Artificial intelligence across the cardiovascular diagnostic pathway: a case-based narrative review

Introduction

Background

Artificial intelligence (AI) is reshaping cardiovascular medicine on a global scale, with applications emerging across the entire continuum of care. Rapid advances in machine learning methods and computational power have accelerated the transition of AI in cardiology from experimental prototypes to important clinical tools. These systems can support routine cardiovascular assessment—including structured history-taking, digital auscultation, electrocardiographic (ECG) interpretation, and cardiac imaging—with the potential to improve diagnostic workflows and patient care. As cardiovascular disease remains the leading cause of morbidity and mortality worldwide, timely recognition and early diagnosis remain major public health priorities (1).

Rationale and knowledge gap

Studies demonstrate that AI integration offers significant gains in diagnostic accuracy and workflow efficiency with several systems approaching or surpassing expert-level performance in select applications (2-6). These innovations hold particular promise in addressing rising challenges such as clinician burnout, increasing public health demands and limited access to subspecialist care (7). As cardiovascular disease continues to represent the leading cause of morbidity and mortality worldwide, timely recognition remains an ongoing public health priority (1). AI-enabled tools have been specifically designed to facilitate early diagnosis at the primary care level, with some extending their utility to home-based screening and remote patient assessments (8,9). These developments are poised to expand access to essential diagnostic services in low-resource settings by supporting scalable, cost-efficient, and reliable screening approaches.

While AI applications have to date focused on individual diagnostic modalities, recent advances in multimodal AI now enable the integration of diverse data sources (e.g., ECGs, imaging, laboratory results, and wearable data), with fusion of predictive capability, to enable more comprehensive and precise diagnostic assessment and prediction of clinical outcomes (10-12). This evolution underscores the need to evaluate AI not as isolated tools, but as components of a continuous diagnostic continuum. Several narrative reviews have summarized AI applications in cardiovascular disease, typically organized by disease phenotype or broad clinical functions. This narrative review synthesizes evidence along the cardiovascular diagnostic pathway—from history-taking and bedside examination through electrocardiography, echocardiography, and advanced imaging—using a case-based framework to illustrate workflow integration, diagnostic performance, and efficiency gains when these data sources are combined (Figure 1). We also discuss important considerations for broad implementation, including regulatory approvals, post-market oversight, governance challenges, and a summary of U.S. Food and Drug Administration (FDA)-authorized cardiovascular AI tools. We conclude with practical barriers and research priorities to enable safe and equitable clinical adoption.

Figure 1 AI-enabled multimodal cardiovascular diagnostic workflow. AI, artificial intelligence.

Objective

This narrative review examines the evidence and real-world use of AI-based diagnostics across major cardiovascular diagnostic domains. We summarize diagnostic performance and workflow impact, review regulatory pathways and current challenges, and outline practical barriers and research priorities to support safe clinical adoption. 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-550/rc).

Methods

To explore AI across the cardiovascular diagnostic pathway, data from clinical trials, retrospective and prospective observational studies and external validation cohorts were retrieved from MEDLINE/PubMed (U.S. National Library of Medicine, Bethesda, MD, USA), Embase (Elsevier, Amsterdam, Netherlands), and Google Scholar (Google LLC, Mountain View, CA, USA). Systematic reviews/meta-analyses were screened to contextualize findings and identify additional primary studies. Searches used combinations of the terms “artificial intelligence”, “machine learning”, “deep learning”, and “large language model”, together with modality-specific terms including “history-taking”, “symptom checker”, “clinical documentation”, “digital stethoscope”, “auscultation”, “phonocardiography”, “electrocardiography/ECG”, “echocardiography”, “point-of-care ultrasound”, “coronary CT angiography/CCTA”, “coronary artery calcium”, “cardiac MRI/CMR”, “SPECT”, “PET”, “angiography”, “plaque quantification”, “FFR”, and “multimodal fusion”, for publications from January 1, 2000 to July 10, 2025. Regulatory resources (e.g., FDA device databases and publicly available clearance summaries) were also reviewed to identify cardiovascular AI software that has received FDA approval. Additional articles were identified by screening reference lists of selected publications. Pertinent studies were critically evaluated by the authors, and inclusion decisions were made by consensus; non-English publications and reports deemed to have insufficient methodological rigor or low credibility were excluded (Table 1).

AI-enhanced history-taking

A 63-year-old man with a new diagnosis of lymphoma is referred to the cardiology clinic for evaluation before starting an anthracycline-based chemotherapy regimen. Upon arrival, he was greeted by an AI-powered assistant accessed via tablet in the examination room. The system elicited clinical information through a structured, interactive series of questions, which identified the presence of orthopnea and rapid weight gain. The algorithm then generated a preliminary note featuring the differential diagnosis, with the primary differential diagnosis as new-onset heart failure. It recommended initial labs, natriuretic peptide testing, and echocardiography based on updated clinical guidelines. The note is incorporated into the electronic health record (EHR) before the patient-physician interaction, allowing the clinician to prioritize clarification of essential clinical details. discussion of the proposed management strategy, with subsequent update of the impression and plan that was AI generated.

Despite significant advances in modern diagnostic modalities, history-taking remains the cornerstone of clinical assessment, contributing approximately 80% to a clinical diagnosis (13). However, the ability to conduct a timely and thorough history using traditional methods is often hindered by the constraints of modern clinical workflows and increasing demands on the clinician. Emerging AI applications are being leveraged to support history-taking by systematically collecting data on patient symptoms and risk factors, either prior or during the patient’s encounter. By improving operational efficiency and the accuracy of collected data, these tools may allow clinicians to dedicate more time to clinical decision-making and meaningful patient engagement.

Early applications of AI in patient assessment included the use of text-based digital symptom checkers, which gather written information on symptoms and medical history to assist individuals in identifying the suitable level of medical care. The United Kingdom’s National Health Service (NHS) introduced the NHS 111 triage service in 2017, an AI based tool aimed to alleviate the increasing burden on primary care providers and emergency departments. Babylon’s AI platform which underpins NHS 111, demonstrated similar triage performance to consultant physicians, with safety and appropriateness rates of 97.0% and 90.0%, respectively, compared to 93.1% and 90.5% for doctors in a prospective validation study (8). Another study evaluating the clinical utility of ChatGPT-4 reported an overall guidance accuracy of 92.63% (95% CI: 90.34–94.93), with internal medicine achieving the highest accuracy (93.51%) and general surgery the lowest (91.46%) (9). Limitations of both of these studies include the use of scenario-based cases rather than real patient encounters and lack of validation in cardiovascular specific populations. An increasing body of evidence highlights the utility of these tools in enhancing healthcare efficiency, primarily by curbing unnecessary emergency department visits and directing patients promptly to suitable care settings.

