Prognostication and clinical opportunities with AI for coronary artery calcium: a scoping review

Abstract

Objective

Coronary artery calcium (CAC) scoring is an established marker of atherosclerotic burden and cardiovascular risk, yet its clinical use remains limited by access, cost and workflow constraints. Artificial intelligence-enabled CAC (AI-CAC) tools allow automated assessment of coronary calcification from chest CT performed for diverse clinical indications, creating opportunities for scalable, opportunistic cardiovascular disease risk stratification.

We conducted a scoping review of AI-CAC models focused on prognostication and real-world deployment, emphasising outcomes, clinical value and implementation considerations.

Methods and analysis

Peer-reviewed original studies using an AI model incorporating CAC from chest CT for clinical prognostication, risk prediction or assessment of clinical interventions, reporting clinical outcomes, prognostic metrics or implementation insights. Studies focused exclusively on technical AI-CAC development or validation without clinical context were excluded.

PubMed/MEDLINE, Embase and Cochrane Library, searched through November 2025, supplemented by expert recommendations from study coauthors.

A single reviewer screened all records. Data were extracted using a standardised 25-variable template and synthesised narratively using an AI translational maturity framework.

Results

Across diverse patient populations and imaging contexts, AI-CAC demonstrated consistent associations with all-cause mortality, cardiovascular mortality, major adverse cardiovascular events and obstructive coronary artery disease. Radiomics-based approaches incorporating high-dimensional imaging features may improve prognostic performance beyond Agatston scoring alone, demonstrating further methodological potential for assessing cardiovascular risk. Emerging prospective and implementation-focused studies suggest that integrating AI-CAC into clinical workflows increases preventive care engagement. However, most evidence remains retrospective with limited evaluation of cost-effectiveness, equity or long-term patient and health system impact.

Conclusion

AI-CAC represents a promising approach for scaling atherosclerotic cardiovascular disease risk stratification. Beyond expanding access to traditional CAC scoring, AI-enabled approaches may support the evolution of the CAC score itself by integrating plaque characteristics, distribution and complementary imaging biomarkers to support more personalised preventive strategies rather than relying on fixed risk thresholds alone. Realising meaningful clinical impact will require prospective implementation studies, standardised reporting and evaluation, integration into care pathways and ongoing monitoring aligned with ethical, operational and financial considerations.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Coronary artery calcium (CAC) scoring is a validated marker of atherosclerotic burden and cardiovascular risk, but remains underused due to access and cost. Artificial intelligence (AI) methods can automate CAC detection and quantification from routine, non-ECG-gated chest CT scans. Prior studies have demonstrated technical feasibility and strong associations between CAC burden and cardiovascular outcomes across diverse populations.

WHAT THIS STUDY ADDS

• This review synthesises evidence on AI-enabled CAC (AI-CAC), focusing on prognostic performance, clinical value and real-world implementation. AI-CAC shows consistent associations with mortality and major adverse cardiovascular events and can be applied opportunistically across imaging contexts, including oncology and screening populations. However, most evidence is retrospective, with limited data on implementation, cost-effectiveness, equity and downstream clinical impact.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• AI-CAC could enable scalable, opportunistic cardiovascular risk stratification and expand access to preventive care without additional imaging. Realising this potential will require prospective implementation studies, standardised reporting frameworks and integration into clinical workflows and care pathways. Policy and health system efforts should address infrastructure, reimbursement and equity to ensure responsible and effective deployment.

Introduction

Coronary artery calcium (CAC) scoring is an established biomarker of atherosclerotic plaque burden and a powerful predictor of future cardiovascular events.¹ The Agatston score, a validated composite metric of CAC density and volume derived from non-contrast CT, is the most widely used method for CAC quantification.² Current US and international guidelines recommend CAC scoring to refine atherosclerotic cardiovascular disease (ASCVD) risk stratification in primary prevention.^3,4 CAC scoring was developed with dedicated non-contrast, ECG-gated cardiac CT scans interpreted by expert readers. Although effective, this process can be resource-intensive, limiting scalability and broad uptake. Despite strong evidence and clear recommendations, clinical use of CAC scoring remains low due to limited clinician awareness, inconsistent insurance coverage, and concerns regarding radiation exposure and incidental findings.⁵ Many eligible individuals never undergo calcium-based risk assessment, reflecting broader underuse of guideline-directed ASCVD risk estimation tools.

Chest imaging obtained for non-cardiac purposes, such as low-dose CT (LDCT) for lung cancer screening or non-contrast chest CT for pulmonary or oncologic evaluation, frequently contains valuable ‘incidental’ coronary calcium information.⁶ Evaluating these existing scans for CAC, termed opportunistic CAC detection, could expand preventive cardiology by leveraging nearly 19 million non-cardiac chest CTs performed annually in the USA, compared with roughly 1 million dedicated CAC scans.⁵ Lipid-lowering therapy based on incidental CAC is now guideline-recommended in the 2026 American College of Cardiology/American Heart Association (ACC/AHA) Guideline on the Management of Dyslipidaemia.⁷ Using existing imaging avoids barriers such as additional radiation, costs or patient visits and can identify subclinical atherosclerosis across more diverse populations.⁸ However, CAC quantification on non-ECG-gated chest CT is not standardised. Variability in CT acquisition parameters, motion artefact and lack of uniform scoring protocols introduce gaps in reproducibility and comparability with guideline-endorsed Agatston scoring from dedicated, ECG-gated studies.⁹ These limitations constrain consistent clinical integration of opportunistic CAC assessment.

