A narrative review on the use of artificial intelligence in cardiovascular medicine

Introduction

Cardiovascular diseases (CVDs) account for nearly 18 million deaths annually and remain the leading cause of morbidity and mortality worldwide despite substantial advances in prevention, diagnosis, and management (1). The global rise in life expectancy, increasing multimorbidity, and the complex nature of cardiovascular pathologies have intensified the need for scalable, efficient, and personalised care strategies (2). Artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL), has emerged as a transformative force that can analyse high-dimensional clinical and imaging datasets, enabling earlier diagnosis, more precise risk stratification, and individualised therapeutic interventions (3-5).

AI applications in cardiology now span a broad range of areas. In diagnostics, AI enhances electrocardiogram (ECG) interpretation, detecting subtle abnormalities and predicting arrhythmias, structural disease, or even biochemical markers from a single tracing (6). In cardiovascular imaging, AI algorithms automate the acquisition, segmentation, and quantification processes across various modalities, including echocardiography, cardiac computed tomography angiography (CTA), and magnetic resonance imaging (MRI). These advances reduce interobserver variability, improve workflow efficiency, and provide more reproducible measurements (7). For clinical decision support, AI-driven risk models outperform traditional scores in predicting outcomes such as heart failure hospitalisation, myocardial infarction, or sudden cardiac death (8). AI is also being integrated into interventional cardiology, supporting intravascular imaging, stent selection, and robotic-assisted procedures (9-11).

AI has extended its reach beyond the hospital setting into cardiac rehabilitation and outpatient monitoring. Wearables equipped with AI-enabled sensors provide continuous rhythm analysis, activity tracking, and early detection of clinical deterioration, facilitating remote supervision and personalised rehabilitation plans as shown in Figure 1 (12).

Figure 1 Applications of AI in cardiovascular medicine. Distribution of AI applications across major domains in cardiovascular medicine, including imaging, ECG interpretation, risk stratification, interventional cardiology, remote monitoring, and rehabilitation. Data are illustrative of proportions seen in recent literature. AI, artificial intelligence; ECG, electrocardiogram.

In the context of congenital heart disease and critical care cardiology, AI is being deployed for anatomical modelling, shock phenotyping, and ventilator optimisation, demonstrating its utility even in specialised or high acuity domains (13,14).

Despite this rapid growth, significant barriers hinder widespread clinical implementation. Most AI models demonstrate excellent performance in internal validation but show reduced accuracy when deployed in external populations due to dataset shift and variability in imaging protocols or electronic health record (EHR) systems (15). Explainability remains a challenge, with many DL approaches operating as “black boxes”, limiting clinician trust and regulatory approval (16). Ethical concerns, including algorithmic bias, data privacy, and accountability in AI-driven decision making, further complicate the integration of AI into clinical workflows (17,18). Additionally, interoperability challenges, regulatory uncertainties, and a lack of prospective trial evidence delay the transition from prototype to practice as shown in Figure 2 (19,20).

Figure 2 AI in heart disease. Proportion of studies utilising different AI techniques in cardiovascular research, including machine learning, deep learning, hybrid approaches, and natural language processing. Percentages reflect general trends in recent publications. AI, artificial intelligence.

AI fundamentals and current clinical value

AI in cardiovascular medicine encompasses a spectrum of computational techniques, broadly categorised into supervised, unsupervised, and reinforcement learning (RL) frameworks. Supervised ML, the most clinically mature form, relies on labelled data to train models for tasks such as arrhythmia detection, left ventricular ejection fraction (LVEF) estimation, and coronary artery stenosis prediction. Popular supervised models include convolutional neural networks (CNNs), gradient boosting machines, and ensemble learning methods, which have demonstrated strong performance in ECG interpretation and imaging-based diagnostics (6,7). Unsupervised learning is increasingly being used in cardiovascular research to uncover novel phenotypes and disease subtypes without predefined labels. Notable applications include phenotype-based clustering of heart failure with preserved ejection fraction (HFpEF), which leads to more nuanced risk stratification and a potential for precision therapeutics (18,21). RL, although less commonly applied than supervised or unsupervised methods, is emerging as a promising approach for adaptive treatment optimisation in critical care and hemodynamic management (Figure 3). RL agents interact with dynamic patient data to learn optimal decision pathways, and early work has demonstrated potential in vasopressor titration and mechanical support strategies for cardiogenic shock (22). Additionally, natural language processing (NLP) models are being deployed to extract structured insights from clinical notes, automate guideline summarisation, and support clinical decision making through contextualised textual analysis (Figure 4) (4). Together, these AI techniques are transforming clinical workflows, enabling predictive modelling, image-based automation, and real-time monitoring across the entire cardiovascular care continuum (Figure 5). We present this article in accordance with the Narrative Review reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-aw-595/rc).

Figure 3 AI techniques in cardiovascular research. AI, artificial intelligence; NLP, natural language processing.

Figure 4 Relationships between AI, ML, and DL. Commonly used ML and DL models are listed, respectively. AI, artificial intelligence; DL, deep learning; ML, machine learning.

Figure 5 Overview of machine learning and deep learning models.

Methods

Study design

The objective was to provide a clinically oriented synthesis of contemporary evidence on AI applications in cardiovascular medicine, focusing on translational relevance rather than systematic evidence aggregation. A structured but flexible literature exploration approach was adopted to identify key studies, landmark trials, and clinically influential publications.

Search strategy

Relevant literature was identified through iterative searches of PubMed/MEDLINE, Scopus, Embase and Web of Science covering publications from January 2015 to August 2025. The search strategy emphasised clinically impactful and methodologically robust studies rather than exhaustive inclusion. Search terms included combinations of controlled vocabulary and free-text keywords: “Artificial Intelligence”, “Machine Learning”, “Deep Learning”, “Cardiology”, “cardiovascular disease”, “Cardiac Imaging”, “Electrocardiography”, “Heart Failure”, “Atrial Fibrillation”, “Cardiac Rehabilitation”, “Telecardiology” and “Digital Cardiovascular Health”.

Additional targeted searches were performed for emerging domains including AI in cardiac imaging, wearable monitoring, risk prediction, and large language models (LLMs) in cardiology. Reference lists of key articles and recent high-impact reviews were also screened to identify seminal studies. The search strategy is described in Table 1.

Study selection approach

As a narrative review, formal inclusion/exclusion criteria were not applied in a systematic screening framework. Instead, studies were selected based on: clinical relevance to cardiovascular practice, methodological rigor, external validation or real-world applicability, influence on the field (landmark or widely cited work), and representation across major AI cardiovascular domains. Priority was given to peer-reviewed original studies, clinical trials, translational research, and authoritative guidelines. Conference abstracts, non-peer-reviewed sources, and studies lacking methodological transparency were not considered.

Data extraction and synthesis

Evidence was narratively synthesized and organized into thematic clinical domains: AI in ECG, AI in cardiovascular imaging, risk prediction and prevention, HR and structural cardiology, remote monitoring and digital cardiology, interventional cardiology and procedural guidance, implementation, governance, and clinical translation. Rather than quantitative pooling, emphasis was placed on: clinical utility, validation quality, generalizability and limitations and translational readiness (Figure 6).

Figure 6 Distribution of studies by theme. The bar chart illustrates the number of included studies categorised by major themes (Imaging, risk prediction, interventional cardiology, cardiac rehabilitation and ECG). ECG, electrocardiogram.

Methodological challenges

Reported limitations and clinical implications

Disagreements were resolved by consensus. Extracted data were narratively synthesized by thematic domain: ECG analysis, imaging interpretation, CVD prevention, cardiac rehabilitation, and remote monitoring.

External validation and generalizability

Many AI cardiovascular studies rely on single-center datasets with limited demographic diversity, raising concerns regarding generalizability across populations, healthcare systems, and device vendors. Multi-center validation remains limited, and local validation is recommended before clinical deployment.