Recent advancements have introduced AI systems that perform clinical interviews prior to physician evaluation, automatically generating structured, symptom-based documentation in real time for clinician review. In a single-center, cardiovascular clinic cohort of 762 patients, the Automated Assessment of Cardiovascular Examination (AACE) platform produced computer-generated histories that were rated significantly higher in quality than those documented by physicians (mean score: 4.2 vs. 2.6; P<0.001). Additionally, when physicians were provided with access to the AI-generated history, consultation times were reduced, and clinicians ordered fewer diagnostic tests and new medications (14). However, early iterations were limited by a lack of conversational fluency and empathy, both of which are essential for patient engagement and satisfaction. Front facing interactive AI algorithms such as MiiHealth (MiiHealth, Inc., Phoenix, AZ, USA) represent a conversational AI assistant that collects patient reported data including symptoms, medical history and medications through specialty specific questioning and interactive dialogue. The platform then generates a structured clinical summary that can be reviewed by the clinician and integrated into the EHR. Designed to improve efficiency, reduce physician burnout and improve documentation accuracy, a health-system pilot is currently underway at Mayo Clinic to evaluate its impact on clinical workflow and physician experience.

While these findings are encouraging, the diagnostic accuracy of these algorithms remains under investigation, necessitating thorough validation and appropriate regulatory measures before broad implementation. A systematic review of 10 studies evaluating digital symptom checkers across various medical conditions found that primary diagnostic accuracy was consistently low, ranging from 19% to 37.9% (15). Moreover, current evidence remains largely observational, with no randomized trials demonstrating clear improvements in clinical outcomes, as study endpoints have largely been limited to diagnostic/triage accuracy, documentation quality, or workflow efficiency. Furthermore, human clinicians excel at listening, empathizing, and interpreting non-verbal cues. In contrast, text-based AI interfaces can feel impersonal, and patients may hesitate to share sensitive information due to privacy concerns with limited evidence on user acceptance of AI enhanced history-taking. Therefore, large, well-designed randomized controlled trials demonstrating net clinical benefit, patient acceptability, and safety are required before routine clinical use.

Physicians have been shown to spend equal or greater time on EHR-related tasks compared to direct patient care, often logging an additional two hours after hours outside of scheduled clinic time (7). This burden is a major contributor to the growing concerns of physician burnout. AI-driven solutions can address the increasing administrative burden on clinicians by automating data acquisition and structured documentation workflows. Additionally, AI can facilitate the conversion of unstructured clinical narratives into standardized coding systems such as Systematized Nomenclature of Medicine Clinical Terms (SNOMED CT) and International Classification of Diseases (ICD)-10, which are often underutilized in manual documentation. However, large-scale primary studies using real-world data are still needed to evaluate the performance of these technologies and establish their non-inferiority to current best practices. Equally important is a timely evaluation of their regulatory oversight and implementation.

AI-enhanced auscultation

After meeting with the patient and discussing the suspected diagnosis, AI-enhanced auscultation was performed during the physical examination. Within seconds, the system identified acoustic features consistent with severe mitral regurgitation (MR) and single lead ECG findings suggesting a reduced ejection fraction (EF) (<40%). A point-of-care alert from the AI auscultation-ECG system indicated suspected heart failure, advising an expedited echocardiogram and early targeted evaluation.

Introduced in the early 19^th century by the French physician René Théophile Hyacinthe Laënnec, cardiac auscultation continues to serve as a fundamental component of the cardiovascular examination (16). Despite its clinical utility, auscultation is often limited by its inherent subjectivity, limited sensitivity, and considerable interobserver variability. AI-enhanced stethoscopes aim to improve diagnostic accuracy by capturing acoustic signals and applying machine learning algorithms to analyze waveforms for subtle diagnostic features. This represents a promising approach to rapidly expanding the diagnostic capabilities of cardiac auscultation, particularly in settings with limited access to advanced imaging technologies or specialized cardiac expertise.

In a study of 369 patients aged 50 years and above without known valvular heart disease, the diagnostic accuracy of AI-enhanced stethoscopes was compared with traditional auscultation performed by primary care physicians (PCPs). When validated against transthoracic echocardiography (TTE), AI-assisted auscultation achieved a sensitivity of 94.1%, more than twice that of conventional auscultation (41.2%) (17). A potential limitation of this study is the exclusion of cardiologists, who are presumed to have greater expertise in cardiac auscultation compared to PCPs. Furthermore, in a randomized virtual clinical trial, Thompson et al. analyzed 3,180 heart recordings from 603 patients using the Johns Hopkins Auscultatory Database. The algorithm achieved a sensitivity of 93% (95% CI: 90–95) and specificity of 81% (95% CI: 75–85) for identifying pathological murmurs (18). Notably, recordings that were deemed “noisy” or lacking audible murmurs were excluded prior to testing, which may limit the generalizability of the findings to real-world clinical settings. While these studies demonstrate improved diagnostic accuracy under experimental conditions, clinical utility in routine practice has yet to be demonstrated.

For the diagnosis of aortic stenosis (AS), in a prospective, blinded, diagnostic accuracy study an AI-integrated stethoscope employing infrasound technology also demonstrated impressive diagnostic performance, achieving a sensitivity of 84% and 90% respectively in two independent cohorts (19). AI-enhanced auscultation has proven equally effective in detecting congenital heart disease. In a prospective, multicenter, cross-sectional diagnostic study of 1,362 children, AI machine learning algorithms demonstrated 97% sensitivity, 89% specificity, and 96% accuracy in identifying abnormal heart sounds compared to expert face-to-face auscultation (20). Despite the need for further validation, these findings demonstrate that the use of AI-integrated stethoscopes during cardiac assessment can yield a substantial increase in diagnostic sensitivity.

In addition, the integration of a single-lead ECG into AI-auscultation systems has expanded their diagnostic scope even further, extending beyond the assessment of valvular heart disease. ECG-enabled commercial stethoscopes (e.g., Eko DUO) have been used in studies integrating single-lead ECG with phonocardiography to screen for reduced EF (21). Bächtiger et al. conducted a prospective multicenter study of 1,050 adults, in which AI predictions of EF <40% based on 15-second single-lead ECG and phonocardiogram recording were compared to echocardiographic measurements. When compared to the four standard precordial sites, diagnostic performance was found to be highest at the pulmonary-valve position. The algorithm achieved an area under the receiver operating characteristic curve (AUROC) of 0.85 for detecting EF <40%, with a sensitivity of 84.8% and specificity of 69.5% (22).

More recently, in a multicenter, largely retrospective observational study involving 2,960 adults across four health systems, a similar 15-second recording obtained using a digital stethoscope was compared to an echocardiogram obtained within one week as the reference standard. An AI model accurately identified patients with a reduced EF (<40%), with an AUC of 0.85, sensitivity of 77.5% and a negative predictive value (NPV) of 98.0%. Among individuals with false-positive results, 87% had either a mildly reduced EF (41–49%) or conduction abnormalities (23). Findings from these multicenter studies support the potential role of AI-enhanced auscultation as a practical and reliable screening approach for heart failure.