Artificial intelligence (AI) has emerged as a powerful enabler of this expansion.^10,11 AI algorithms can automate CAC detection and accurately quantify calcium burden,^12,13 potentially accelerating interpretation and improving reproducibility from non-gated CT scans. Automation could transform opportunistic imaging data into actionable cardiovascular risk insights, improving efficiency and equity in preventive care.

More than 60 published and proprietary AI algorithms for CAC detection and scoring on chest imaging (AI-CAC) have been reviewed in the last 2 years,^12–15 varying in populations, imaging protocols, training datasets, validation approaches and model architectures. Reviews generally focused on algorithmic performance rather than clinical implementation. As AI-CAC transitions from research to practice, important questions arise about workflow integration, interpretability, data interoperability, clinician engagement and cost-effectiveness.¹⁶ AI implementation for real-world impact depends not only on algorithm accuracy but also on designing, testing, monitoring and adapting systems to ensure proper reporting, follow-up and outcomes.¹⁷ Evaluating prognostic utility and clinical impact is essential to advance these tools toward prospective, patient-centred applications.¹⁸

In response to these gaps, this scoping review aims to map and characterise the existing literature on AI-CAC that extends beyond algorithm development to demonstrate prognostic or clinical implications. Specifically, we:

1. Describe populations and clinical contexts in which these tools have been applied.

2. Summarise reported prognostic associations and clinical impacts.

3. Highlight implementation considerations, including workflow integration, ethical challenges and economic value.

4. Identify next steps for advancing AI-CAC toward prospective, scalable implementation.

By synthesising this evidence, we aim to clarify the current state of clinically oriented AI-CAC research, assess readiness for clinical use and outline opportunities for responsible integration into preventive cardiovascular care.

Methods

This scoping review followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) framework. A scoping rather than systematic review design was selected to map the breadth of a rapidly evolving and heterogeneous literature on the clinical implications and implementation of AI-enabled CAC assessment rather than to quantitatively synthesise diagnostic performance outcomes. See online supplemental appendix 1 for the a priori protocol, search strategy, data charting, synthesis framework and PRISMA-ScR checklist. The review was designed to characterise how AI-CAC models have been applied beyond detection and quantification, with particular focus on studies using AI-CAC for prognostication or to inform clinical decision-making. Eligible studies were peer-reviewed original research that met all of the following criteria:

1. Used or incorporated an AI model involving CAC derived from chest CT imaging.

2. Applied the AI-CAC model for clinical prognostication, risk prediction or assessment of potential clinical interventions (eg, treatment initiation, workflow improvement).

3. Reported either clinical outcomes, prognostic metrics or implementation-related insights.

Eligibility assessment was guided by a predefined screening framework. Studies were excluded if they focused solely on the technical development or validation of AI-CAC detection or scoring algorithms without evaluation of prognostic, clinical or implementation-related applications. Studies without human imaging data, conference abstracts without sufficient methodological detail, narrative reviews and editorials were also excluded.

A comprehensive literature search was conducted in PubMed, Embase and Cochrane databases through November 2025, supplemented by manual search of study references and expert recommendations from the author team for three studies not yet indexed at the time of search. Nested Knowledge’s AI-assisted relevance ranking was used to prioritise the order of title and abstract screening (https://nested-knowledge.com/), but all retrieved records were manually reviewed and no study was excluded solely on the basis of algorithmic ranking. A single reviewer screened all titles and abstracts and full-text decisions were made by the same reviewer (VJFM). Outcome definitions varied substantially across studies; therefore, clinical endpoints were extracted as reported and coded into predefined thematic categories to enable structured synthesis. Data were charted using a structured 25-variable template using an AI-assisted strategy with independent review of a 10% sample of the data. See online supplemental table 1 for all charted data. Results were synthesised using a translational maturity framework.¹⁷ Patients and/or members of the public were not involved in the design, conduct, reporting or dissemination plans of this research.

Results

We identified 695 total records across three databases and three expert recommendations. Following the removal of duplicates and initial title and abstract screening, we excluded 560 records as not relevant. We assessed the remaining 135 articles in full for eligibility. Of these, 97 studies described AI-CAC model development and performance but provided no discussion of implementation considerations, clinical integration or downstream implications beyond model performance metrics. As this body of work has been synthesised in recent reviews,^13–15 including one meta-analysis,¹² these studies were excluded from the present analysis. A total of 38 studies met the inclusion criteria and were included in the final review; all were published between 2020 and 2025. The study selection process is summarised in the PRISMA-ScR flow diagram (figure 1). This approach allowed us to focus on AI-CAC studies that extend beyond algorithm validation, emphasising clinical use cases.

Figure 1. PRISMA-ScR flow diagram summarising the screening and inclusion process for studies evaluating artificial intelligence (AI) models related to coronary artery calcium (CAC). CDSR, Cochrane Database of Systematic Reviews; CENTRAL, Cochrane Central Register of Controlled Trials; PRISMA-ScR, Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Review.

29 studies evaluated the retrospective application of AI-CAC models for prognostication of outcomes, including mortality, major adverse cardiovascular events (MACE) and ASCVD. Five studies reported the retrospective application of radiomic feature-based AI models incorporating CAC for outcome prediction. Four studies employed AI-CAC models to identify opportunities for clinical improvement, namely statin prescriptions. Table 1 summarises key study characteristics.