Calibration and interpretability

Model calibration is often underreported despite its importance for clinical decision-making. Additionally, the “black box” nature of some DL models may limit interpretability, clinician trust, and regulatory transparency

Bias and equity

Underrepresentation of specific sex, ethnic, and socioeconomic groups in training datasets may introduce algorithmic bias and worsen health disparities. Subgroup performance reporting remains inconsistent across studies.

Emerging directions in AI cardiology

LLMs in cardiology

LLMs demonstrate potential in: summarizing complex clinical data, supporting guideline-based clinical decisions and enhancing clinician-patient communication.

However, challenges remain including hallucination risk, data privacy concerns, and the need for domain-specific clinical fine-tuning.

Advanced survival modelling

Modern time-to-event models such as DeepSurv and Random Survival Forests may outperform traditional Cox regression in selected cardiovascular datasets, particularly for censored outcomes. However, adoption requires rigorous handling of missing data and external validation.

Limitation

Generalizability & bias

Models often perform well on their training data but exhibit reduced accuracy in external settings. Diverse, well-curated training/validation is non-negotiable. A model developed at one centre may underperform at another (different devices, populations, prevalence). Always validate locally before wide use. If certain groups are under-represented in training data, performance can be uneven. Monitor metrics by subgroup.

Explainability & trust

Many high-performing models function as “black boxes”. Use interpretable outputs or add explainers when decisions affect care. Favor tools that show saliency/feature importance and come with plain-language model cards.

Prospective proof

Lots of retrospective wins; fewer prospective randomized controlled trials showing better outcomes/costs. Adoption should track evidence.

Data Infrastructure and integration

Standards, device drift, and EHR integration still trip people up. If inputs are messy, outputs will be too. Poor-quality inputs inevitably produce unreliable outputs. You’ll need clean interfaces, consistent labelling, and governance for model updates.

Regulatory & medico-legal

Who’s responsible for an AI-assisted miss? Keep a human-in-the-loop and log decisions. Check clearance (e.g., FDA/CE), intended use, and post-market monitoring plans.

Privacy

Favor federated learning, on device processing, and strict governance for cross-site models. Robust prospective evaluation and transparent reporting remain essential before widespread clinical adoption.

Applications of AI in cardiovascular medicine

AI in ECG interpretation

The ECG, a cornerstone of cardiovascular diagnosis for over a century, continues to evolve through the integration of AI. Traditional ECG interpretation is limited by interobserver variability and reduced sensitivity for detecting subclinical disease. However, DL, particularly CNNs, has dramatically expanded the diagnostic and prognostic capabilities of ECG analysis by revealing latent patterns that are imperceptible to the human eye (16,19). A landmark DL-enabled ECG algorithm developed by Attia et al. was able to identify left ventricular systolic dysfunction (LVEF ≤35%) from a normal sinus rhythm ECG with an area under the curve (AUC) exceeding 0.90, despite the absence of overt clinical signs (19). Subsequent advancements have demonstrated that DL models can also predict incident atrial fibrillation (AF) from routine sinus rhythm recordings before clinical onset (20) and detect silent myocardial ischemia through waveform recognition, outperforming traditional ST-segment criteria in multiple external cohorts (23). In addition to arrhythmias and ischemia, AI-ECG has shown the ability to diagnose structural heart diseases such as hypertrophic cardiomyopathy, amyloidosis, and pulmonary hypertension without the need for dedicated imaging. These developments have enabled population-scale cardiovascular screening, with AI-powered ECG models now integrated into smartphones, wearable devices, and cloud platforms, thereby democratising access to cardiovascular risk detection in both clinical and community settings (24-26). Despite these advances, model generalizability remains a critical challenge. Variability in ECG acquisition, population demographics, and comorbidities can degrade performance, underscoring the need for local validation, calibration, and transparent governance frameworks to ensure safe and equitable deployment (27).

AI in cardiovascular imaging

Cardiovascular imaging has been one of the earliest and most successful areas for integrating AI, due to the inherently high-dimensional and image-based nature of imaging data. AI models have consistently demonstrated high accuracy, efficiency, and reproducibility for a wide range of imaging tasks, including

cardiac chamber segmentation, ejection fraction estimation, tissue characterization, and coronary plaque quantification (10,28-30). In echocardiography, fully automated DL models now perform chamber and valvular assessments at an expert level. This drastically reduces both intra- and interobserver variability while boosting processing speed (31). Similarly, in cardiac CTA, AI enables automated plaque burden scoring, coronary calcium quantification, and non-invasive physiologic assessment using fractional flow reserve derived from CT (FFR-CT), reducing reliance on invasive pressure-wire assessment (32-34). AI-based cardiac MRI pipelines now support the precise quantification of ventricular volumes, strain, and tissue characterisation, providing reproducible results that are independent of operator input. Texture-based DL approaches have also demonstrated strong performance in detecting myocardial fibrosis and scar patterns, enhancing diagnostic and prognostic accuracy in cardiomyopathies (35). In intravascular imaging modalities such as intravascular ultrasound (IVUS) and optical coherence tomography (OCT), ML models distinguish between fibrotic, necrotic, and lipid-rich plaque subtypes with granular detail. This level of precision has critical implications for percutaneous coronary intervention (PCI) planning, enabling more accurate stent sizing, optimisation, and risk prediction (36-38).

As shown in Table 2, ML and DL models have been successfully applied across several cardiovascular imaging modalities, including CTA, MRI, echocardiography, and ECG. AI has demonstrated substantial potential in cardiovascular diagnostics by improving the accuracy, speed, and automation of imaging and electrocardiographic interpretation. DL models applied to large ECG datasets have shown strong capability for detecting latent disease states; for example, an algorithm trained on over 180,000 ECGs was able to predict AF during sinus rhythm with an AUC of 0.87, suggesting the potential for early identification of patients at risk for AF (6). Similarly, deep neural networks trained on more than 91,000 single-lead ECG recordings have achieved high performance in automated arrhythmia detection, with reported F1 scores of approximately 0.84, although these models were largely developed using ambulatory monitoring data (16).

In cardiac imaging, DL approaches have enabled automated quantification of cardiac structure and function. Video-based DL models applied to echocardiographic datasets exceeding 10,000 studies have demonstrated rapid and accurate estimation of LVEF and global longitudinal strain (GLS), with mean absolute errors of approximately 4.1% for EF estimation (7). CNNs have also been used to automate left ventricular segmentation and EF measurement from echocardiography, achieving high diagnostic accuracy with AUC values around 0.94 and sensitivities and specificities exceeding 90% when compared with expert readers, although performance may decline in studies with poor image quality or single-center training datasets (10). Additional work has shown that DL models can accurately classify standard echocardiographic views with accuracies exceeding 96%, facilitating automated workflow and quality control in echocardiography laboratories (39).

Advanced cardiac imaging modalities have also benefited from AI-based analysis. DL models applied to cardiac MRI have been used for automated myocardial scar detection, achieving sensitivities of approximately 94% and specificities near 88%, although external validation across diverse populations remains limited (19). Similarly, DL-based segmentation techniques have demonstrated high accuracy in automated left and right ventricular segmentation from cardiac MRI, with Dice similarity coefficients around 0.91 (20). In coronary imaging, ML models applied to coronary computed tomography angiography (CCTA) have been used to predict long-term cardiovascular outcomes, outperforming traditional risk scores in predicting 5-year major adverse cardiovascular events (MACE), although such models often require large high-quality imaging datasets and may lack interpretability (27). Hybrid DL and ML approaches have also been developed to automatically quantify coronary stenosis from angiography, demonstrating strong correlation with quantitative coronary angiography (QCA) measurements (r ≈ 0.93), though high-quality imaging remains necessary for optimal performance (30).

Beyond diagnostic interpretation, AI-assisted imaging platforms are increasingly being explored to guide interventional procedures, including transcatheter aortic valve implantation (TAVI), percutaneous coronary intervention (PCI), and device sizing, with the potential to reduce procedural time, radiation exposure, and operator variability; however, many of these applications remain conceptual and require outcome-based validation in clinical trials (23). Collectively, these developments highlight the rapidly expanding role of AI in cardiovascular diagnostics, offering opportunities to enhance clinical decision-making while emphasizing the need for broader validation, multi-center datasets, and integration into routine clinical workflows.