Advances in wearable AI-assisted stethoscopes have broadened the scope of auscultation outside traditional clinical settings, with intuitive, patient-focused designs that improve their usability. Commercial wearables/handhelds such as Stethee Pro and ECG-enabled stethoscopes (e.g., Eko DUO) exemplify designs aimed at intuitive patient use with app-based guidance and cloud analytics (21,24). Wickramathilaka et al. demonstrated that patients can effectively operate an AI-integrated stethoscope to detect reduced EF, achieving diagnostic accuracy similar to clinicians, with an intraclass correlation coefficient (ICC) of 0.817 and 0.888 respectively (25). These findings lend support to decentralization and democratization of healthcare delivery, enabling patients to perform accurate self-screening at home. The integration of AI into stethoscopes also provides a unique educational advantage by offering real-time analytic feedback to learners, transforming a historically subjective skill into a quantifiable, teachable component of the cardiovascular assessment.

Despite encouraging diagnostic performance, current AI-assisted auscultation systems are supported almost entirely by non-randomized diagnostic accuracy studies and observational cohort validations. These results primarily support concordance with expert readers under controlled conditions rather than establish real-world effectiveness and impact on patient outcomes. Clinical benefit and cost effectiveness will depend on sustained diagnostic accuracy across diverse environments, body habitus and arrhythmia burden. As an emerging tool, broad adoption should await multicenter, pragmatic randomized trials demonstrating improvements in management and patient outcomes.

AI-assisted auscultation holds considerable potential as a transformative tool in cardiovascular diagnostics, offering a rapid, bedside screening tool for valvular pathology, heart failure, and cardiac arrhythmia. This may prove to be especially valuable in underserved regions with limited access to advanced diagnostics, allowing for the early detection of clinically significant cardiac abnormalities and facilitating timely escalation to specialist care. Given the significant burden heart failure places on public health systems worldwide, early detection and management at the primary care level have remained a long-standing public health priority. Backed by the UK’s NHS, the TRICORDER study is an ongoing two-arm, multi-center randomized controlled trial designed to evaluate whether AI-enabled auscultation in primary care improves early detection of heart failure, atrial fibrillation and valvular disease. This trial will help inform both the cost-effectiveness and practical implementation of this pivotal new technology to routine clinical practice.

AI-enhanced electrocardiography

The patient then undergoes AI-enhanced12-lead ECG that produced an automated report identifying subtle electromechanical signatures of a low EF. The algorithm confirms normal sinus rhythm and a low probability of arrhythmias, including atrial fibrillation or atrial flutter. No features suggestive of chamber enlargement, hypertrophic cardiomyopathy, or cardiac amyloidosis (CA) were detected per ECG AI prediction. A prognostic engine subsequently generated a high-risk profile for cardiovascular deterioration if cardiotoxic chemotherapy is initiated without further optimization. These findings were automatically integrated into the EHR with an alert sent to the cardio-oncology team, which facilitated referral to a formal transthoracic echocardiogram for further evaluation.

The ECG is an indispensable diagnostic tool in the evaluation of acute cardiovascular events and a routine component of cardiovascular assessment. It remains one of the most accessible and affordable diagnostic tools, with proven reliability across diverse care settings, including hospitals and primary care. While automated ECG interpretation has long been available, diagnostic accuracy remains quite limited, hindering its use and reliability in today’s clinical practice. As a result, ECG analysis has become a primary target for AI integration, aiming to improve timely decision-making and risk stratification, as well as provide risk prediction for specific cardiac conditions (Figure 2).

Figure 2 Mayo Clinic AI-ECG dashboard displaying probability scores across multiple cardiovascular conditions from a single ECG. The algorithm flagged high probabilities for low EF and cardiac amyloidosis; subsequent echocardiography confirmed a left ventricular EF of 24%. AF, atrial fibrillation; AI, artificial intelligence; ECG, electrocardiogram; EF, ejection fraction; HCM, hypertrophic cardiomyopathy.

Myriad machine learning models have been developed and validated for the rapid evaluation of ECGs, enabling the detection of complex electrical patterns that may not be readily discernible to the human eye. A retrospective validation study by Herman et al. demonstrated the diagnostic accuracy of six specific deep neural network (DNN) models using more than 900,000 12-lead ECGs (2). Each model was assessed across six diagnostic domains, including arrhythmia, acute coronary syndrome, conduction abnormalities, ectopy, chamber enlargement, and axis determination, achieving excellent F1 scores between 0.893 and 0.972. Limitations include validation from the same telehealth group using a computerized ECG algorithm which limits generalizability. Furthermore, Ribeiro et al. evaluated a deep learning (DL) model on more than 2 million ECGs, demonstrating diagnostic accuracy surpassing that of resident physicians, when detecting arrhythmias such as atrial fibrillation, sinus tachycardia, sinus bradyarrhythmia, right and left bundle branch block, and first-degree atrioventricular (AV) block (3). An important limitation is the use of resident physicians as the reference standard which limits the strength of the comparison and may overstate relative model performance. A prospective pilot study evaluating 2,261 subcutaneous ECG recordings demonstrated a diagnostic accuracy of 95.4% for the detection of cardiac arrhythmias, and an NPV of 98.55%. Subsequent to these investigations, numerous studies have corroborated the high diagnostic accuracy of AI algorithms in identifying a wide range of cardiac arrhythmias (26,27).

Another core clinical application of the standard ECG is the identification of high-risk segmental changes that may correspond to ischemic myocardial injury. This was evaluated in the recently published prospective multicenter ROMIAE study, which included 8,493 adult patients presenting with chest pain, in whom acute myocardial infarction (AMI) was suspected (4). The area under the curve (AUC) of ECG-AI to detect AMI was 0.878 with a sensitivity of 77% and specificity of 85%. These results were comparable or superior to previously validated risk prediction scores that involve multiple parameters and cardiac biomarkers. Similar findings were reported in the ARISE study, a cluster-randomized controlled trial conducted in Taiwan, which demonstrated that ECG-AI achieved an NPV of 99.9% for STEMI detection, with significantly shorter time to intervention in AI-positive cases (28). Recent work has extended the application of AI to post-percutaneous coronary intervention (PCI) risk prediction. Using convolutional neural network (CNN) with block attention, Thakur et al. demonstrated AUROC scores of 0.849, 0.913, and 0.794 for predicting heart failure, all-cause mortality, and stroke post-PCI in a retrospective, multi-center cohort of PCI patients (10).