Table 1. Characteristics of included AI-CAC prognostication and implementation studies.

Author	Location	Cohort size	Mean age (years (SD*))	Per cent female	Primary imaging indication	Imaging protocol	AI architecture/software and radiomics software (open source vs proprietary)
Machine learning and deep learning studies
Dekker et al (2020)33	Netherlands	150	67.6 (11.5*)	36	Suspected obstructive CAD	CTAC from PET MPI	Unnamed deep learning algorithm (proprietary)
Han et al (2020)52	South Korea	4915	54.9±9.7	34	Health check-up and screening in asymptomatic individuals	Non-ECG-gated non-contrast chest CT	Boosted ensemble algorithm (LogitBoost) with decision stumps as base classifiers (open)
Stemmer et al (2020)32	USA	12 332	61.8 (5.1*)	41	Lung cancer screening	LDCT	Fully automated machine learning algorithm: CCS-Alg (proprietary Zebra Medical Vision)
de Vos et al (2021)51	USA	1113	61	38	Lung cancer screening	LDCT	Two convolutional neural networks: (1) atlas registration network for image alignment, (2) per-slice direct calcium scoring network (proprietary)
Dekker et al (2021)44	Netherlands	747	67±10	50	Myocardial perfusion imaging (82Rb-PET/CT)	PET MPI	Deep learning algorithm (validated method from Lessmann et al) (proprietary)
Gal et al (2021)19	Netherlands	15 915	59.0 (range 22–95)	100	Breast cancer radiotherapy planning	Non-ECG-gated non-contrast chest CT	Deep learning calcium scoring algorithm (proprietary)
Atkins et al (2022)70	USA	428	DL-CAC=0: 59 (median) DL-CAC ≥1: 69 (median)	49	Radiotherapy planning for locally advanced lung cancer	RT planning CT	Convolutional neural network with 3 consecutive networks for segmentation (proprietary)
Kim et al (2022)20	South Korea	1111	50 (24–87) median	100	Breast cancer treatment	RT planning CT or PET-CT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Balbi et al (2023)71	Italy	4098	60 (55–64)	39	Lung cancer screening	LDCT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Miller et al (2023)72	USA and Canada	2271	66.4	49	Suspected CAD	CTAC from SPECT MPI	Convolutional LSTM for CAC vessel segmentation and DenseNet for cardiac silhouette (proprietary)
Peng et al (2023)23	USA	5678	60.5±16.2	51	Various indications for routine non-ECG-gated chest CT	Non-ECG-gated non-contrast chest CT	Deep learning algorithm (proprietary algorithm from Bunkerhill Health)
Pieszko et al (2023)73	USA and Canada	9543	65 (56–73 IQR)	47	Assessment of known or suspected CAD	CTAC from PET MPI	ConvLSTM algorithm (proprietary)
Ruggirello et al (2023)50	Italy	6495	Females: 58 (54–63); Males: 59 (54–64) median (IQR)	37	Lung cancer screening	LDCT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Shen et al (2023)24	China	1468	<60 years: 851 (58%); ≥60 years: 617 (42%)	47	Pretreatment staging for diffuse large B-cell lymphoma	Non-ECG-gated non-contrast chest CT	3D U-Net architecture (proprietary AI-CACS software by ShuKun Technology)
Wang et al (2023)29	China	1273	58±14	46	Suspected CAD in hypertensive patients	CCTA	XGBoost (also tested: Random Forest, SVM, Neural Network, Logistic Regression) (code unavailable)
Dobrolinska et al (2024)45	Netherlands	572	58±9	46	Suspected CAD	CTAC from PET MPI	Deep learning approach for automatic scoring; Visual estimation for LDCT (proprietary)
Hu et al (2024)28	China	469	64 (56–70) median	38	Suspected CAD in T2DM patients	CACS or CCTA	3D U-net architecture (Proprietary CACScoreDoc, ShuKun Network Technology, Beijing)
Jian et al (2024)49	China	1157	59.0 (11.2*)	Not reported	Suspected obstructive CAD	CACS and CCTA	SVM with 9 clinical features: age, gender, hypertension, diabetes, diabetes duration >5 years, CVD history, RCC, LVESD, BNP (open)
Marjanski et al (2024)21	Poland	193	67.4±6.9	7	Surgical treatment for NSCLC	PET CT and LDCT	Syngo.via Siemens Healthcare (proprietary); Cardiac age via Hoff method (proprietary)
Miller et al (2024)46	13 centres (USA, Canada, Israel, Italy, UK, Switzerland)	45 252	64.7±11.8	44	Suspected or known CAD	CTAC from SPECT	3D U-net for CAC segmentation; nnUnet/TotalSegmentator for cardiac chamber segmentation (open)
Miller et al (2024)47	USA	7613	64 (56–72) median	51	Suspected CAD	CTAC from SPECT MPI	3D U-net for CAC segmentation; nnUnet/TotalSegmentator for cardiac chamber segmentation (open)
Park et al (2024)22	South Korea	1002	62.4±5.4	1	Lung cancer screening	LDCT from PET/CT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft); Syngo.via Siemens Healthcare (proprietary)
Williams et al (2024)74	USA	11 063	62	48	Suspected CAD; myocardial perfusion imaging	CACS; CTAC from SPECT or PET MPI	Convolutional LSTM for CAC vessel segmentation and DenseNet for cardiac silhouette
Cho et al (2025)48	South Korea	289	66 (59–73) median	42	Suspected CAD	CTAC from PET MPI	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Hagopian et al (2025)75	USA	8847	Not reported	Not reported	Lung cancer screening; suspected CAD	LDCT; paired non-ECG-gated non-contrast and CACS	2D Swin-UNETR (U-net variant) (open)
Kim et al (2025)76	South Korea	193	61.6±5.2	0	Lung cancer screening	LDCT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Kim et al (2025)27	South Korea	3153	60–79 (57%)	37	Oncologic imaging (cancer staging/ surveillance)	Non-ECG-gated non-contrast chest CT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Sabia et al (2025)30	Italy	536	68.6 (62.4–74.1) median	44	Stage I lung cancer surgical resection	Preoperative chest or chest abdominal CT	3D U-net architecture (proprietary AVIEW-CAC, Coreline Soft)
Zhou et al (2025)34	USA	1713	55±9	41	Asymptomatic adults with intermediate risk of CAD	Non-contrast CAC CT	XGBoost classification algorithm integrating: distance from lesion to reference point, lesion location, reference point location, vessel identity; cardiac structure segmentation via nnUNet/TotalSegmentator; vessel-specific CAC segmentation via deep learning (open)
Radiomics studies
Hoori et al (2024)35	USA	2457	60.7±9.6	48	CAC scoring	Non-contrast cardiac CT	Research software for radiomic feature extraction; elastic net+Cox model for predictive modelling (open)
Marcinkiewicz et al (2024)25	USA	24 401	61	41	Lung cancer screening	LDCT	TotalSegmentator for structure segmentation; PyRadiomics for radiomic feature extraction; XGBoost for predictive modelling (open)
Singh et al (2024)36	USA	1998	60.8	50	CAC scoring	Gated cardiac CT	Research software for radiomic feature extraction; DeepFat for automated EAT segmentation; Cox regression+elastic net for predictive modelling (proprietary)
Liang et al (2025)37	China	459	58	50	Suspected CAD with CACS <100 AU	LDCT and CCTA	PyRadiomics for radiomic feature extraction; random logistic and LASSO methods for predictive modelling (open)
Marcinkiewicz et al (2025)26	USA and Canada	10 480	65	45	Suspected CAD	SPECT MPI	TotalSegmentator for structure segmentation; PyRadiomics for radiomic feature extraction; XGBoost for predictive modelling (open)
Studies prospectively identifying prevention opportunities
Sandhu et al (2023)41	USA	173	70.8	52	Pulm. nodule eval., lung cancer screening, other	Non-gated, non-contrast chest CT	Deep learning algorithm (proprietary algorithm from Bunkerhill Health)
Dudum et al (2026)42	USA	202	67.9	57	Not discussed	Non-gated, non-contrast chest CT	Deep learning algorithm (proprietary algorithm from Bunkerhill Health)
Bouladian et al (2025)43	USA	163	Not reported	Not reported	Pulm. nodule eval., lung cancer screening, other	Non-gated, non-contrast chest CT	Deep learning algorithm (proprietary algorithm from Bunkerhill Health)
Ladak et al (2026)40	Canada	300	60±17	54	Various indications	Non-gated chest CT	LLMs: GPT-4o (v2024-05-13) (proprietary), Meta Llama-3 70B-Instruct, Meta Llama-3 8B-Instruct (open)

