A comprehensive scoping review examining a decade of cardiovascular imaging research indicates that artificial intelligence algorithms demonstrate strong performance in predicting patient outcomes and identifying vascular pathology during angiographic procedures, according to findings published in the Journal of Cardiovascular Imaging. Investigating literature published between January 2015 and March 2025, researchers Shreeyash Shashank Tulpule and Sammita Jagdish Jadhav synthesized evidence from across the biomedical and engineering spectrum to understand how machine learning systems evaluate coronary artery disease. Since its initial introduction by Dr. Mason Sones in 1958, coronary angiography has served as the diagnostic gold standard for assessing coronary artery disease by detailing the severity and anatomical location of arterial blockages as well as the extent of collateral circulation. Modern cardiac catheterization laboratories generate massive streams of diagnostic information, including high-frame cine-angiographic image sequences, intravascular pressure recordings, and real-time hemodynamic measurements. The authors systematically evaluated how computational architectures process these complex data streams to assist clinicians with diagnostic accuracy, procedural risk stratification, and patient monitoring, highlighting both the technical sophistication of modern automated models and the structural barriers that currently hinder routine deployment in daily bedside care.
To rigorously map the emerging literature, the investigators structured their scoping review using the methodological framework developed by Arksey and O’Malley in 2005 alongside the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews guidelines. The investigative team carried out comprehensive bibliographic queries across four prominent electronic indexing databases, specifically PubMed, Scopus, Web of Science, and IEEE Xplore, combining Medical Subject Headings and free-text terms including artificial intelligence and angiography. Gray literature searches were supplemented using Google Scholar to capture pertinent research outside traditional repository channels. From an initial search yield of 5,695 articles, the investigators employed Rayyan screening software to systematically eliminate 2,485 duplicate records and exclude 2,213 irrelevant citations during preliminary title and abstract assessment. The resulting 1,024 publications underwent thorough full-text evaluation by two independent researchers who achieved substantial inter-rater reliability, evidenced by a Cohen’s kappa coefficient of 0.78. Following the exclusion of 965 papers that lacked relevant predictive modeling, human subject cohorts, or standard quantitative metrics, 59 studies met all predefined criteria and formed the final analytic synthesis.
Methodological quality across the 59 included studies was formally appraised using the Critical Appraisal Skills Programme checklist, establishing that all selected literature satisfied minimum rigorous investigative standards with scores of 16 or higher out of a maximum possible score of 24. Within this evaluated corpus, 36 studies attained quality scores ranging between 16 and 20, while 21 papers achieved high-tier scores between 20 and 24. Publication distribution displayed a pronounced upward trajectory over the decade, concentrating heavily between 2022 and 2024, with 2024 recording the single largest volume of publications in the survey period. This upward momentum continued into the first half of 2025, which contributed nine newly published studies demonstrating sustained scientific momentum. The examined body of work comprised 58 primary empirical investigations alongside one dedicated case study, spanning an array of cardiovascular conditions that included coronary artery disease, heart failure, cardiac arrhythmias, and valvular disorders. Across these investigations, algorithms were tasked with diverse responsibilities ranging from real-time computer-assisted diagnostic support to postprocedural prognostic modeling, illustrating an expanding reliance on algorithmic decision-support platforms within modern interventional cardiology workflows.
The thematic analysis identified four primary clinical domains, the largest of which comprised 34 studies, representing 60 percent of the reviewed corpus, directed at forecasting postprocedural adverse cardiovascular events. These investigations targeted acute in-hospital and perioperative complications, such as major bleeding episodes, acute renal failure, contrast-induced nephropathy, emergent arrhythmias, myocardial infarction, and all-cause mortality. To train these predictive architectures, researchers compiled expansive multidimensional feature sets integrating baseline demographic profiles, including age, biological sex, body mass index, ethnicity, and tobacco exposure, with continuous dynamic physiological telemetry such as serial blood pressure curves, heart rate tracings, and oxygen saturation waveforms. These datasets were further linked with routine laboratory biomarkers, encompassing serum creatinine, serial hemoglobin concentrations, high-sensitivity cardiac troponin, and C-reactive protein. Prior to model exposure, raw clinical data underwent rigorous data conditioning steps, including outlier mitigation, multiple imputation procedures for missing observations, temporal signal summarization, and numerical normalization, enabling ensemble architectures like random forests and gradient boosting machines to capture intricate non-linear interactions across diverse clinical inputs.