As summarized by Palermi et al., contemporary AI-ECG models extend far beyond arrhythmia and ischemic disease, with validated applications in detecting subclinical left ventricle (LV) dysfunction, hypertrophic cardiomyopathy, valvular heart disease, pulmonary hypertension, and systemic phenotypes such as hyperkalemia and cardiometabolic risk (11). Ko et al. was a retrospective case control study which evaluated a CNN-based ECG-AI model for the detection of hypertrophic cardiomyopathy using a dataset of 51,153 ECGs, reporting an AUC of 0.96, with a sensitivity of 87% and specificity of 90% (27,29). Comparable diagnostic performance has been observed in the detection of CA, with ECG-AI models achieving AUC values exceeding 0.85 across multiple large, independent cohort studies (30-32). In a pre-procedural context, an ECG-AI algorithm was applied to a retrospective cohort of 1,426 transcatheter aortic valve replacement (TAVR) patients, identifying 24.4% as high probability for CA. These individuals were subsequently found to be independently associated with increased all-cause mortality [hazard ratio (HR) 1.40, P=0.046] and heart failure hospitalizations (HR 1.58, P=0.008) at 1 year (33).

In patients receiving cardiotoxic agents such as chemotherapy, AI-enhanced ECG may assist in identifying those at increased risk for heart failure deterioration, enabling more targeted therapeutic and monitoring approaches. In a study of approximately 100,000 patients, an ECG-AI model accurately identified reduced left ventricular ejection fraction (LVEF) (≤35%) with a sensitivity of 93% and an AUC of 0.93. Notably, elevated ECG-AI risk scores even among those with normal baseline echocardiograms predicted a four-fold higher risk of future systolic dysfunction (21). Among 3,439 patients treated with anthracyclines, elevated post-treatment ECG-AI scores demonstrated meaningful diagnostic performance for detecting severe cardiac dysfunction (AUC 0.761) and were independently associated with increased mortality at 1 year (HR 2.19, 95% CI: 1.92–2.51) and 5 years (HR 1.69, 95% CI: 1.54–1.87) (34). In a retrospective, multi-center cohort of 2,258 patients treated with immune checkpoint inhibitors, Ayoub et al. created a multimodal fusion AI model combining ECG and clinical parameters, demonstrating robust diagnostic accuracy for myocarditis and major cardiovascular events (AUC 0.72, NPV 0.98) (12). These tools can help inform important treatment discussions in patients with a heightened risk for resultant cardiotoxicity.

AI-ECG is also poised to play an important role in sports cardiology, where integration of serial ECGs with imaging and wearable-derived data may help distinguish physiological athlete’s heart from early cardiomyopathy, particularly as misclassification can lead to unnecessary exclusion from sport or missed sudden cardiac death risk (35). This potential is supported by AI-ECG models originally developed in the general population to detect LV dysfunction, silent AF, and hypertrophic cardiomyopathy, which Palermi et al. highlight as particularly relevant for pre-participation screening as they account for a substantial proportion of sudden cardiac death in athletes (35).

Despite the abundance of data demonstrating the diagnostic accuracy of AI-ECG systems, current evidence remains dominated by retrospective or non-randomized cohort studies and technical validations, often performed within single health systems. These models have demonstrated excellent performance in ECG pattern recognition, detection of hidden structural disease, and risk stratification, with several cohorts demonstrating that AI-ECG scores are strongly associated with subsequent heart failure, Major adverse cardiovascular events (MACE), or cancer-therapy—related cardiotoxicity. However, randomized trials have yet to demonstrate that AI-ECG—guided screening or management improve mortality, heart failure hospitalizations, or other patient-centered outcomes.

Beyond algorithm accuracy, ECG interoperability and electronic medical record (EMR) integration is hindered by inconsistent waveform standards [standard communications protocol (SCP)-ECG, Digital Imaging and Communications in Medicine (DICOM)-ECG, Health Level Seven International (HL7) augmented electrocardiography (aECG)], device heterogeneity and lead placement, which can affect model calibration across sites. In the PhysioNet/Computing in Cardiology 2020 Challenge, an international ECG benchmarking initiative, algorithms scored lower on hidden test sets from unseen sources, demonstrating a discrepancy in performance across sites and the need for external validation (36). Accordingly, while AI-ECG should be regarded as a powerful clinical adjunct, wider clinical implementation will require more robust evidence.

ECG-AI enables faster, more accurate interpretation, which can facilitate efficient clinical workflows and rapid clinical decision making. Advances in AI-driven risk prediction are transforming the ECG from a snapshot diagnostic tool into a longitudinal digital biomarker that enables dynamic, individualized risk stratification (11). Formal integration of ECG-AI into clinical pathways, guided by robust validation studies and inclusion in practice guidelines, is especially necessary in acute care, where rapid and accurate interpretation is crucial for patient outcomes. Spanning 48 practices and 22,641 patients, the EAGLE trial confirms that AI-ECG increased newly detected low-ejection-fraction cases by up to 32% without an increase in imaging burden. Concurrently, the FDA has cleared multiple standalone algorithms validating a regulatory pathway for future clinical adoption. Key next steps include robust implementation studies and dynamic regulatory guidance that balance regular algorithm updates while ensuring diagnostic accuracy and patient safety.

AI-enhanced echocardiography

An echocardiogram was performed using an AI-guided platform that provided real-time feedback to the sonographer, suggesting subtle anatomical maneuvers to improve image quality and markedly shorten the time required to complete the study. Automated analysis immediately demonstrates a dilated LV with globally mildly reduced systolic function, severe functional MR with a central jet and secondary left atrial enlargement. The AI system generates a structured report that synthesizes these findings into a likely diagnosis of dilated cardiomyopathy with severe secondary MR and suggests guideline-based considerations to be reviewed by the on-call cardiologist.

Echocardiography serves as the cornerstone of cardiovascular imaging, allowing for a non-invasive, real-time evaluation of cardiac structure and function. Given its accessibility, the amount of information it provides, and lack of radiation, it has an expanding role across a range of cardiac conditions. Echocardiography has quickly become the most widely utilized cardiac imaging modality, with demand continuing to rise substantially. Additionally, advancements such as strain imaging, tissue Doppler, and three-dimensional (3D) imaging have increased both the complexity of echocardiographic assessment and the time required for image acquisition and interpretation. As a result, there has been growing interest in AI-enhanced echocardiography and its potential to automate image interpretation and quantitative analysis, improving diagnostic accuracy and workflow efficiency.

AI-enhanced echocardiography aims to improve both the speed and quality of image acquisition, through machine learning algorithms that detect non-diagnostic or misaligned views, providing real-time corrective feedback to the user (37). A multicenter, randomized prospective study including 240 patients by Narang et al. demonstrated that nurses without prior experience in echocardiography were able to acquire images of diagnostic quality comparable to those obtained by experienced sonographers, after a short training period using an AI-guided system (5). In a mixed prospective and retrospective validation study by Olaisen et al., AI-supported measurements of left ventricular volumes and end-diastolic function led to a 77% reduction in total acquisition and processing time (from 7 minutes 30 seconds to just 1 minute 54 seconds) in 50 patients, without compromising accuracy (6). These studies were limited by a restricted set of views acquired and a focus on measurement agreement not patient outcomes.