*SD.

AI, artificial intelligence; AU, Agatston units; BNP, B-type natriuretic peptide; CAC, coronary artery calcium; CACS, CAC scan; CAD, coronary artery disease; CCS-Alg, coronary calcium score algorithm; CCTA, coronary CT angiogram; ConvLSTM, convolutional long short term memory; CTAC, CT attenuation correction; CVD, cardiovascular disease; 3D, three-dimensional; DL-CAC, deep learning CAC; EAT, epicardial adipose tissue; LASSO, least absolute shrinkage and selection operator; LDCT, low dose CT; LLMs, large language models; LVESD, left ventricular end-systolic diameter; MPI, myocardial perfusion imaging; NSCLC, non-small cell lung cancer; PET, positron emission tomography; RCC, red cell count; RT, radiotherapy; SPECT, single-photon emission CT; SVM, support vector machine; T2DM, type 2 diabetes mellitus.

Study characteristics

Patient population demographics

The included studies were conducted across North America, Europe and Asia. The USA (17 studies) and South Korea (6) were the most common locations, followed by China (5), Canada and the Netherlands (4 each) and Italy (3). Additional studies originated from Poland and a multicentre international consortium spanning 13 sites across the USA, Canada, Israel, Italy, the UK and Switzerland. Patient populations included large-scale screening or registry datasets such as the US National Lung Screening Trial and hospital-based or academic medical centre populations. Cohort sizes ranged from 150 to 45 252 participants, with the majority including between 400 and 5000 participants.

