included in final invasive angiographic analysis. AI-assisted interpretation demonstrated numerically higher agreement with ICA compared with physician interpretation (78.6% vs 55.0%). Exploratory subgroup analysis also demonstrated numerically higher sensitivity and overall diagnostic accuracy for AI-assisted interpretation.
Discordance patterns were primarily driven by false-negative physician interpretations and AI-assisted stenosis overestimation relative to invasive angiography. The left circumflex artery represented the most common discordant vascular territory.
Within the ICA subgroup, significant CAD was identified in 73.7% of patients, and 52.6% ultimately underwent revascularization. Coronary artery bypass grafting was performed in 36.8% of patients, and percutaneous coronary intervention was performed in 15.8% (Figure 2).
Figure 2.
Diagnostic Concordance and Discordance Patterns
(A) Agreement with invasive coronary angiography among the invasive coronary angiography subgroup (n = 19), demonstrating numerically higher agreement for artificial intelligence (AI)–assisted interpretation compared with physician interpretation (78.6% vs 55.0%). (B) Summary of major discordance patterns observed between AI-assisted and physician interpretation. (C) Exploratory diagnostic performance metrics within the invasive angiography validation subgroup.
Project Outcome, Impact, and Future Directions/Next Steps
This quality improvement initiative identified clinically meaningful variation in inpatient CCTA interpretation with important downstream implications for invasive testing utilization and revascularization decisions. The findings suggest that hybrid AI-physician workflows may support improved diagnostic consistency in selected inpatient populations.
Importantly, these findings should not be interpreted as evidence that AI replaces physician expertise. Rather, the observed variation highlights potentially complementary strengths between AI-assisted and physician interpretation. AI-assisted analysis may support lesion detection consistency and standardized stenosis characterization, whereas physician interpretation may prioritize specificity and individualized clinical decision making in medically complex hospitalized patients.
Findings from this initiative were reviewed with institutional cardiologists to identify opportunities for workflow standardization and selective incorporation of AI-assisted plaque analysis into inpatient CCTA interpretation pathways. In particular, observed discordance patterns informed ongoing discussions regarding hybrid review strategies for intermediate or potentially high-risk lesions being considered for ICA.
Future directions include prospective evaluation of hybrid AI-physician workflows, assessment of interreader variability reduction strategies, and incorporation of standardized decision support pathways aimed at improving consistency in inpatient CAD evaluation, while minimizing unnecessary invasive testing (Figure 3).
Figure 3.
Downstream Revascularization Outcomes
(A) Prevalence of significant coronary artery disease among patients undergoing same-admission invasive coronary angiography (ICA). (B) Overall revascularization rate within the invasive angiography subgroup. (C) Distribution of revascularization strategy, including coronary artery bypass grafting (CABG) and percutaneous coronary intervention (PCI) outcomes.
Discussion
As health care systems increasingly emphasize value-based care and resource stewardship, reducing unwarranted diagnostic variation has become a major priority across cardiovascular imaging pathways.4 Our findings provide real-world insight into inpatient CCTA interpretation practices and demonstrate how discordance between AI-assisted and physician interpretation may influence downstream care pathways.
The numerically higher agreement observed between AI-assisted interpretation and invasive angiography suggests a potential role for AI as an adjunctive decision support tool within inpatient imaging workflows. At the same time, physician interpretation demonstrated a more conservative stenosis characterization pattern, reflecting the clinical complexity frequently encountered in hospitalized patients.
Several factors likely contributed to observed discordance patterns. Left circumflex lesions represented the most common area of disagreement, which may reflect anatomical variability, image quality limitations, and vessel-specific interpretive challenges. Additionally, the invasive angiography subgroup represented a clinically selected high-risk population rather than a screening cohort, limiting broader generalizability and affecting specificity assessment. Additionally, because ICA was clinically selected rather than systematically performed across the entire cohort, verification bias may have influenced exploratory diagnostic performance estimates.
Despite these limitations, this project highlights an actionable opportunity for workflow optimization. Hybrid AI-physician interpretation pathways may represent a pragmatic strategy to improve consistency while preserving physician oversight and individualized patient-centered decision making.
Importantly, the potential value of hybrid AI-physician interpretation extends beyond diagnostic accuracy alone. In contemporary inpatient practice, CCTA increasingly functions as a gatekeeper for invasive angiography and downstream resource utilization. Accordingly, strategies that improve interpretive consistency may have broader implications for patient safety, procedural appropriateness, health care utilization, and value-based cardiovascular care delivery.
Conclusions
Clinically meaningful variation exists between AI-assisted and physician CCTA interpretation in hospitalized patients. Hybrid AI-physician interpretation workflows may support more consistent inpatient CAD evaluation pathways, while balancing detection of high-risk disease against avoidance of unnecessary invasive testing.
Visual Summary.
Hybrid AI-Physician CCTA Interpretation Workflows Were Evaluated in Hospitalized Patients Undergoing Inpatient CCTA Assessment
Artificial intelligence (AI)–assisted interpretation demonstrated higher agreement with invasive coronary angiography compared with physician interpretation within the exploratory invasive angiography subgroup. Discordance patterns were primarily driven by false-negative physician interpretations and AI-assisted stenosis overestimation. Hybrid review pathways may support more consistent inpatient coronary artery disease evaluation and downstream invasive testing decisions. CCTA = coronary computed tomography angiography.
Funding Support and Author Disclosures
The authors have reported that they have no relationships relevant to the contents of this paper to disclose.
Take-Home Messages
-
•
Clinically meaningful variation exists in inpatient CTTA interpretation and may influence downstream invasive testing and revascularization decisions.
-
•
Hybrid AI-physician interpretation workflows may support more consistent inpatient CAD evaluation, while preserving physician-guided clinical decision making.
Acknowledgments
The authors thank the Division of Cardiovascular Medicine and cardiovascular imaging teams involved in patient care and imaging workflow review.
Footnotes
The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center.
References
- 1.Budoff M.J., Dowe D., Jollis J.G., et al. Diagnostic performance of 64-multidetector row coronary computed tomographic angiography for evaluation of coronary artery stenosis in individuals without known coronary artery disease: results from the ACCURACY trial. J Am Coll Cardiol. 2008;52(21):1724–1732. doi: 10.1016/j.jacc.2008.07.031. [DOI] [PubMed] [Google Scholar]
- 2.Dey D., Slomka P.J., Leeson P., et al. Artificial intelligence in cardiovascular imaging: JACC state-of-the-art Review. J Am Coll Cardiol. 2019;73(11):1317–1335. doi: 10.1016/j.jacc.2018.12.054. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Cury R.C., Leipsic J., Abbara S., et al. CAD-RADS™ 2.0 – 2022 coronary artery disease reporting and data system. J Cardiovasc Comput Tomogr. 2022;16(6):536–557. doi: 10.1016/j.jcct.2022.07.002. [DOI] [PubMed] [Google Scholar]
- 4.Hecht H.S., Narula J., Fearon W.F. Fractional flow reserve and coronary computed tomographic angiography-guided management of coronary artery disease. J Am Coll Cardiol. 2016;67(13):1565–1572. [Google Scholar]