AI models have also shown promise in point-of-care ultrasound (POCUS) further extending cardiac imaging into primary care and frontline settings. This can solve key problems with POCUS including high operator dependence and huge variability in image quality and diagnostic accuracy. In a prospective, multicenter validation study Mor-Avi et al. embedded a deep-learning guidance algorithm (UltraSight) into a handheld device and demonstrated that nurses and medical residents with minimal ultrasound training could acquire 10 standard transthoracic views of diagnostic quality in the vast majority of patients [≥99% for LV size and function, 93% for right ventricle (RV) size, and 100% for pericardial effusion], with diagnostic interpretations concordant with expert sonographer images in 83–96% of cases. Limitations included the use of single handheld platform and AI solution in controlled echo-lab setting which limits generalizability (38). By enabling non-experts to acquire reliable cardiac views, AI-guided POCUS can make accurate bedside echocardiography routine in everyday care, allowing for rapid assessment of hemodynamics, LV function and quantifying a pericardial effusion to help reduce dependence on formal studies for initial decision-making.

Beyond image capture and analysis, AI is now being applied to the reporting phase of echocardiography. Using over 97,000 echocardiography reports from the Mayo Clinic, a fine-tuned Llama-2 model (“EchoGPT”) generated reports with win rates of 87–99%. EchoGPT’s outputs were significantly preferred for conciseness (P<0.001) without differences in completeness, correctness, or clinical utility (39). Collectively, these tools illustrate an end-to-end AI driven workflow, offering faster turnaround, reduced workload, and consistent diagnostic quality.

AI-based echocardiography has demonstrated particular value in scenarios where diagnosis depends on recognizing subtle motion patterns or rare disease-specific features that may be challenging to visually appreciate. A DL model based on the standard apical four-chamber echocardiographic view was developed to differentiate constrictive pericarditis (CP) from CA. Trained on 381 cases in a multicenter retrospective study, the ResNet50-based model achieved excellent performance with an AUC of 0.97 internally and 0.84 on an external dataset. GradCAM analysis revealed predominant attention to the ventricular septal region, aligning with known pathophysiological features of CP (40). A recent multicenter study developed a DL tool that effectively differentiated CA from phenotypic mimickers using a single apical four-chamber clip with 85% sensitivity and 93% specificity, demonstrating consistent performance across CA. The model outperformed traditional TTE-based methods, maintaining accuracy across racial/ethnic subgroups. Future integration of these algorithms into routine echocardiography could expedite differentiation among phenotypically overlapping cardiomyopathies, decrease downstream invasive testing, and enable timely referral to disease-directed therapies (Figure 3).

Figure 3 Apical four-chamber echo with AI-assisted classification and strain. (A) Standard apical four-chamber view. (B) AI classifier for thick-wall cardiomyopathies predicted cardiac AMY with high probability (0.97); Grad-CAM attention map highlights ventricular myocardium as the most discriminative regions. (C) Standard TTE reveals apical-sparing longitudinal strain pattern, consistent with cardiac AMY. AI, artificial intelligence; AMY, amyloidosis; ANT, anterior; Grad-CAM, Gradient-weighted Class Activation Mapping; HCM, hypertrophic cardiomyopathy; HTN, hypertension; INF, inferior; LAT, lateral; NC, normal cardiac; POST, posterior; SEPT, septal; TTE, transthoracic echocardiography.

Advances in AI now allow algorithms to extract information extending beyond overt structural abnormalities, broadening their diagnostic scope. These models enable the detection of phenotypic patterns and incidental findings that may not be readily apparent to the human eye. In a large retrospective study using a CNN trained on >10,000 echo studies, Ouyang et al. demonstrated that AI could predict patient demographics including age (R²=0.46), sex (AUC =0.88), weight (R²=0.56), and height (R²=0.33) (41). A retrospective study using an AI algorithm to predict cardiac age from over 2.6 million echo videos, found that risk stratification by age prediction revealed associations with increased risk of coronary artery disease (CAD), heart failure, and stroke (42). These findings suggest the presence of hidden biomarkers that are predictive of systemic phenotypes and outcomes, offering new avenues for cardiovascular risk reduction and screening at scale.

While the majority of AI research has focused on TTE, advancements are also being extended to transesophageal echocardiography (TEE) and 3D echocardiography. A recent prospective method study has shown AI-integrated 3D TEE can enable continuous monitoring of LV function in postoperative cardiac patients, by the automated measurement of mitral annular plane systolic excursion. The findings were consistent with manual echocardiographic measurements and showed significant associations with N-terminal pro B-type natriuretic peptide (ρ =−0.37, P=0.008) and high-sensitivity troponin-T (ρ =−0.28, P=0.047), supporting its potential role as a valuable adjunct to hemodynamic monitoring (43). Given the complexity and volume of 3D echocardiographic data, AI offers a promising solution for extracting meaningful clinical information and automating the labor-intensive task of manual valve topographic mapping.

While significant progress has been made, several limitations and barriers to the integration of AI-enhanced echocardiography remain. In the presence of artifacts or foreshortened views, AI algorithms may misidentify the endocardial border and proceed with analysis of suboptimal images, whereas an experienced human operator would recognize the limitation and adjust the acquisition accordingly (44). Furthermore, current systems may have difficulty distinguishing fine anatomical structures, such as papillary muscles and chordae tendineae, potentially resulting in overestimation of wall thickness (44). Similar to other applications, AI echocardiography systems have been evaluated in non-randomized observational and validation studies, with only a single randomized trial showing improved acquisition quality for novice operators and several larger cohorts demonstrating excellent agreement with expert measurements and promising prognostic association. These investigations primarily rely on endpoints involving image quality and measurement reproducibility, as we await trials which assess their ability to improve patient outcomes. These challenges highlight the need for robust quality control mechanisms and rigorous prospective, outcome-focused trials to ensure safe and accurate clinical implementation.

The current landscape of AI-enhanced echocardiography reflects a transition from research to clinical practice, with an increasing number of echocardiograms now incorporating some degree of AI assistance. Several AI-based echocardiography tools have received FDA approval, including those for automated strain analysis, EF measurement, and assessment of CA (45,46). Robust multicenter clinical trials and continued collaboration between clinicians and regulatory bodies are essential to ensure proper validation and the safe integration of AI into clinical workflows. Current literature suggests AI stands to transform echocardiography by optimizing workflow efficiency, reducing provider fatigue, and enhancing diagnostic capabilities, including the generation of new data and phenotypic insights previously beyond the reach of conventional imaging modalities.

AI-enhanced advanced cardiac imaging

To further clarify the etiology and ischemic burden before proceeding with high-risk chemotherapy, the patient is referred for advanced imaging. Coronary computed tomography (CT) angiography is performed on an AI-enabled platform that automatically quantifies total and component plaque volume, identifying mild, non-flow-limiting atherosclerotic plaque. An AI-assisted cardiac MR protocol provides automated quantification of ventricular volumes and predicts late gadolinium enhancement (LGE) burden from cine images, suggesting a non-ischemic dilated cardiomyopathy. A multimodal fusion model integrates data from the AI-derived history, auscultation, ECG, echocardiography, and advanced imaging to generate an individualized risk profile for heart failure decompensation, arrhythmias, and treatment-related cardiotoxicity over the coming years. This information is used to support a shared decision-making conversation with the oncologist and patient regarding treatment options and risk.