Role of AI in echocardiography

Echocardiography is one of the most widely used cardiovascular imaging modalities due to its safety, portability, and low cost. However, its reliance on operator expertise introduces variability in image acquisition, probe positioning, view selection, and manual border tracing, contributing to diagnostic inconsistency. AI, particularly DL, addresses these limitations across the entire echo workflow from acquisition to final interpretation (31). Contemporary DL models can autonomously recognise standard echocardiographic views such as the apical four-chamber and parasternal long-axis with accuracy exceeding 95%, offering real-time guidance and improving image standardisation across clinical settings, including handheld and point-of-care ultrasound (POCUS) devices (10,31). Automated quality control and view scoring help reduce nondiagnostic studies and improve consistency across operators and institutions (32). Once acquired, AI-powered segmentation tools facilitate reproducible quantification of cardiac chambers, automatically computing end-diastolic and end-systolic volumes, stroke volume, and ejection fraction (EF) with minimal interobserver variability (33). CNN-based architectures further enable the accurate measurement of wall thickness, valvular gradients, and Doppler-derived indices, which were previously dependent on manual tracing (10,31). Beyond volumetric analysis, AI has extended to strain imaging, particularly GLS, which is often more sensitive than EF in detecting early myocardial dysfunction. AI-based speckle tracking tools provide rapid, automated GLS calculations with consistent reproducibility, even in low-quality image settings (40). Similarly, ML models using Doppler indices (e.g., E/e', left atrial volume) are being employed to standardise diastolic function grading, a common source of clinical variability (40,41). More advanced applications include phenotyping and disease detection. Unsupervised clustering of echo-derived features has been used to identify phenotypes of HfpEF, each with distinct biomarker and outcome associations supporting precision cardiology (21,40). AI models have also demonstrated the capacity to discriminate cardiomyopathies (e.g., hypertrophic cardiomyopathy vs. athlete’s heart), detect cardiac amyloidosis, and flag studies with suspected severe ventricular dysfunction for urgent review (33,42). Recent bibliometric analyses also reveal a rapid rise in AI-echo research, with dominant themes including view classification, segmentation, strain analysis, and risk prediction. However, gaps remain, including limited multicenter validation, variable performance across ultrasound vendors, and concerns regarding black-box interpretability (43,44). Looking ahead, emerging strategies such as federated learning may address data privacy and diversity by enabling multicenter model training without data sharing (45).

Role of AI in CCTA and cardiac magnetic resonance imaging (CMR)

AI has transformed cardiovascular imaging by enabling automated, quantitative, and reproducible analysis across CTA and CMR, reducing reliance on time intensive manual interpretation. Applications include plaque quantification, tissue characterization, functional analysis, and prognostic modelling, delivering both diagnostic and workflow benefits (29-34). DL applied to cardiac MRI and ECG data has also improved identification of high-risk sudden cardiac death and patients most likely to benefit from implantable cardioverter defibrillators (ICD), with better precision than guideline-only criteria (46).

AI is moving coronary CT angiography (CCTA) beyond stenosis-centric assessment toward quantitative, biology informed profiling that integrates plaque morphology, distribution, and vessel geometry. A pivotal study derived and externally validated an AI-guided ischemia model that fused CCTA-derived atherosclerotic burden and vascular morphometrics to detect vessel-specific ischemia against invasive reference standards and to stratify prognosis, illustrating how machine-extracted plaque signatures can complement and sometimes outperform anatomy-only reads for clinical decision-making (47). By operationalising volumetric plaque metrics into a unified risk construct, this approach supports upstream selection for functional testing, intensification of preventive therapy, and longitudinal monitoring of disease biology rather than focusing solely on luminal narrowing. In parallel, AI-quantitative plaque analysis (AI-QCT) is beginning to resolve sex-specific risk gradients. A recent analysis showed that CCTA-derived features (e.g., total plaque volume, non-calcified plaque, and percent atheroma volume) may carry differential prognostic weight in women versus men, arguing against one size fits all thresholds. Collectively, recent literature suggests a shift toward precision prevention guided by quantitative coronary phenotype, provided deployment includes calibration checks, workflow integration, and equity monitoring so gains in sensitivity and throughput translate into patient-relevant outcomes across populations and scanner vendors (48). At the same time, explainable AI (XAI) techniques may enhance clinician trust and regulatory approval (49).

In CTA, AI algorithms facilitate automated plaque quantification, identification of high-risk morphological features (e.g., low attenuation, positive remodelling), and estimation of functional significance through CT-derived fractional flow reserve (CT-FFR). These applications enhance diagnostic precision, outperforming conventional CTA metrics for risk stratification and showing strong correlation with invasive QCA (27,30,32). AI-based tools, such as those by Popescu et al. (2020) and Sardar et al. (2019), achieved high diagnostic accuracy, reporting 89% accuracy for coronary artery disease detection and a correlation coefficient of r=0.93 between AI-derived and invasive stenosis quantification, respectively. Such automated approaches also streamline image interpretation and reduce observer variability, which is critical for large-scale clinical integration (30).

In CMR, DL-based segmentation models have achieved near-human accuracy, with Dice similarity coefficients of around 0.90, enabling fully automated ventricular contouring, volume computation, and tissue characterisation (33,35). AI applications extend to the detection of myocardial scar, fibrosis, and oedema, with models such as those developed by Lopez-Jimenez et al. (2020) achieving 94% sensitivity and 88% specificity for scar identification (19). Similarly, Westcott and Tcheng (2021) reported a Dice score of 0.91 for automated segmentation of the left (LV) and right ventricles (RV), demonstrating robustness in tissue and volumetric analysis (20).

Advanced ML models have also been applied for risk prediction and prognostic analytics derived from imaging. Motwani et al. (2017) used ensemble learning on 10,030 CTA scans to predict 5-year major adverse cardiac events (MACE), outperforming traditional risk scores (30). Beyond static image analysis, conceptual AI applications in interventional imaging guidance, including transcatheter valve implantation and PCI planning, were emphasised by Göçer and Durukan (2023), highlighting the evolution of AI from diagnostic to procedural domains (23).

Overall, AI-enabled cardiovascular CT and MRI represent among the most mature applications of medical AI, automating complex measurements, improving reproducibility, enhancing prognostic precision, and increasing workflow efficiency. However, many studies remain retrospective, single-centre, and vendor-dependent, limiting generalizability; prospective multicentre validation and standardised benchmarking across imaging devices remain essential for scalable clinical translation (43,44). An emerging direction is multimodal AI that fuses cine imaging, mapping, late gadolinium enhancement (LGE), and clinical data to differentiate overlapping phenotypes (e.g., hypertrophic cardiomyopathy versus amyloidosis subtypes), where accurate classification has major implications for therapy, surveillance, and family screening. Recent perspectives emphasise transparent cohort reporting, pre-specified calibration, and rigorous external testing before routine adoption, alongside site level recalibration, drift surveillance, and subgroup performance reporting to ensure fairness and clinical utility (49). A multicentre study across eight centres reported a two-stage system (cine-based screening followed by cine + LGE diagnosis) achieving near expert performance across multiple CVDs with external validation, highlighting how video-native architectures can improve generalizability and standardise reporting to address CMR scalability bottlenecks (49).

CMR provides highly detailed and reproducible quantification of cardiac function, volumes, and tissue characteristics, often serving as the gold standard for many parameters. AI is being used to streamline the cumbersome manual analysis steps associated with CMR. DL models, especially CNNs, have achieved near perfect accuracy in automated segmentation of the LV and RV, reducing a 20–30-minute manual task to seconds (42).