Participants were predominantly middle-aged to older adults, with a cohort-weighted average age of 61.9 years. Age distributions were slightly younger in general population or screening cohorts and older in cohorts with pre-existing cardiovascular disease or oncologic conditions. The studies were mixed-gender populations, averaging 40%–45% women, except four cohorts that were either male-dominant or focused exclusively on one sex.^19–22 Few studies reported race/ethnicity and, when reported, it was variable.^23–26

The study populations were highly heterogeneous with respect to comorbidities, including hypertension, diabetes mellitus, dyslipidaemia and smoking history variably reported across cohorts. Seven studies focused on specialised populations, including individuals with breast cancer,^19,20 colorectal cancer,²⁷ diabetes²⁸ or hypertension²⁹ as inclusion criteria; and postoperative cohorts.^21,30

Imaging indications and protocols

The studies included an array of cardiovascular and non-cardiovascular imaging indications (see primary imaging indication in table 1). 14 studies enrolled patients undergoing imaging for suspected or known CAD, including those with hypertension, diabetes or stable angina, as well as patients with known low Agatston scores (<100 Agatston units). 10 studies leveraged LDCT or non-ECG-gated chest CT performed for lung cancer screening or pulmonary nodule evaluation. Seven studies used CT scans obtained for cancer care, including breast cancer radiotherapy planning, lung cancer or lymphoma staging, oncologic surveillance and surgical planning for lung cancer. Four assessed asymptomatic individuals for cardiovascular risk. Three studies included routine non-gated chest CT performed for other clinical indications, opportunistically assessed for CAC scoring.

AI algorithm/architecture

AI methods for CAC detection and scoring were dominated by deep learning algorithms, particularly 3D U-net convolutional neural network (CNN) architectures for CAC segmentation and Agatston scoring (see AI architecture/algorithm in table 1). Other models included atlas-based segmentation networks, multistage CNNs and hybrids integrating cardiac chamber segmentation. A smaller number used machine learning approaches such as XGBoost, support vector machines and ensemble approaches to combine clinical features (eg, age, sex, comorbidities, lab values) with CAC or radiomics features to predict outcomes.

Radiomics is the rapidly evolving field of extracting high-dimensional quantitative metrics, also called radiomics features, from medical images, including complex data that may not be recognisable to the human eye.³¹ The radiomics-based studies move beyond CAC detection/quantification to extract additional quantitative layers of data from imaging, including features from CAC, epicardial fat or surrounding thoracic structures.

Clinical impact

Outcomes and prognostic value

The included studies targeted a wide range of primary outcomes, primarily centred on mortality and MACE, identification of high-risk phenotypes and statin prescription rates and opportunities. Secondary outcomes spanned adverse cardiovascular events, functional imaging metrics, surgical or oncologic outcomes and implementation-focused measures (figure 2). Studies included both diagnostic ECG-gated CAC imaging and opportunistic calcium assessment from non-gated chest CT and CT attenuation correction (CTAC) imaging; findings are therefore presented descriptively without modality-specific stratification. Table 2 provides a summary of representative studies illustrating the range of prognostic outcomes reported across AI-CAC applications.

Figure 2. Summary of primary and secondary outcomes for AI-CAC prognostication and implementation studies. AI, artificial intelligence; BMI, body mass index; CAC, coronary artery calcium; CAD, coronary artery disease; CCTA, coronary CT angiography; CT-FFR, CT-fractional flow reserve; CV, cardiovascular; MACE, major adverse cardiac event; MI, myocardial infarction.

Table 2. Prognostication results of select AI-CAC models.

Outcome (study)	Predictor	Imaging indication	Effect size/HR/OR	Key insights
ACM70	AI-CAC	RT planning for lung cancer	CAC >1, HR 1.51	CAC detectable predicts mortality in lung cancer patients
CVD mortality32	AI-CAC	Lung cancer screening	CAC 101–400 OR 1.72; CAC >400 OR 2.62	CAC strongly predicts CVD mortality in lung cancer screening
Obstructive/haemodynamically significant CAD28	AI-CAC	Suspected CAD in T2DM patients	Adjusted OR obstructive CAD 1.005, CT-FFR ≤0.8 OR 1.003	CAC predicts functional CAD in T2DM patients
Postoperative mortality after lung cancer resection30	AI-CAC	NSCLC postresection	CAC 100–399 HR 1.78 CAC ≥400 HR 2.37	Higher CAC associated with increased risk of postoperative mortality
Cancer therapy–related cardiac dysfunction24	AI-CAC	Staging CT for DLBCL	CAC 1–100 OR 2.59 CAC >100: 5.24	CAC predicts CTRCD post-treatment with anthracyclines
ACM25	AI-CAC, AI-EAT and radiomics features of chest structures	Lung cancer screening	AUC 0.723 (0.71, 0.736) for 10-year ACM across all models	CAC had highest significance for ACM but other features capture risk from many disease processes
MACE35	Calcium-omics	CAC scoring for CLARIFY programme screening	HR 3.62 for 39 calcium features predicting MACE	Number of calcifications and LAD mass improve MACE prediction
Statin prescription within 6 months of notification41,42	Patient and clinician notification based on AI-CAC	Various	51.2% in notification arm vs 6.9% P<0.001 44% in notification arm vs 9.8% P<0.001	Notifications re AI-CAC improve preventive therapy adherence
Statin persistence at 18 months postnotification43	Patient and clinician notification based on AI-CAC	Various	87% of patients initiated on statin after notification persisted on statin at 18 months	Patients who started lipid-lowering therapy after notification re AI-CAC persisted on therapy at 18 months postnotification

AI, artificial intelligence; AUC, area under curve; CAC, coronary artery calcium; CAD, coronary artery disease; CT-FFR, CT-fractional flow reserve; CTRCD, cancer therapy related cardiac dysfunction; DLBCL, diffuse large B-cell lymphoma; EAT, epicardial adipose tissue; LAD, left anterior descending; NSCLC, non-small-cell lung cancer; RT, radiotherapy; T2DM, type 2 diabetes mellitus.