AI is revolutionizing cardiovascular imaging by streamlining complex workflows, enhancing the detection of subtle pathological features, and enabling prognostic risk stratification. Cardiac magnetic resonance imaging (CMR), coronary CT angiography (CCTA), and invasive coronary angiography are commonly used imaging modalities that provide detailed structural and functional information. Recent advances in AI now enable fully automated analysis of these images, improving diagnostic accuracy, reducing interpretation time, and supporting clinical decision-making across a range of cardiovascular conditions (Figure 4).

Figure 4 AI enhances coronary CT, angiography, and cardiac MRI by speeding interpretation, improving accuracy, and revealing new prognostic markers. AI, artificial intelligence; CT, computed tomography; MR, magnetic resonance; MRI, magnetic resonance imaging.

CCTA

CCTA is a noninvasive imaging modality widely used to assess CAD. Recent advancements in AI have further enhanced both the anatomical characterization and functional evaluation of CAD on CT imaging. In a recent international, multicenter observational study involving 921 patients across 11 sites, AI-algorithms used to quantify total plaque volume and percent stenosis showed excellent agreement with expert readers (intraclass correlation 0.96 for plaque volume) and intravascular ultrasound (IVUS) measurements (ICC ~0.95 for plaque volume) (47). Notably, the algorithm demonstrated remarkable efficiency analyzing images in just 5.6 seconds compared to 26 minutes for manual experts. Another retrospective study evaluated an AI-based CCTA analysis against invasive quantitative coronary angiography (QCA) in 841 vessels, which achieved high per-patient sensitivity (~94%) and NPV (~97%) for detecting obstructive CAD (≥70% stenosis) (48). These studies demonstrate that such tools can rapidly quantify coronary plaque burden and stenosis with accuracy approaching invasive methods. However, notable limitations include small, highly selective sample sizes including patients undergoing both CCTA and invasive angiography, limiting generalizability.

A major advancement in CCTA has been the introduction of fractional flow reserve (FFR) using computational fluid dynamics (CFD). More recent approaches have leveraged machine learning algorithms, which demonstrated excellent correlation (R=0.997) with CFD-based FFR, enabling functional assessment of coronary stenosis on standard workstations without the need for off-site processing or supercomputing resources. Additionally, in an international multicenter observational study assessing the prognostic value of DL-based plaque quantification in 1,611 patients, a total plaque volume of ≥238.5 mm³ was independently associated with an increased risk of fatal or non-fatal myocardial infarction over long-term follow-up (HR 5.36, 95% CI: 1.70–16.86; P=0.004) (49). This highlights the potential of AI-derived imaging biomarkers to serve as prognostic indicators beyond traditional visual assessment.

DL algorithms can now quantify coronary artery calcium (CAC) directly on a routine, non-ECG-gated chest CT to opportunistically detect CAC and aortic valve calcification (AVC) on thoracic CT scans obtained for noncardiac indications. In a multicenter, retrospective study an AI system achieved excellent agreement with manual CAC scoring on ECG-gated CT (R²=0.989–0.995; ICC =0.990–0.997; κ ≈0.97; AUROC =0.99) and sustained performance on non-gated low-dose chest CT (R²=0.926–0.988; ICC =0.948–0.994; κ =0.71–0.96; AUROC =0.89–0.99) (50). In the randomized NOTIFY-1 project, AI-enabled identification and clinician notification of incidental CAC on prior CT scans increased 6-month statin prescriptions to 51.2% vs. 6.9% with usual care and increased downstream CAD testing (15.1% vs. 2.3%), demonstrating tangible clinical benefits to its implementation (51).

Identifying AVC on thoracic CT scans also provides meaningful benefit, with observational data demonstrating that AVC identified on lung-screening CT independently predicts all-cause and cardiovascular mortality (52). A machine-learning software (Bunkerhill AVC; Bunkerhill Health, San Francisco, CA) for automated detection and quantification of AVC on non-gated, non-contrast chest CT received FDA clearance in January 2025, highlighting a promising pathway for opportunistic valve screening on routine imaging (53).

In addition to its applications in sophisticated cardiac imaging, AI is demonstrating considerable promise in conventional modalities like chest X-ray and mammography to extract clinically relevant data for cardiovascular risk assessment. AI models have been applied to chest radiography (CXR), demonstrating the ability to detect a wide range of cardiovascular condition including heart failure, pulmonary hypertension, CAD, valvular disease, aortic pathology, and atrial fibrillation (54). Across numerous studies, DL algorithms have outperformed traditional CXR interpretation, identifying subtle imaging features beyond human perception and achieving diagnostic accuracies comparable to echocardiography or invasive testing in certain settings. Prognostic models derived from CXR have also stratified patients by risk for adverse cardiovascular events, enabling targeted referral for advanced imaging (54). Another AI model quantified breast arterial calcification (BAC) from screening mammography, demonstrating that higher scores were associated with an increased long-term risk of myocardial infarction, stroke, and cardiovascular death. Despite only moderate correlation with coronary calcium scores, BAC burden remained a strong predictor of adverse cardiovascular outcomes over 15 years (55).

Nuclear imaging

Single-photon emission computed tomography (SPECT) and positron emission tomography (PET) cardiac imaging provide valuable insights into myocardial perfusion and viability, though remains limited by image noise, attenuation artefacts and lengthy acquisition protocols. A single center prospective study developed a DL diagnostic tool based on SPECT images that achieved 88% accuracy with AUC ≈0.91 and area under the precision-recall curve (AUPRC) 0.87 for identifying CAD. AI-assisted readers achieved 80% accuracy (from 65% for cardiologists alone) while reducing mean reading time from 31 to 12 minutes (56). Additionally, in a multicenter retrospective study of 1,683 patients without known CAD undergoing SPECT studies, a DL model outperformed conventional quantitative assessments of total perfusion defect in detecting obstructive CAD at both the per-patient and per-vessel levels, demonstrating robust generalizability across multiple sites (57). For PET, emerging denoising and low-count methods aim to preserve myocardial blood flow quantification while reducing noise and dose, with early studies reporting promising retention of image quality (58). Machine-learning models are also yielding promising performance in predicting major adverse cardiac events, indicating the prognostic potential of regional perfusion patterns (59). These advancements led to the recent FDA approval of a deep-learning reconstruction algorithm (Clarify DL; GE Medical Systems, LLC, Waukesha, WI) for SPECT imaging, designed to lower noise, maintain quantitative accuracy, and shorten scan/read times (60).