AI in intravascular imaging (IVUS and OCT) and robotics

Intravascular imaging modalities, including IVUS and OCT, provide high-resolution visualisation of coronary artery structure, facilitating precise characterisation of plaque morphology, vessel wall composition, and stent deployment. However, their interpretation traditionally requires expert level training and remains time-consuming. AI-driven analysis has transformed IVUS and OCT by enabling automated plaque segmentation, tissue classification, and quantitative feature extraction, thereby reducing the burden of manual review and improving reproducibility (37). ML algorithms have demonstrated high accuracy in differentiating between fibrous, lipid rich, and calcified plaques, providing clinically actionable insights for percutaneous coronary intervention (PCI) planning (27,37). In OCT, CNNs facilitate automated measurement of luminal area, stent expansion, malposition, and neointimal hyperplasia, supporting real time decision making; emerging platforms are integrating these capabilities for automated stent sizing, lesion length estimation, and vessel contour detection to reduce operator variability. While adoption remains early, hybrid imaging-AI systems linked to procedural guidance may enable more individualized revascularisation strategies (50). AI is increasingly reshaping interventional cardiology, with applications spanning lesion assessment, patient selection, procedural guidance, and robotic-assisted navigation. AI algorithms have enabled real-time intracoronary imaging interpretation, automated stent sizing, and computational modelling of coronary physiology during PCI (37). ML models have also shown promise in predicting post-intervention complications, such as no reflow, in-stent restenosis, and stent thrombosis, using angiographic, intravascular, and clinical variables (38,40). AI-powered robotic PCI systems, such as the CorPath GRX platform, incorporate motion stabilisation, force-feedback reduction, and remote capability, improving operator safety by reducing radiation exposure and ergonomic strain (41). These systems also provide a foundation for remote intervention, with the potential to enhance Cath lab access in underserved regions. AI-guided intravascular imaging platforms including IVUS and OCT have further improved procedural precision by automating stent sizing, vessel contour tracing, malapposition detection, and tissue characterisation (50).

As shown in Table 3, AI is increasingly being applied in interventional cardiology to support procedural decision-making, improve imaging interpretation, and predict procedural complications. Unsupervised ML clustering applied to cardiogenic shock registries has identified distinct clinical phenotypes that may assist in guiding percutaneous coronary intervention (PCI) decision-making in complex shock scenarios, although prospective validation is still required before routine clinical implementation (21). DL techniques have also demonstrated strong performance in intravascular imaging analysis; for example, automated plaque tissue characterization from IVUS images has achieved high segmentation accuracy with Dice scores around 0.91, though such models may remain dependent on specific imaging devices or platforms (35). Similarly, ML-based fractional flow reserve (FFR) estimation using angiographic images has shown promising diagnostic accuracy, with reported AUC values near 0.91 in retrospective datasets of approximately 300 angiograms, potentially enabling functional lesion assessment during PCI without invasive pressure wires (37). Predictive models have also been developed to anticipate procedural complications, such as the no-reflow phenomenon after PCI, with gradient boosting algorithms applied to large procedural datasets achieving AUC values around 0.89, although the absence of comprehensive biomarker data may limit predictive precision (38). In addition, ensemble ML models trained on datasets of up to 10,000 PCI patients have been proposed for predicting stent thrombosis risk, demonstrating high discrimination with AUC values approaching 0.92, though external validation remains lacking (39). Beyond decision support and risk prediction, AI has also been integrated with robotic-assisted PCI systems to improve operator ergonomics and reduce radiation exposure in the catheterization laboratory, although high costs and workflow integration challenges remain important barriers to adoption (41). Furthermore, OCT-based AI systems have been used to optimize stent deployment during PCI by detecting malapposition and other procedural issues, with studies reporting improved procedural optimization in cohorts of around 1,000 patients, although these findings are largely derived from single-center investigations (50). Together, these advances highlight the expanding role of AI in improving procedural precision, safety, and efficiency in interventional cardiology.

AI in congenital and structural heart disease

Risk prediction remains a central challenge in cardiovascular medicine, particularly in complex conditions such as heart failure (HF), cardiomyopathies, and post-percutaneous coronary intervention (PCI) outcomes. Traditional prognostic models such as the Framingham and TIMI scores, while historically valuable, are limited by their reliance on linear assumptions and fixed variable sets, which restrict their predictive performance in heterogeneous and real-world populations (43-45). Heart failure with reduced ejection fraction (HFrEF) poses a considerable clinical challenge, affecting an estimated 26 million people globally and leading to high morbidity and mortality, with approximately 50% of patients dying within five years of diagnosis (46). AI solutions in congenital cardiology are beginning to address challenges of anatomical variability, multi-modality imaging, and limited datasets. DL applied to fetal echocardiography has demonstrated effectiveness in detecting congenital heart defects (CHDs) with diagnostic accuracy comparable to that of expert readers (51). Feature-learning models have been developed to automate shunt quantification, chamber segmentation, and surgical planning for complex CHDs (42). Emerging multimodal systems integrate genetic, imaging, and hemodynamic data to support personalised management strategies for rare structural heart conditions (52).

AI in critical care cardiology

In cardiac intensive care, AI is being deployed for early detection of impending arrest, hemodynamic collapse, and multiorgan failure (53). ML models applied to real-time telemetry and lab feeds have demonstrated the capacity to improve the early recognition and triage of cardiogenic shock, thereby prompting the timely application of mechanical circulatory support (54). In a recent review, Zweck et al. highlighted unsupervised clustering as a promising strategy to phenotype cardiogenic shock patients, potentially guiding precision-based treatments and clinical trial stratification (22).on IVUS for plaque characterization (Dice score 0.91) (37), while Garcia et al. (2024) identified shock phenotypes via unsupervised clustering to support PCI decisions (21).

Research landscape and bibliometric trends in cardiovascular AI: bibliometric evolution of AI in cardiovascular imaging (2009–2024)

Bibliometric analyses reveal a rapid acceleration in AI-based cardiovascular imaging research since 2018, reflecting the field’s transition from experimental algorithms to clinical validation studies (11,23,43). Between 2009 and 2024, more than 3,400 peer-reviewed publications were identified in the domain of AI-assisted echocardiography alone, of which nearly 88% were original research contributions (23,43). China and the United States currently lead global output, with major contributions from academic centres such as the Mayo Clinic, Stanford University, and University College London (19,23). Cluster analyses using tools such as VOSviewer and CiteSpace have identified recurring themes, including deep learning, convolutional neural networks (CNNs), segmentation, speckle tracking, and view classification as dominant research foci (12,32,43). Emerging research patterns indicate sustained growth in areas such as quantification of left ventricular function, myocardial strain analysis, and AI-assisted assessment of valvular disease (27,32,33). Despite this strong momentum, notable limitations persist. Fewer than 15% of studies incorporate multicenter external validation, and vendor-specific imaging dependencies continue to restrict generalizability (30,35,43). Moreover, the underrepresentation of minority populations and developing regions underscores a pressing need to address algorithmic fairness and global equity in dataset curation (55-57).

AI in risk stratification and predictive analytics

AI-based models, particularly those leveraging ML and DL, address these limitations by processing high-dimensional, nonlinear data from EHR, genomic profiles, and imaging modalities. AI-driven risk scores often incorporate hundreds of variables, including unstructured clinical notes (via NLP) and complex image features (via radiomics), to generate more accurate and dynamic predictions. For instance, ML models have been developed to predict incident heart failure, predicting new-onset HF up to five years in advance using EHR data, significantly outperforming the Framingham risk score (AUC improvement of ~0.10) (58). AI-enhanced tele-rehabilitation programs have shown improvements in functional recovery and reduced readmissions, reinforcing their role in decentralised, patient-centered cardiovascular care (54,59). XAI is a burgeoning field that provides tools to help interpret model outputs, such as saliency maps (which highlight influential pixels on an image) or feature importance scores (which rank clinical variables) (60). Because most cardiovascular prevention and risk-stratification tools are predictive AI models, their clinical value depends not only on discrimination performance but also on real-world deployment, calibration, governance, and workflow integration. AI deployment in cardiology should follow a safety-first path that links model accuracy to measurable clinical impact under clear governance. Before activating any EHR/PACS system, teams should confirm prospective calibration and perform local recalibration to minimise dataset-shift problems (61,62). A practical rollout involves silent-mode testing with local adjudication, decision-curve and workflow analyses to demonstrate net clinical benefit, followed by a guarded go-live with human override, audit trails, and drift monitoring (61). Clinician-led, LLM-assisted builds that integrate existing, guideline-endorsed scores into a single step illustrate evidence-preserving integration, reducing workflow friction without altering the underlying science (63). Regulatory clearance is not equivalent to clinical utility; real-world evaluation and post-market surveillance remain essential, particularly for adaptive systems (64). Routine equity checks with performance reported by sex, age, ethnicity, and vendor help prevent unfair error patterns. Explainability can aid troubleshooting and trust, but does not replace rigorous validation and governance (17). Current society guidance emphasises oversight, pre-specified update policies, and multidisciplinary stewardship to scale AI safely across cath labs, echo services, and ambulatory care (65). High quality, annotated data are essential for reliable clinical AI, yet real-world cardiovascular datasets are fragmented, incomplete, and variably labeled, leading to biased or degraded performance when models are deployed outside their development settings (66).