The second most prevalent research theme encompassed 18 studies, accounting for 32 percent of the analyzed papers, which focused on the non-invasive detection and functional evaluation of hemodynamically significant coronary artery disease. These investigations primarily centered on high-resolution imaging modalities, predominantly coronary computed tomography angiography, frequently augmented with invasive catheterization frames or intravascular ultrasound runs. Volumetric imaging data were conditioned through automated vessel centerline extraction, isotropic voxel resampling, and image intensity normalization before being ingested by deep convolutional neural network backbones, including U-Net, ResNet-50, and DenseNet-121 architectures. These neural networks performed either voxel-level atherosclerotic plaque segmentation or comprehensive vessel-level obstruction classification, with diagnostic reference standards established by expert panel consensus, quantitative coronary angiography, or invasive fractional flow reserve measurements. Furthermore, approximately one-third of these investigations incorporated multimodal data fusion strategies, utilizing late-fusion vector concatenation or parallel dual-stream deep neural networks to merge imaging representations with systemic cardiovascular risk profiles, including lipid panels, glycated hemoglobin, and blood pressure, producing comprehensive predictive models capable of acting as non-invasive gatekeepers prior to catheterization.
In-stent restenosis prediction formed a highly specialized third thematic focus, evaluated across four studies representing seven percent of the examined literature. These investigations concentrated on post-percutaneous coronary intervention cohorts to identify individuals at elevated risk of luminal renarrowing on surveillance angiography. The modeling pipelines combined static anatomical and procedural metrics extracted from catheterization logs, such as implanted stent length, nominal stent diameter, lesion segment characteristics, and post-deployment residual stenosis percentages, with longitudinal postprocedural follow-up indicators. These dynamic longitudinal inputs incorporated patient adherence records regarding dual antiplatelet therapy regimens, interval lipid profile measurements, and pharmacogenomic data concerning clopidogrel metabolism. Because clinical restenosis occurs in a relatively small proportion of stented vessels, datasets exhibited severe positive-class sparsity. To counteract extreme class imbalance, investigators leveraged tree-based ensembles, including gradient boosting algorithms and random forests, which support internal class weighting, and supplemented their workflows with synthetic minority oversampling techniques and focal loss functions. Feature importance analyses consistently demonstrated that mechanical stent dimensions and antiplatelet medication adherence exerted greater predictive weight over vessel patency than standard demographic covariates.
The fourth major clinical theme, addressed in 12 studies representing 21 percent of the literature, evaluated automated coronary artery identification, vessel segmentation, and lesion characterization across invasive coronary angiography, computed tomography angiography, and sequential chest radiographs. Employing diverse deep learning architectures, such as convolutional networks, vision transformers, and specialized hybrid frameworks including UTNets, MI-MS ConvMixer, YOLO, and U-Net variants, these studies demonstrated high visual recognition fidelity. Performance metrics across these segmentation investigations documented Dice similarity coefficients exceeding 0.90 in vessel contouring tasks, accompanied by area under the receiver operating characteristic curve values surpassing 0.90 for detecting significant luminal stenosis. Image preprocessing pipelines operated on raw pixel-level angiographic sequences, utilizing radiopaque contrast media enhancement and multi-frame temporal alignment to ensure continuous vessel edge tracing throughout the cardiac cycle. Investigators frequently utilized quantitative coronary angiography as an objective benchmark to validate automated lumen diameter calculations, while exploratory investigations extended predictive screening paradigms to non-invasive chest radiography scoring to assess baseline cardiovascular risk prior to invasive intervention.