Cardiac MR

CMR is an increasingly utilized modality but remains highly dependent on manual parameter selection, contributing to the increased technician workload and inter-operator variability. AI integration streamlines this process by reducing scan and post-processing times and improving diagnostic accuracy. Edalati et al. demonstrated that AI-driven shimming improved magnetic field uniformity and boosted signal-to-noise ratio by 12.5%, while a separate AI algorithm automated slice alignment and reduced scan time and operator dependence (61). Similarly, Wood et al. used U-Net-based segmentation and 3D-DenseNet landmark detection to automatically identify the cardiac resting phase, accelerating trigger delay calculation for coronary magnetic resonance angiography (MRA) (62).

AI algorithms have also been trained to recognize specific cardiac diseases on CMR. A recent multi-center study developed a DL model to screen cardiac abnormalities and to classify 11 different cardiomyopathies using cine and LGE images. The AI-based screening and diagnostic models demonstrated excellent performance, achieving AUC values of 0.988%±0.3% for screening and 0.991%±0.0% for diagnosis in a cohort of 9,719 patients. Notably, these models outperformed cardiologists in the diagnosis of pulmonary arterial hypertension (63).

Cardiac catheterization

In the cardiac catheterization lab, AI applications can assist in real-time lesion assessment and decision-making during invasive coronary angiography. A 2021 study used a DL model to classify intermediate coronary lesions (50–70% stenosis by eye) as FFR-positive vs. negative, achieving an AUC ~0.81 for predicting FFR ≤0.80 (64). These models aim to provide real-time interpretation of angiographic lesions, flagging areas that might be significant and require intervention. Additionally, using IVUS or optical coherence tomography (OCT), AI can automatically characterize plaque composition distinguishing fibrous from calcific lesions and quantifying lesion length to facilitate stent planning.

Despite significant progress, the evidence base for AI-enhanced cardiovascular imaging remains largely observational. Most studies of AI applied to coronary CT, nuclear imaging, CMR, invasive angiography, and CXR are retrospective; prospective cohorts to date have tended to prioritize diagnostic agreement, plaque burden, image quality, or risk prediction, rather than patient-centered endpoints. A few systems demonstrate that AI-derived imaging biomarkers are strongly associated with future events, and one randomized quality-improvement trial (NOTIFY-1) showed that AI-enabled opportunistic coronary calcium detection on prior chest CT can increase statin initiation (65). However, no AI imaging tool in the current literature has yet been proven to reduce myocardial infarction, heart failure hospitalizations, or mortality when incorporated into routine care.

Next-generation AI systems are anticipated to integrate data from multiple imaging modalities in conjunction with clinical variables to generate comprehensive diagnostic and prognostic assessments. AI is poised to become increasingly embedded within the imaging workflow itself. For instance, during CCTA or coronary angiography, real-time AI algorithms could provide augmented reality overlays to flag potentially concerning findings, thereby assisting operators during the procedure. In the magnetic resonance imaging (MRI) setting, AI could oversee image acquisition in real time, prompting repeat scans of suboptimal slices or automatically adjusting pulse sequences based on preliminary findings. The next decade will be pivotal in translating AI’s technical advancements into meaningful, patient-centered outcomes.

Limitations

While AI stands as a transformative tool in cardiovascular diagnostics, limitations need to be addressed to permit safe deployment. Firstly, AI systems are predominantly reliant on data derived from training populations, in which datasets may be affected by measurement variability (e.g., troponin levels from assays with differing sensitivity, LVEF estimated from poorly standardized echocardiography), underrepresentation of minority subgroups and variations in disease prevalence. Such imbalances have tangible clinical consequences. For instance, one study reported that an AI model exhibited significantly reduced performance in minority populations, with notable disparities in automated ventricular segmentation on CMR across different racial groups (66). This challenge is further compounded by the tendency of AI models to hinge on idiosyncrasies within the training dataset rather than true underlying pathophysiology, raising concerns about reliance on spurious correlations over clinically meaningful features.

Supporting this concern, a recent study demonstrated that model performance is frequently overestimated by up to 20% on average due to “shortcut learning” of hidden data acquisition biases (67). This further extends to the current literature on AI, which mainly relies on retrospective single-center studies with inconsistent and unvalidated reference standards. These concerns highlight the need for high quality, well-curated training datasets that fully represent diverse populations, accompanied by transparent reporting of model strengths and limitations. Safe deployment also requires external and prospective validation across multiple centers and populations with prespecified endpoints and clear reporting standards.

Furthermore, evidence from the radiology literature indicates that clinicians are vulnerable to automation bias when using AI assistance. In a randomized mammography study, radiologists across experience levels made more errors when AI provided incorrect cues, which can be exacerbated by time pressure, low disease prevalence and assertive AI messaging. Psychological studies in human-machine interaction show that when presented with high-confidence AI outputs, users tend to accept recommendations without adequate verification, raising critical patient safety concerns (68). While AI models may achieve performance comparable to that of expert clinicians, they remain fundamentally computational tools and exclusive reliance on their outputs in cardiovascular diagnostics risks introducing preventable errors without clinician oversight.

Over the past several years, numerous AI-enabled cardiovascular models have received U.S. FDA clearance for clinical use (Table S1). Regulatory oversight must also keep pace with continuously learning models that continue to evolve after deployment, adapting to new data and improving diagnostic capabilities. Without the appropriate safeguards, these updates risk introducing unintended and unvalidated changes that may impact performance and overall risk. To address this, emerging frameworks such as the FDA’s Total Product Lifecycle approach which recommends predetermined change plans, continuous monitoring, and clear documentation of algorithm updates, ensuring that iterative improvements do not compromise patient safety (69). Dedicated analysis must also assess data drift, where the characteristics of incoming clinical data may lead to subtle but clinically significant reductions in diagnostic accuracy if left unmonitored. While AI delivers unprecedented gains in cardiovascular diagnostics, it also introduces novel risks that necessitate vigilant and dynamic oversight.

Conclusions

Integrating AI across the diagnostic pathway from history-taking and automated documentation, physical exam augmentation and the interpretation of ECG, echocardiography and advanced cardiac imaging creates a longitudinal layer of support for clinicians. These innovations have shown the potential to address increasing workforce demands, enable earlier disease detection, and expand access to advanced diagnostic capabilities within primary care and resource-limited settings. The integration of AI components into multimodal, workflow-embedded systems that continuously aggregate clinical and imaging data will enhance individualized risk prediction and treatment recommendations. The current phase should prioritize multicenter prospective outcome trials and integration studies that also address quality control and regulatory oversight, potential overreliance, shortcut learning, and data drift. This paradigm shift can deliver considerable improvements in equity, access, and cardiovascular outcomes.