Post PCI outcomes, predicting in-stent restenosis or future myocardial infarction by fusing intravascular imaging features with clinical and procedural data (67,68).

As shown in Table 4, AI has shown substantial potential in CVD prevention and risk stratification by improving prediction accuracy beyond traditional clinical risk models. Early work applying ML techniques such as random forests, support vector machines, and neural networks to population datasets like NHANES demonstrated modest but meaningful improvements in predictive performance for incident CVD, increasing the AUC by approximately 0.02–0.05 compared with traditional risk scores, although concerns about dataset representativeness and overfitting remain (11). Subsequent hybrid AI approaches applied to larger clinical datasets have been used to predict MACE, achieving predictive performance with AUC values around 0.85, though some studies were limited by relatively short follow-up durations (30). Similarly, ML models trained on large patient registries have been developed to predict heart failure hospitalization risk, demonstrating good discrimination with AUC values near 0.82; however, the absence of socio-economic variables may limit model generalizability (37). In large registry-based analyses including more than 750,000 patients with acute myocardial infarction (AMI), ensemble ML models such as XGBoost have shown strong predictive ability for post-AMI mortality, achieving AUC values of approximately 0.88 and improving net reclassification compared with conventional risk models, though registry-based datasets often require substantial missing data imputation (44). In addition, multimodal ML frameworks integrating imaging and clinical variables have been evaluated in population-based cohorts such as Multi-Ethnic Study of Atherosclerosis (MESA), demonstrating improved risk prediction compared with established tools like the Framingham risk score, albeit at the cost of increased complexity and resource requirements for multimodal screening (45). Together, these findings suggest that AI-driven risk prediction models may enhance personalized cardiovascular prevention strategies, although further validation in diverse populations and real-world settings is needed.

DL models, while powerful, often operate as “black boxes”, obscuring the rationale behind their predictions. This lack of transparency is a major barrier to clinical adoption and regulatory approval, as clinicians require certainty and interpretability before trusting a life-critical decision support tool (69). XAI is transforming the acceptance of AI by building clinician trust and improving diagnostic confidence, for instance, by showing which ECG leads or which areas of a CT scan contributed most to a specific diagnosis (Figure 7) (70).

Figure 7 Overview of explainable AI approaches. This conceptual diagram illustrates the workflow of AI models and the two primary categories of XAI methods. The process begins with Input Data, which feeds into either Transparent Models (e.g., Decision Trees) or Black-Box Models (e.g., Deep Learning). Both model types generate an Output (Prediction). This output then flows into a Built-in XAI Method. The XAI method’s approach is “Based on” either Model Specificity (Intrinsic or Model-Specific XAI) or Interpretation Types (model-agnostic or post-hoc method). AI, artificial intelligence; XAI, explainable AI.

Cardiovascular care is inherently multimodal, integrating ECG, lab results, imaging, genomics, and EHR data. Single-modality AI models miss the synergistic value of these diverse data streams. Multimodal AI platforms are designed to fuse and analyse this heterogeneous data. Early successes include integrating echocardiography, ECG, and clinical variables to predict HFpEF phenotypes with higher accuracy than any single-modality model (71,72). This approach holds the promise of constructing a comprehensive ‘digital twin’ for each patient, enabling highly accurate, dynamic risk assessment and therapy selection (60). Regulatory agencies are now developing frameworks for “Software as a Medical Device” (SaMD), including adaptive algorithms, post-market surveillance requirements, and real-world monitoring (73).

The ‘digital twin’ concept involves creating a virtual representation of a patient’s heart by integrating multi-scale data (from genetic markers to organ-level hemodynamics) within a computational model (74,75). AI, particularly physics-informed DL, is crucial for simulating disease progression and therapeutic response in this twin. For example, a digital twin can model the effect of a new anti-hypertensive drug on LV strain or predict the precise hemodynamic changes following valve replacement (57). Early applications have focused on simulating coronary flow and ventricular mechanics, moving beyond simple static prediction to a dynamic, ‘what-if’ scenario testing platform (76).

AI in wearables, remote monitoring and cardiac rehabilitation

eAI is transforming cardiac rehabilitation by addressing long-standing challenges related to adherence, patient engagement, and access to care, particularly among underserved populations. AI-enabled wearable technologies have demonstrated remarkable scalability, as shown in the Apple Heart Study by Perez et al., where smartwatch-based AF detection was validated in over 420,000 users, enabling continuous rhythm monitoring across diverse demographic groups (28). In the hospital setting, AI-driven command centre monitoring systems could predict clinical deterioration 3–6 hours before standard vital sign alerts, facilitating early intervention and supporting post-discharge monitoring for high-risk cardiovascular patients (54).

The proliferation of wearable technologies has ushered in a new era of cardiovascular care, enabling continuous physiological monitoring outside traditional healthcare settings. Wearable-based platforms continue to evolve, with AI-integrated devices capable of issuing early alerts for hemodynamic changes related to heart failure or hypertension for outpatient and home-based monitoring programs (16,26). In multiple studies, these systems have demonstrated superior predictive capabilities and responsiveness compared to standard follow-up strategies, though workflow integration and multicenter validation remain ongoing challenges (27). Modern devices equipped with photoplethysmography (PPG) and ECG sensors now collect real-time biosignals that can be analysed by ML algorithms to detect arrhythmias, estimate hemodynamic parameters, and assess cardiovascular fitness (28,42,58). Unsupervised learning techniques have also enabled refined phenotyping in conditions like aortic stenosis and HFpEF, revealing clinically meaningful subgroups with distinct prognoses and therapeutic responses (28). AI-driven platforms integrated into smartwatches and mobile applications have demonstrated clinically validated performance in detecting AF, estimating blood pressure, and monitoring oxygen saturation, supporting earlier diagnosis and risk-based preventive strategies (28,46,58). In one of the largest population-based studies, PPG-based algorithms achieved a positive predictive value (PPV) of over 80% in detecting irregular pulse notifications, ultimately leading to confirmed AF diagnosis via subsequent ECG recordings (28).

In the post-percutaneous coronary intervention (PCI) context, Westcott and Tcheng reported improved medication adherence and physical activity levels through the combined use of AI algorithms and wearable sensors integrated into remote monitoring platforms, leading to better recovery trajectories and reduced rehospitalisation (20). Similarly, Itchhaporia described the potential of AI-supported “virtual cardiac rehab” programs, where remote coaching, exercise tracking, and real-time engagement reinforcement offer meaningful alternatives to traditional centre-based rehabilitation for those unable to attend in person (10).

Wearable-integrated remote patient monitoring systems now support longitudinal tracking of vital signs, physical activity, and cardiac rehabilitation adherence, enabling timely identification of heart failure decompensation and post-myocardial infarction or PCI complications (22,53,55).