Across heterogeneous imaging modalities and clinical contexts, studies generally reported associations between AI-CAC and outcomes including obstructive coronary artery disease (CAD), MACE and mortality, although effect sizes and study populations varied. Risk increased with CAC burden, for example, CVD for CAC 101–400, OR 1.72; CAC >400, OR 2.62.³² AI-CAC was predictive of haemodynamically significant CAD (CT-Fractional Flow Reserve ≤0.8),²⁸ postoperative complications,²¹ and cancer therapy-related cardiac dysfunction.²⁴ AI-CAC models generally relied on traditional categorical CAC (eg, Agatston score risk categories of 0, 1–99, 100–399, ≥400) along with clinical features as core inputs, although several approaches extended quantification to the level of individual coronary vessels or segments to capture regional burden and distribution.^33,34

Radiomics-based approaches sought to further refine risk stratification by extracting higher-dimensional imaging features beyond the Agatston score, including plaque texture, spatial distribution, density characteristics and lesion morphology.^{25,26,35–37} However, radiomic features are highly sensitive to CT acquisition and reconstruction parameters; features identified as strongly predictive in one scanner environment or protocol may not generalise to another.^38,39 Moreover, no included radiomics studies were prospective; these approaches should be treated as hypothesis-generating and require prospective validation before clinical deployment can be considered.

In real-world applications, large language models accurately identified patients for preventive interventions from review of patient charts and CT reports (F1 0.99, positive predictive value 98%–100%).⁴⁰ Patient and clinician notifications based on AI-CAC results improved statin prescription and downstream lipid testing, statin persistence, low-density lipoprotein cholesterol (LDL-C) reduction and CAD evaluation.^41–43

Clinical implications

AI-CAC enables opportunistic cardiovascular risk detection from routine chest CT. Algorithms generally perform two core tasks: automated detection of CAC and quantification of calcium burden, typically using qualitative or categorical CAC scoring (eg, 0, 1–99, 100–399, ≥400) or continuous Agatston-equivalent estimates.

Integration of AI-CAC with CTAC scans associated with single-photon emission CT (SPECT) or PET CT myocardial perfusion imaging improved the detection of obstructive or haemodynamically significant CAD, reducing false-negative results and enhancing diagnostic accuracy. AI-CAC also identified high-risk CAD phenotypes in diabetes and hypertensive populations, which could potentially guide CCTA utilisation and early intervention.

Across multiple cohorts, AI-CAC independently predicted all-cause mortality and MACE with HRs typically ranging from 1.5 to 3.6 for high CAC categories (≥100–400).^{20,22,27,34,35,44–48} AI-CAC quantification provided prognostic value beyond traditional risk factors, even in asymptomatic or non-cardiac imaging populations, validating its role in opportunistic risk stratification.

Agentic workflows can be used at the panel-management level to identify patients in the electronic health record with elevated CAC without statin prescription. Notification of patients and clinicians based on these results has been shown to increase cardiovascular preventive discussions, statin prescription rates and statin persistence.^40–43 Integration of AI-CAC into clinical workflows could thus improve population-level preventive therapy delivery. In oncology care, automatic CAC detection may inform cardioprotective therapy planning and radiotherapy optimisation.

Deployment considerations

Workflow integration

These studies demonstrated that AI-CAC could be technically integrated into existing imaging workflows with minimal or no additional imaging or radiation exposure for CAC scoring. In the NOTIFY-1 and NOTIFY-PICTURE trials, AI-CAC was integrated using a hybrid workflow combining automated detection and EHR screening with partial manual validation, with potential for further automation at scale.^41,42 Study characteristics related to workflow integration are described in table 3.

Table 3. AI-CAC workflow integration factors and benefits of implementation.

Characteristic	Examples	Benefits to workflow
Integration with existing imaging protocols	Non-ECG-gated chest CT Attenuation correction CT (eg, with PET or SPECT) Low-dose lung cancer screening CTs	Enable CAC quantification without new imaging Avoid additional radiation
Processing time and automation efficiency	1–2 min per patient on standard CPU/GPU Full automation and batch processing Automatic vessel labelling and Agatston categorisation	Significant reductions in time/labour compared with manual scoring Maintaining excellent reproducibility Seamless integration into PACS or structured reporting systems
Scalability and applicability	Large scale or retrospective screening potential Deployed on intrahospital servers Incorporated into routine workflows	High-throughput, reproducible analyses Potential application to millions of chest CTs performed annually in USA
Reporting and interoperability	Auto-generated CAC reports transferred to PACS Commercial and FDA-approved software options	Standardisation of reports*
Population and clinical workflow innovations	Auto-generated patient notification by large language model Increased statin prescription rates for incidental CAC	Automated identification of opportunities for preventive therapy

^*Studies using commercial software (eg, Siemens Syngo.via, Coreline AVIEW) highlighted the need for standardisation across vendors and scanners for consistent results.