Despite widespread technical successes, the scoping review uncovered critical methodological vulnerabilities across the analyzed literature, most notably a pervasive reliance on small cohorts and single-center datasets. Thirty-two studies, representing more than half of the included evidence, were restricted to single hospital environments, which sharply limits sample size, reduces statistical power, and risks overfitting algorithms to localized scanner hardware, imaging protocols, and patient demographics. Furthermore, retrospective observational designs characterized the vast majority of studies, with six publications explicitly detailing how retrospective electronic health record mining introduces profound cohort selection bias and unmeasured confounding. In retrospective contexts, clinical measurements are gathered opportunistically rather than systematically, obscuring critical chronological sequences, such as whether therapeutic interventions preceded or followed subtle physiological deteriorations. Additionally, five studies acknowledged evaluating their models exclusively on internal data partitions without external validation cohorts. The authors noted that without testing across external healthcare systems or differing regional patient populations, algorithmic generalizability remains entirely speculative, as models may learn institution-specific workflow artifacts rather than genuine pathological markers.
Translational barriers within clinical workflow implementation, system interpretability, and regulatory governance represented another major gap highlighted in the scoping review. Among the 58 evaluated primary studies, very few explored concrete implementation pathways, such as real-time electronic health record interoperability, automated alerting protocols, or user-friendly interfaces tailored to cardiac catheterization laboratory staff. Furthermore, fewer than 10 studies incorporated formal explainable artificial intelligence methodologies, such as Shapley Additive Explanations or Local Interpretable Model-agnostic Explanations, relying instead on opaque black-box deep neural networks and XGBoost algorithms that obscure their predictive reasoning from treating clinicians. Economic and regulatory analyses were completely absent from the surveyed literature, with zero studies presenting formal cost-effectiveness data, payer reimbursement engagement strategies, or documented compliance pathways for United States Food and Drug Administration clearance or European Conformity certification. Additionally, head-to-head comparisons against established clinical risk scores, including the Global Registry of Acute Coronary Events, TIMI, and SYNTAX scores, were conducted in only five studies, demonstrating that while artificial intelligence consistently exhibited superior predictive discrimination, standardized benchmark evaluations remain exceptionally scarce.
Data governance and international representation presented additional systemic challenges, with patient privacy concerns forming a significant barrier to large-scale data pooling across institutions. The review highlighted that despite data de-identification and masking procedures, significant re-identification vulnerabilities persist, citing external evidence from a 2018 United States National Health and Nutrition Examination Survey analysis in which machine learning algorithms successfully re-identified 85.6 percent of adults and 69.8 percent of children from physical activity tracking cohorts lacking direct personal identifiers. Geographical publication trends revealed marked disparities as well, with the vast majority of research originating from single regions such as China, while substantial gaps in research output were noted across South America, Africa, and major parts of Asia, including India where coronary artery disease prevalence is particularly severe. Ultimately, the authors noted that while artificial intelligence models possess considerable potential to refine risk stratification and streamline cardiac catheterization lab workflows, their routine clinical adoption hinges on prospective multicenter clinical trials, rigorous standardized reporting frameworks, transparent interpretability, and robust regulatory oversight.
Subject of Research: Scoping review of artificial intelligence applications in coronary angiographic outcome prediction and diagnostics
Article Title: Utility of artificial intelligence in predicting clinical outcomes of angiographic procedures: a comprehensive scoping review
Article References: Tulpule, S. S., & Jadhav, S. J. (2026). Utility of artificial intelligence in predicting clinical outcomes of angiographic procedures: a comprehensive scoping review. Journal of Cardiovascular Imaging, 34(1), Article 24. https://doi.org/10.1186/s44348-026-00064-x
Image Credits: AI Generated
DOI: 10.1186/s44348-026-00064-x
Keywords: artificial intelligence, coronary angiography, scoping review, cardiovascular imaging, coronary artery disease, in-stent restenosis, hemodynamically significant stenosis, convolutional neural networks, machine learning, random forest, computed tomography angiography, clinical risk scores, fractional flow reserve, explainable artificial intelligence, patient data privacy, model validation, percutaneous coronary intervention
News Source: Blake Davidson. (November 18, 2026). Artificial Intelligence Shows Diagnostic and Prognostic Value Across Coronary Angiography Studies. Scienmag.