Systematic reviews of tele-rehabilitation models confirm that AI-enabled cardiac rehabilitation can improve functional outcomes, reduce readmissions, and extend access to medically supervised exercise programs, particularly in resource-limited and elderly populations (58,59). These technologies enable personalized exercise prescriptions adjusted by ML algorithms, predictive modelling of dropout risk using EHR and wearable data, early detection of arrhythmias and heart failure decompensation through continuous monitoring, and automated reminders and behavioral nudges to enhance long-term adherence. Despite these advances, widespread adoption remains limited by challenges including the scarcity of multicenter randomized controlled trials, concerns about generalizability across different healthcare systems, and variable technology acceptance among older adults (Figure 8) (60).

Figure 8 Workflow of AI integration in cardiovascular care. Conceptual workflow from data sources through AI methods and applications to clinical outputs and outcomes. AI, artificial intelligence; ECG, electrocardiogram; EHR, electronic health record; NLP, natural language processing.

Future research must prioritise equitable implementation strategies, model calibration across diverse populations, and accessible user interfaces to realise the potential of AI-enabled cardiac rehabilitation fully.

As shown in Table 5, AI is increasingly being integrated into cardiac rehabilitation and remote monitoring systems to enhance patient surveillance, adherence, and early detection of complications. ML algorithms applied to wearable ECG devices have demonstrated strong performance in identifying arrhythmias during cardiac rehabilitation, achieving sensitivities of approximately 93% and specificities of 90% in cohorts of around 500 patients, although performance may vary depending on device-specific algorithms (19). Remote monitoring platforms that combine wearable sensors with cloud-based analytics have also been used to track activity levels and vital signs in patients after percutaneous coronary intervention (PCI), with early studies suggesting improved adherence to rehabilitation programs, though randomized controlled trial evidence remains limited (20). In addition, DL-based motion analysis has been explored for assessing gait and functional recovery during rehabilitation, showing promising accuracy of about 88% in small single-center cohorts (23). Large-scale remote rhythm monitoring using PPG signals from consumer smartwatches has further demonstrated the feasibility of AI-driven AF screening, with studies involving over 400,000 participants reporting a PPV of approximately 84% for irregular pulse notifications, although confirmatory ECG remains necessary and participant populations may be self-selected (28). Together, these developments highlight the growing role of AI-enabled wearable technologies in extending cardiac rehabilitation beyond traditional clinical settings, while also underscoring the need for validation in larger, controlled studies.

Translational implementation, governance, methodological challenges, and roadmap for the future in AI-enabled cardiovascular medicine

AI is reshaping cardiovascular medicine, but its translation into clinical practice faces substantial challenges related to data quality, bias, validation, interpretability, regulation, and equity.

Data quality, heterogeneity, and bias

AI models depend on large, well-annotated datasets; however, real-world cardiovascular data are often fragmented, incomplete, and inconsistently labeled, resulting in biased or poorly generalizable models (72,77). Population-level disparities in age, sex, and ethnicity can distort predictions, especially in heart failure and risk stratification models (55). A systematic review revealed up to 30% performance variation across sex and race groups for coronary artery disease prediction models (56), while some AI-ECG tools overestimated abnormalities in younger patients and underestimated risk in older or diverse populations (58).

Lack of model transparency and explainability

High-performing DL models often act as “black boxes”, limiting clinician trust and regulatory acceptance (46,67). XAI techniques such as saliency maps, feature attribution, and rule-based surrogates have been developed to visualise decision pathways and enhance interpretability, but few have been validated in real-world cardiovascular settings (68).

Insufficient external validation and reproducibility

Despite the proliferation of AI models, fewer than 10% undergo external validation, and less than 1% are tested in prospective trials (43). Most high-accuracy AI-ECG models are trained and validated in a single institution, raising overfitting concerns (43). The lack of dedicated registries and audit systems hampers regulatory approval and real-world integration.

Workflow integration and human-Machine interaction

AI is increasingly deployed to optimise workflows, automate reporting, and reduce administrative load (45,60). Systems that prioritise cardiovascular imaging and automate documentation have reduced diagnosis times and administrative delays (60,77-79). However, interoperability, clinician trust, and human-centred design remain essential to prevent workflow disruption or alert fatigue (45,80).

Calibration, interpretability, and equity in AI for cardiovascular medicine

Model calibration ensures predicted risks align with observed outcomes, yet it is inconsistently reported (72). Poor calibration can mislead clinical decisions. DL “black-box” models limit interpretability, though XAI methods such as saliency maps and feature-attribution models aim to improve transparency (6). Equity remains critical: underrepresentation of women, older adults, and minorities reduces AI reliability and may worsen disparities (71,72). Without rigorous attention to these factors, calibration, interpretability, and equity failures risk exacerbating harm in cardiovascular medicine (2,13,29).

Regulatory, legal, and ethical barriers

Regulatory systems for cardiovascular AI remain underdeveloped. Most tools do not yet meet FDA or European Medicines Agency (EMA) requirements for post-market monitoring or lifecycle management (81,82). Ethical concerns include data privacy and data ownership, informed consent for EHR use, and accountability for algorithmic errors (83). Bias against underrepresented groups remains a key issue (72). Generative AI and large language models (LLMs) add complexity, given unclear data lineage and dynamic outputs (84). Emerging safeguards include human in the loop frameworks, federated learning, XAI, and standardised validation protocols. Bias in training datasets can lead to disparities in care, especially among underrepresented populations. Data governance represents a central challenge for clinical AI deployment. Cardiovascular datasets are often fragmented across institutions, variably labeled, and subject to differing regulatory protections. Compliance with international data protection standards, including the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA), is critical when handling patient-level data. Issues surrounding data ownership, cross-institutional data sharing, and secondary use of EHR data remain unresolved in many healthcare systems and require clear governance structures (43,45).

Economic and global health considerations

AI innovation demands high financial and computational resources, limiting access in low- and middle-income countries (LMICs) and deepening the “AI care divide” (70). Current models are largely developed in high-income regions using non-representative data. Federated learning (FL) and cloud-based cardiovascular AI platforms may help promote equitable global deployment (76,77).

Future directions in cardiovascular AI

The next phase of cardiovascular AI is transitioning from single-task tools to adaptive, multimodal systems integrating imaging, ECG, EHR, and omics data. FL enables multi-institutional collaboration without centralised data sharing, improving generalisability and privacy (45,49). XAI remains essential to enhance trust and regulatory acceptance (41). Multimodal models combining imaging, genomics, and clinical data improve early detection and precision therapy, outperforming unimodal predictors (40). Future clinical trials should evaluate AI based on clinical utility and patient outcomes rather than accuracy metrics alone (82).

Emerging applications in multi-omics, generative AI, and continuous learning systems promise to advance precision cardiology (3,77,82). Generative models can simulate cardiovascular anatomy and pathology, aiding training and procedural planning (83,84). Continuous learning AI systems allow adaptive updates with new data and clinician feedback, maintaining accuracy over time (76,77).

The evolving paradigm of clinician AI partnership envisions AI as a decision co-pilot enhancing, not replacing, human expertise through hybrid intelligence (78-80). Regulatory innovation, such as the FDA’s Predetermined Change Control Plan (PCCP), will allow post-approval AI model evolution under strict safety criteria (61), with upcoming mandates for transparency, auditability, and fairness in cardiovascular AI approval processes within the next 5–7 years (62).

Conclusions

AI is rapidly transforming cardiovascular medicine, with growing applications in diagnostics, prognostics, and therapeutic planning. AI-enabled tools such as multimodal models and federated learning systems are driving a shift toward more precise, automated, and personalised care. Rather than replacing clinicians, the future of cardiovascular medicine lies in AI complementing human expertise, enhancing decision-making while expanding access to high-quality care. However, widespread adoption depends on overcoming challenges in generalizability, transparency, and algorithmic bias. Strategies like federated learning, XAI, standardised reporting, and outcome-focused clinical trials will be essential to ensure safe, equitable, and clinically meaningful integration. Ultimately, AI should serve as an intelligent collaborator, advancing precision cardiovascular care while upholding safety, fairness, and clinical trust.

Unlike prior broad conceptual reviews, including the JACC State of the Art review by Khera et al., this narrative review focuses specifically on clinical translation, real-world validation, implementation challenges, and workflow integration of AI tools in cardiovascular practice (44).