AI, artificial intelligence; CAC, coronary artery calcium; CPU/GPU, central processing unit/graphics processing unit; FDA, U.S. Food and Drug Administration; PACS, picture archiving and communication system; PET, positron emission tomography; SPECT, single-photon emission CT.

These studies demonstrated rapid processing time, scalability for retrospective application of models to previously acquired CT scans, and ease of integration into radiology reporting systems. Studies did not discuss workflow elements of confirming AI-CAC reports, acting on results, or monitoring and correcting drift across longitudinal model performance, which is important to ensure that deployment leads to the intended outcomes and maintains robust performance over time.¹⁸ Broader implementation gaps including infrastructure requirements, human-in-the-loop versus automated workflows, stakeholder engagement and capacity for follow-up care at scale remain unaddressed and represent important directions for future investigation.

Bias and transparency considerations

All studies reported institutional review board or ethics committee approval, with some discussion of bias, fairness, and model transparency. A few used explainable methods such as Shapley additive explanations to enhance interpretability.^{24–26,29,34,35,37,49} One framework referenced the Proposed Requirements for Cardiovascular Imaging–Related Machine Learning Evaluation (PRIME) checklist for cardiovascular imaging AI as an emerging reporting standard, reflecting increasing awareness of the need for standardised evaluation and transparent reporting.⁴⁷

Value and economic impact

Six studies highlighted the potential of AI-CAC models to enable low-cost cardiovascular risk assessment using existing non-contrast, non-ECG-gated chest CTs, without requiring additional imaging or radiation exposure.^{19,29,36,40,41,50} Four reported that fully automated pipelines could reduce radiologist labour time and improve throughput compared with manual scoring, supporting scalability for opportunistic screening.^20,25,28,51 Only a minority quantified economic impact, including one estimating a cost of US$0.006 per screened patient using automated notification workflows (≈US$6485 per million patients)⁴⁰ and another reporting opportunistic CAC analysis costs under US$100 compared with substantially higher charges for echocardiography.⁵² Formal cost-effectiveness analyses were not performed.

Discussion

We examined the clinical value and deployment considerations of AI-CAC models used to prognosticate health outcomes across diverse patient populations. Across studies, AI-CAC demonstrated strong and consistent associations with all-cause mortality, cardiovascular mortality, MACE and obstructive CAD, including in surgical and oncology cohorts. Radiomics models show potential to extend beyond traditional CAC scoring by incorporating additional imaging features, although no included studies were prospective, limiting clinical translation. AI-CAC models are being used to identify and engage patients and clinicians in preventive therapy opportunities. Overall, the literature suggests that AI-CAC has reached advanced stages of technology readiness,¹⁸ reflecting substantial technical development and validation, but is still in relatively early stages of translational maturity with limited implementation and downstream clinical impact.

The development of AI-CAC mirrors that of many clinical AI tools, with early development dominated by retrospective validation focused on accuracy and reproducibility. These studies established that deep learning models can reliably quantify CAC and stratify cardiovascular risk across heterogeneous imaging protocols and patient populations. More recently, prospective applications have emerged, enabling opportunistic screening from chest CTs performed for various indications. This progression reflects an ongoing translational arc from technical performance into prospective real-world validation and implementation science.

AI-CAC models that generate Agatston scores and guideline-aligned CAC categories integrate naturally with existing ASCVD prevention frameworks, supporting threshold-based decisions regarding statin prescription or intensification.⁵³ However, research has demonstrated the importance of evolving the Agatston score in the era of evidence-based statin and aspirin prescription, non-statin lipid-lowering therapy and the shift towards personalised risk assessment.⁵⁴ In this context, AI-CAC enables scalable assessment of vessel-specific calcium burden and plaque density and composition.⁵⁵ AI-CAC could support evolution toward a more informative, physiologically grounded CAC score, preserving clinical familiarity and, combined with radiomics features or additional image-derived markers such as epicardial adipose tissue, cardiac chamber volumetry, as well as complementary biomarkers such as thoracic aortic calcification and body composition markers if these are shown to be predictive of increased cardiovascular disease,^55–57 could enable more personalised CVD risk stratification. AI methods may also detect subthreshold coronary calcification not captured by conventional Agatston scoring, as in the Agatston 2.0 framework, which uses a relative-density approach with internal calibration to extend assessment below the conventional 130-HU threshold, potentially refining risk stratification within the zero CAC category.⁵⁸ Such approaches remain hypothesis-generating pending prospective validation.

Beyond single-time point assessment, AI-CAC also creates the potential for longitudinal cardiovascular phenotyping. Because many patients undergo repeated thoracic imaging over time, AI-derived CAC measurements could enable serial assessment of CAC progression alongside related imaging features such as plaque distribution, myocardial structure or cardiometabolic markers. Tracking these changes across imaging encounters may allow development of individualised cardiovascular risk trajectories, shifting opportunistic CAC assessment from a static risk marker toward a dynamic tool for monitoring broader cardiovascular disease risk rather than ASCVD alone.

By integrating multidimensional features with longitudinal assessment, AI-CAC could support more personalised preventive strategies, including individualised LDL-C targets, tailored pharmacologic intensity and precision medicine approaches such as polypill deployment. Rather than categorising patients into discrete risk bins, composite AI-derived risk trajectories may better reflect cardiovascular risk as a continuum. Developing and validating such enhanced CAC constructs, however, will likely require evaluation approaches that extend beyond traditional randomised clinical trials, incorporating real-world data, adaptive learning frameworks and longitudinal monitoring to assess clinical utility and downstream outcomes.