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-aw-595/rc

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-aw-595/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-aw-595/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Roth GA, Mensah GA, Johnson CO, et al. Global Burden of Cardiovascular Diseases and Risk Factors, 1990-2019: Update From the GBD 2019 Study. J Am Coll Cardiol 2020;76:2982-3021. [Crossref] [PubMed]
2. Virani SS, Alonso A, Aparicio HJ, et al. Heart Disease and Stroke Statistics-2021 Update: A Report From the American Heart Association. Circulation 2021;143:e254-e743. [Crossref] [PubMed]
3. Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med 2019;25:44-56. [Crossref] [PubMed]
4. Johnson KW, Torres Soto J, Glicksberg BS, et al. Artificial Intelligence in Cardiology. J Am Coll Cardiol 2018;71:2668-79. [Crossref] [PubMed]
5. Shameer K, Johnson KW, Glicksberg BS, et al. Machine learning in cardiovascular medicine: are we there yet? Heart 2018;104:1156-64. [Crossref] [PubMed]
6. Attia ZI, Noseworthy PA, Lopez-Jimenez F, et al. An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction. Lancet 2019;394:861-7. [Crossref] [PubMed]
7. Ouyang D, He B, Ghorbani A, et al. Video-based AI for beat-to-beat assessment of cardiac function. Nature 2020;580:252-6. [Crossref] [PubMed]
8. Kwon JM, Lee SY, Jeon KH, et al. Deep Learning-Based Algorithm for Detecting Aortic Stenosis Using Electrocardiography. J Am Heart Assoc 2020;9:e014717. [Crossref] [PubMed]
9. Zweck E, Li S, Burkhoff D, et al. Profiling of Cardiogenic Shock: Incorporating Machine Learning Into Bedside Management. J Soc Cardiovasc Angiogr Interv 2025;4:102047. [Crossref] [PubMed]
10. Zhang J, Gajjala S, Agrawal P, et al. Fully Automated Echocardiogram Interpretation in Clinical Practice. Circulation 2018;138:1623-35. [Crossref] [PubMed]
11. Krittanawong C, Zhang H, Wang Z, et al. Artificial Intelligence in Precision Cardiovascular Medicine. J Am Coll Cardiol 2017;69:2657-64. [Crossref] [PubMed]
12. Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in medical image analysis. Med Image Anal 2017;42:60-88. [Crossref] [PubMed]
13. Liu X, Glocker B, McCradden MD, et al. Ethical issues in artificial intelligence for health: Challenges and opportunities for research. Health Data Sci 2021;2021:614292.
14. Lee CS, Lee AY. Clinical applications of continual learning machine learning systems in healthcare. JAMA Netw Open 2020;3:e203033.
15. U.S. Food and Drug Administration. Proposed regulatory framework for modifications to AI/ML-based software as a medical device. 2021. Available online:
16. Hannun AY, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection using a deep neural network. Nat Med 2019;25:65-9. [Crossref] [PubMed]
17. Ribeiro MT, Singh S, Guestrin C. Why should I trust you? Explaining the predictions of any classifier. Proc ACM SIGKDD Int Conf Knowl Discov Data Min 2016:1135-44.
18. Jentzer JC, van Diepen S, Barsness GW, et al. Machine-learning phenotypes of cardiogenic shock. J Am Coll Cardiol 2020;76:2319-31.
19. Lopez-Jimenez F, Attia Z, Kapa S, et al. Artificial intelligence in cardiology: Present and future. J Am Coll Cardiol 2020;75:1609-25. [PubMed]
20. Westcott T, Tcheng JE. Artificial intelligence and machine learning in cardiology: Let the fun begin. J Soc Cardiovasc Angiogr Interv 2024;3:100112.
21. Garcia R, Li X, Chen Y, et al. Clustering-based shock phenotype analysis for precision cardiology. Crit Care Med 2024;52:78-89.
22. Romiti GF, Pastori D, Rivera-Caravaca JM, et al. Artificial intelligence and cardiovascular diseases: An unexpected alliance. Cardiol Res Pract 2020;2020:6132487. [Crossref] [PubMed]
23. Göçer H, Durukan AB. The use of artificial intelligence in interventional cardiology. Turk Gogus Kalp Damar Cerrahisi Derg 2023;31:420-1. [Crossref] [PubMed]
24. Armoundas AA, Adams J, Arboix A, et al. Artificial intelligence in improving outcomes in heart disease. Circ Cardiovasc Qual Outcomes 2024;17:e008233. [PubMed]
25. Skalidis I, Cagnina A, Luangphiphat W, et al. ChatGPT takes on the European exam in core cardiology. Eur Heart J Digit Health 2023;4:279-81. [Crossref] [PubMed]
26. Avram R, Olgin JE, Kuhar P, et al. A digital biomarker of diabetes from smartphone-based vascular signals. Nat Med 2020;26:1576-82. [Crossref] [PubMed]
27. Motwani M, Dey D, Berman DS, et al. Machine learning for prediction of all-cause mortality in patients with suspected coronary artery disease: a 5-year multicentre prospective registry analysis. Eur Heart J 2017;38:500-7. [PubMed]
28. Perez MV, Mahaffey KW, Hedlin H, et al. Large-Scale Assessment of a Smartwatch to Identify Atrial Fibrillation. N Engl J Med 2019;381:1909-17. [Crossref] [PubMed]
29. Narula S, Shameer K, Salem Omar AM, et al. Machine-learning algorithms to classify calcium in intravascular OCT images. JACC Cardiovasc Imaging 2016;9:1241-9.
30. Sardar P, Abbott JD, Kundu A, et al. Impact of Artificial Intelligence on Interventional Cardiology: From Decision-Making Aid to Advanced Interventional Procedure Assistance. JACC Cardiovasc Interv 2019;12:1293-303. [Crossref] [PubMed]
31. van Assen M, Boerman LM, Zhuang T, et al. Feature tracking echocardiography using deep neural networks. J Am Soc Echocardiogr 2020;33:809-19.
32. Al’Aref SJ, Maliakal G, Singh G, et al. Machine learning of clinical variables and coronary artery calcium scoring. JACC Cardiovasc Imaging 2020;13:1-10. [PubMed]
33. Bloomfield GS, Baldridge A, Agarwal A, et al. Global trends in cardiovascular disease. J Am Coll Cardiol 2022;80:543-53.
34. Stone GW, Maehara A, Lansky AJ, et al. Natural-history study of coronary atherosclerosis. N Engl J Med 2011;364:226-35. [Crossref] [PubMed]
35. Backhaus SJ, Lange T, Hellenkamp K, et al. Deep learning plaque characterization in intravascular ultrasound. Eur Heart J 2023;44:1047-56.
36. Liu Y, Chen PC, Krause J, et al. How to read articles that use machine learning. JAMA 2019;322:1806-16. [Crossref] [PubMed]
37. Seckanic M, Risteski M, Widimsky P, et al. AI-driven fractional flow reserve during coronary angiography. Front Cardiovasc Med 2023;10:1023452.
38. Xu B, Maddali S, Wang L, et al. Predicting no-reflow after PCI using machine learning. Catheter Cardiovasc Interv 2021;98:260-8.
39. Madani A, Arnaout R, Mofrad M, et al. Deep learning for echocardiographic view classification. NPJ Digit Med 2018;1:6. [Crossref] [PubMed]
40. Siontis GCM, Buechel RR, Räber L, et al. Machine learning predicting stent thrombosis after PCI. Eur Heart J 2022;43:1124-33.
41. Mahmud E, Schmid F, Kalmar P, et al. Robotic-assisted PCI. J Soc Cardiovasc Angiogr Interv 2023;3:100184. [PubMed]
42. Muffoletto M, Xu H, Burns R, et al. Evaluation of deep learning estimation of whole heart anatomy from automated cardiovascular magnetic resonance short- and long-axis analyses in UK Biobank. Eur Heart J Cardiovasc Imaging 2024;25:1374-83. [Crossref] [PubMed]
43. Mortazavi BJ, Ko YA, Bucholz EM, et al. Machine learning in heart failure: Beyond predictive modelling. J Card Fail 2021;27:558-69.
44. Khera R, Haimovich J, Hurley NC, et al. Use of machine learning models to predict death after acute myocardial infarction. JAMA Cardiol. 2021;6:633-41. [Crossref] [PubMed]
45. Ambale-Venkatesh B, Yang X, Wu CO, et al. Deep learning and ECG for sudden cardiac death prediction. J Am Heart Assoc 2022;11:e025678.
46. Kwon JM, Kim KH, Jeon KH, et al. Clinical implementation of artificial intelligence-based ECG analysis for heart failure. Eur Heart J 2021;42:4445-54.
47. Nurmohamed NS, Danad I, Jukema RA, et al. Development and validation of a quantitative coronary CT angiography model for diagnosis of vessel-specific coronary ischemia. JACC Cardiovasc Imaging 2024;17:894-906. [Crossref] [PubMed]
48. Miller RJH, Manral N, Lin A, et al. Patient-specific myocardial infarction risk thresholds from AI-enabled coronary plaque analysis. Circ Cardiovasc Imaging 2024;17:e016958. [Crossref] [PubMed]
49. Wang YR, Yang K, Wen Y, et al. Screening and diagnosis of cardiovascular disease using artificial intelligence-enabled cardiac magnetic resonance imaging. Nat Med 2024;30:1471-80. [Crossref] [PubMed]
50. Chamié D, Nakamura M, Mattesini A, et al. Optical coherence tomography-based artificial intelligence for PCI optimisation. Circ Cardiovasc Interv 2022;15:e011685.
51. Sun H, Zhang X, Li Y, et al. Artificial intelligence-enabled prediction of refractory cardiogenic shock: Insights from multicenter registries. J Soc Cardiovasc Angiogr Interv 2025;4:102119.
52. Huang L, Zhao W, Li Y, et al. A multimodal deep learning model for structural congenital heart disease support. Eur Heart J 2023;44:1351-63.
53. Joshi M, Ashrafian H, Aufegger L, et al. Wearable sensors to improve detection of patient deterioration. Expert Rev Med Devices 2019;16:145-54. [Crossref] [PubMed]
54. Yuan S, Yang Z, Li J, et al. AI-Powered early warning systems for clinical deterioration significantly improve patient outcomes: a meta-analysis. BMC Med Inform Decis Mak 2025;25:203. [Crossref] [PubMed]
55. Cirillo D, Catuara-Solarz S, Morey C, et al. Sex and gender differences and biases in artificial intelligence for biomedicine and healthcare. NPJ Digit Med 2020;3:81. [Crossref] [PubMed]
56. Nestor B, Martinez-Mateo V, Lopez-Lopez JA, et al. Bias in artificial intelligence models used in cardiology: A systematic review. J Am Coll Cardiol 2024;83:884-97.
57. Wahl B, Cossy-Gantner A, Germann S, et al. Digital health and artificial intelligence in low- and middle-income countries: Scaling up equitable innovations. Lancet Glob Health 2023;11:e33-45.
58. Galloway CD, Valys AV, Shreibati JB, et al. ECG-based detection of cardiac abnormalities using deep learning across age and racial groups. JAMA Cardiol 2023;8:410-8.
59. Chen IY, Szolovits P, Ghassemi M. Can artificial intelligence help reduce disparities in general medical and mental health care? AMA J Ethics 2019;21:167-79. [Crossref]
60. Benjamens S, Dhunnoo P, Meskó B. The state of artificial intelligence-based FDA-approved medical devices and algorithms. Nat Med 2020;26:1212-5. [PubMed]
61. Kelly CJ, Karthikesalingam A, Suleyman M, et al. Key challenges for delivering clinical impact with artificial intelligence. BMC Med 2019;17:195. [Crossref] [PubMed]
62. Obermeyer Z, Powers B, Vogeli C, et al. Dissecting racial bias in an algorithm used to manage the health of populations. Science 2019;366:447-53. [Crossref] [PubMed]
63. Egaimi M, Corvo P, Al Houri H, et al. Development and Technical Validation of an Integrated Risk Calculator for Acute Coronary Syndrome Using ChatGPT-Assisted Coding. Cureus 2025;17:e83410. [Crossref] [PubMed]
64. Muehlematter UJ, Daniore P, Vokinger KN. Approval of artificial intelligence and machine learning-based medical devices in the USA and Europe (2015-20): a comparative analysis. Lancet Digit Health 2021;3:e195-203. [Crossref] [PubMed]
65. Lata K, Rooney C, Tukaye D, et al. Governing the unbound: SCAI’s role in the future of artificial intelligence. J Soc Cardiovasc Angiogr Interv 2025;4:102517. [Crossref] [PubMed]
66. Dou Q, So TY, Jiang M, et al. Federated deep learning for cardiac imaging diagnostics. IEEE Trans Med Imaging 2022;41:2973-85.
67. Mittelstadt BD, Floridi L. The ethics of artificial intelligence in healthcare: A mapping review. Big Data Soc 2022;9:1-21.
68. Selvaraju RR, Cogswell M, Das A, et al. Grad-CAM: Visual explanations from deep networks via gradient-based localisation. Int J Comput Vis 2020;128:336-59. [Crossref]
69. Wiens J, Saria S, Sendak M, et al. Do no harm: a roadmap for responsible machine learning for health care. Nat Med 2019;25:1337-40. [Crossref] [PubMed]
70. van de Velde S, Heselmans A, Delvaux N, et al. Clinical decision support systems in heart failure: A review of human-machine interaction. ESC Heart Fail 2021;8:4461-72.
71. Rajkomar A, Hardt M, Howell MD, et al. Machine learning in medicine: Addressing physician burnout. Lancet Digit Health 2022;4:e137-9.
72. Desai W, Shah PB, Yeh RW, et al. Decision support tools in PCI: Clinician acceptance and AI feedback loops. JACC Cardiovasc Interv 2023;16:1191-202.
73. U.S. Food and Drug Administration. Proposed regulatory framework for AI/ML-based software as a medical device. FDA 2024. Available online:
74. Cohen IG, Amarasingham R, Shah A, et al. The lawyer’s guide to the ethics of AI in medicine. J Law Med Ethics 2023;51:23-34.
75. Ghassemi M, Oakden-Rayner L, Beam AL, et al. The false hope of current approaches to explainable AI in health care. Lancet Digit Health 2021;3:e745-50. [Crossref] [PubMed]
76. Esteva A, Chou K, Yeung S, et al. A guide to continuous learning for deployed medical AI systems. Nat Med 2021;27:1056-62.
77. Pavlovski A, Ding WY, Gupta D, et al. Continuous learning ECG platforms in emergency medicine. Emerg Med J 2024;
78. Collins GS, Moons KGM. Hybrid intelligence in healthcare: The future of decision-making. BMJ 2023;382:e074198.
79. Westaby J, Salinas P, Cook CM, et al. Cardiologist–AI collaboration in catheter laboratory imaging: A prospective trial. JACC Cardiovasc Interv 2024;17:387-97.
80. Lee C, Ouyang D, He B, et al. Hybrid intelligence in echocardiography: A randomized workflow study. Eur Heart J 2023;44:1040-6.
81. Bae S, Park H, Lee J, et al. Generative models for three-dimensional cardiac anatomy simulation. IEEE Trans Vis Comput Graph 2023;29:403-13.
82. McKinney SM, Sieniek M, Godbole V, et al. International evaluation of an artificial intelligence system for breast cancer screening. Nature 2020;577:89-94. [Crossref] [PubMed]
83. Elia N, Toma M, Siontis GCM, et al. Algorithmic fairness in cardiovascular artificial intelligence: A framework for clinical deployment. J Am Coll Cardiol 2024;83:1285-98.
84. Balu N, Narayan SM, Kahn AM, et al. Physics-informed generative modelling of coronary flow. Circulation 2024;149:671-84.