Despite promising prognostic performance, most AI-CAC models remain at early stages of clinical implementation. The studies reviewed here are early in the AI translational pathway, with only a few moving beyond AI model output to actual clinical reporting and care pathways, and none addressing longer-term model performance monitoring and outcomes (figure 3). Using established frameworks for AI maturity in cardiovascular care,¹⁷ the majority of studies remain at stage 1 (Initial Development and Validation), with only limited progress to stage 2 (Implementation in Real World Settings). Advancing toward these stages will require prospective deployment within clinical environments, with attention to usability, interpretability and integration into decision-support pathways.

Figure 3. AI-CAC translational pathway. AI, artificial intelligence; CAC, coronary artery calcium; CCTA, coronary CT angiogram; EAT, epicardial adipose tissue; LDCT, low-dose CT; MPI, myocardial perfusion imaging; NGNC CT, non-gated non-contrast CT; PET, positron emission tomography; RT, radiotherapy; SPECT, single-photon emission CT.

Key implementation challenges include technological infrastructure and interoperability between image archiving and electronic health record systems; decisions regarding embedded versus cloud-based processing; and automated versus clinician-triggered workflows. Equally important are questions of responsibility and actionability (ie, who owns AI-CAC findings, how they are communicated and what clinical actions they prompt), particularly when applied opportunistically across broad imaging populations. Finally, sustained implementation will require capacity for follow-up care, including preventive cardiology services and coordinated pathways to translate AI-identified risk into meaningful intervention. A practical AI-CAC integration pathway may involve sequential steps: automated triggering of the AI-CAC model by the imaging or records system; structured result reporting; clinician and patient notification with contextual risk framing; linkage to guideline-based preventive care recommendations and follow-up tracking. Each step introduces implementation challenges that future prospective studies should address.

Ethical considerations surrounding AI-CAC include bias, privacy and consent, transparency and explainability.⁵⁹ Approaches such as federated machine learning may improve inclusivity while preserving data privacy across institutions.⁶⁰ Guidance from the Stanford HEAL AI framework suggests that patient disclosure and consent to AI should depend on both the risk of physical harm and the degree of patient agency afforded by the tool.⁶¹ In terms of transparency, AI models often process input data in ways that no human interpreter can easily explain, but interpretability is crucial for physicians and patients to trust and use the outputs in clinical decision-making.⁶² Given the inherent opacity of many AI models and the risk of automation bias, interpretability methods such as Shapley additive explanations may help build clinician and patient trust in AI-CAC outputs.^63,64

Economic value and workflow utility are similarly critical to the successful integration of AI-CAC models into clinical practice. Although most studies cited time and labour savings compared with manual CAC scoring, none performed formal cost–benefit analyses. Frameworks such as realised net benefit and the APLUS usefulness assessment could be applied to quantify preventive benefit relative to workflow and resource constraints.^65,66 Borrowing from industry, target product profiles may also guide AI-CAC development toward clinically useful, guideline-aligned and financially viable implementations.⁶⁷

Future AI-CAC implementation efforts should align with emerging guidance for responsible evaluation and monitoring of clinical AI tools. Recent AHA recommendations emphasise four guiding principles: alignment with organisational strategic priorities, ethical design and use, demonstrable usefulness and effectiveness and sustainable financial performance.⁶⁸ Applying these principles to AI-CAC shifts the focus from algorithmic accuracy alone toward long-term clinical impact, accountability and integration within learning health systems. Notably, CMS reimbursement pathways now exist for AI-detected CAC,⁶⁹ which may further accelerate the economic case for adoption.

Limitations

This review was a scoping rather than a systematic review, mapping existing literature on AI-CAC models for prognostication and clinical value rather than formally assessing study quality. Most included studies were retrospective and heterogeneous in their study design, demographics, imaging protocols and outcomes, which limits direct comparison or pooled synthesis of results. Reporting of prospective data and of broader deployment and evaluation, including ethical implications, economic value and utility, remains limited. Some included studies were based on overlapping source populations (eg, the National Lung Screening Trial^32,51), and thus the observed consistency of findings across studies should not be interpreted as reflecting entirely independent datasets. Single-reviewer screening is an additional limitation; a dual-reviewer design would have strengthened methodological rigour. Three included studies were identified through coauthor expert recommendation because they were not yet indexed at the time of search; although all met prespecified eligibility criteria, this approach may introduce selection bias. Three included studies^41–43 employed an AI algorithm (Bunkerhill Health) in which co-investigator CPL holds equity interest.

Conclusions

AI-CAC is rapidly evolving as a scalable approach for opportunistic ASCVD risk stratification and identification of preventive care opportunities across diverse patient populations. Across imaging modalities and clinical contexts, AI-CAC demonstrates consistent associations with cardiovascular outcomes, supporting its potential for low-cost risk assessment for the millions of patients who undergo chest CT in the USA annually. Translation to clinical impact remains limited by the predominance of retrospective studies and limited prospective implementation or health economic evaluation. Advancing AI-CAC towards routine clinical use will require greater standardisation, integration into real-world workflows and ongoing monitoring to demonstrate not only predictive accuracy but meaningful improvements in preventive care delivery and patient outcomes.

Data availability statement

All data relevant to the study are included in the article or uploaded as supplementary information.