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-550/rc

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-aw-550/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-550/coif). C.A. serves as an unpaid editorial board member of Cardiovascular Diagnosis and Therapy from April 2025 to March 2027. C.A. reports a stipend role as Digital Media Section Editor for Circulation: Cardiovascular Imaging; unpaid leadership/service roles with the American Society of Echocardiography, including Chair, AI Curriculum Committee, and membership on the Education Committee, AI Task Force, and Industry Round Table Committee. The other 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. Di Cesare M, Perel P, Taylor S, et al. The Heart of the World. Glob Heart 2024;19:11. [Crossref] [PubMed]
2. Herman R, Demolder A, Vavrik B, et al. Validation of an automated artificial intelligence system for 12 lead ECG interpretation. J Electrocardiol 2024;82:147-54. [Crossref] [PubMed]
3. Ribeiro AH, Ribeiro MH, Paixão GMM, et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat Commun 2020;11:1760. [Crossref] [PubMed]
4. Lee MS, Shin TG, Lee Y, et al. Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: the ROMIAE multicentre study. Eur Heart J 2025;46:1917-29. [Crossref] [PubMed]
5. Narang A, Bae R, Hong H, et al. Utility of a Deep-Learning Algorithm to Guide Novices to Acquire Echocardiograms for Limited Diagnostic Use. JAMA Cardiol 2021;6:624-32. [Crossref] [PubMed]
6. Olaisen S, Smistad E, Espeland T, et al. Automatic measurements of left ventricular volumes and ejection fraction by artificial intelligence: clinical validation in real time and large databases. Eur Heart J Cardiovasc Imaging 2024;25:383-95.
7. Arndt BG, Beasley JW, Watkinson MD, et al. Tethered to the EHR: Primary Care Physician Workload Assessment Using EHR Event Log Data and Time-Motion Observations. Ann Fam Med 2017;15:419-26.
8. Baker A, Perov Y, Middleton K, et al. A Comparison of Artificial Intelligence and Human Doctors for the Purpose of Triage and Diagnosis. Front Artif Intell 2020;3:543405.
9. Wang C, Wang F, Li S, et al. Patient Triage and Guidance in Emergency Departments Using Large Language Models: Multimetric Study. J Med Internet Res 2025;27:e71613.
10. Thakur A, Agasthi P, Chao CJ, et al. Joint fusion of EHR and ECG data using attention-based CNN and ViT for predicting adverse clinical endpoints in percutaneous coronary intervention patients. Comput Biol Med 2025;189:109966.
11. Palermi S, Vecchiato M, Ng FS, et al. Artificial intelligence and the electrocardiogram: A modern renaissance. Eur J Intern Med 2025;140:106329.
12. Ayoub C, Appari L, Pereyra M, et al. Multimodal Fusion Artificial Intelligence Model to Predict Risk for MACE and Myocarditis in Cancer Patients Receiving Immune Checkpoint Inhibitor Therapy. JACC Adv 2025;4:101435.
13. Fukuzawa F, Yanagita Y, Yokokawa D, et al. Importance of Patient History in Artificial Intelligence-Assisted Medical Diagnosis: Comparison Study. JMIR Med Educ 2024;10:e52674.
14. Erden I, Sen A, Erden IE. The impact of automatic history-taking software on data quality in the cardiology outpatient clinic: Retrospective observational study. Digit Health 2024;10:20552076241260155. [Crossref] [PubMed]
15. Wallace W, Chan C, Chidambaram S, et al. The diagnostic and triage accuracy of digital and online symptom checker tools: a systematic review. NPJ Digit Med 2022;5:118. [Crossref] [PubMed]
16. Roguin A. Rene Theophile Hyacinthe Laënnec (1781-1826): the man behind the stethoscope. Clin Med Res 2006;4:230-5. [Crossref] [PubMed]
17. Bamford P, Abdelrahman A, Malkin CJ, et al. Artificial intelligence in heart valve disease: diagnosis, innovation and treatment. A state-of-the-art review. Br J Cardiol 2024;31:031.
18. Thompson WR, Reinisch AJ, Unterberger MJ, et al. Artificial Intelligence-Assisted Auscultation of Heart Murmurs: Validation by Virtual Clinical Trial. Pediatr Cardiol 2019;40:623-9. [Crossref] [PubMed]
19. Ghanayim T, Lupu L, Naveh S, et al. Artificial Intelligence-Based Stethoscope for the Diagnosis of Aortic Stenosis. Am J Med 2022;135:1124-33.
20. Lv J, Dong B, Lei H, et al. Artificial intelligence-assisted auscultation in detecting congenital heart disease. Eur Heart J Digit Health 2021;2:119-24.
21. Food US, Drug A. 510(k) Summary: Eko Model E5 System (Eko DUO) — K170874. U.S. Food and Drug Administration. 2017. Available online: https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K170874
22. Attia ZI, Dugan J, Rideout A, et al. Automated detection of low ejection fraction from a one-lead electrocardiogram: application of an AI algorithm to an electrocardiogram-enabled Digital Stethoscope Eur Heart J Digit Health 2022;3:373-9.
23. Guo L, Pressman GS, Kieu SN, et al. Automated Detection of Reduced Ejection Fraction Using an ECG-Enabled Digital Stethoscope: A Large Cohort Validation. JACC Adv 2025;4:101619.
24. Food US, Drug A. 510(k) Summary: Stethee Pro 1 — K172296. U.S. Food and Drug Administration. 2017. Available online: https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K172296
25. Wickramathilaka A, Kelshiker MA, Griffin J, et al. Abstract 12917: Reproducibility of Use by Physicians and Patients of a Smart Stethoscope for Artificial-Intelligence Enhanced Diagnosis of Heart Failure. Circulation 2022;146:A12917.
26. Katal N, Gupta S, Verma P, et al. Deep-Learning-Based Arrhythmia Detection Using ECG Signals: A Comparative Study and Performance Evaluation. Diagnostics (Basel) 2023;13:3605.
27. Siontis KC, Noseworthy PA, Attia ZI, et al. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nat Rev Cardiol 2021;18:465-78.
28. Cho Y, Kwon JM, Kim KH, et al. Artificial intelligence algorithm for detecting myocardial infarction using six-lead electrocardiography. Sci Rep 2020;10:20495.
29. Ko WY, Siontis KC, Attia ZI, et al. Detection of Hypertrophic Cardiomyopathy Using a Convolutional Neural Network-Enabled Electrocardiogram. J Am Coll Cardiol 2020;75:722-33.
30. Tison GH, Zhang J, Delling FN, et al. Automated and Interpretable Patient ECG Profiles for Disease Detection, Tracking, and Discovery. Circ Cardiovasc Qual Outcomes 2019;12:e005289.
31. Grogan M, Lopez-Jimenez F, Cohen-Shelly M, et al. Artificial Intelligence-Enhanced Electrocardiogram for the Early Detection of Cardiac Amyloidosis. Mayo Clin Proc 2021;96:2768-78.
32. Khan LA, Shaikh FH, Khan MS, et al. Artificial intelligence-enhanced electrocardiogram for the diagnosis of cardiac amyloidosis: A systemic review and meta-analysis. Curr Probl Cardiol 2024;49:102860.
33. Pereyra Pietri M, Farina JM, Mahmoud AK, et al. The prognostic value of artificial intelligence to predict cardiac amyloidosis in patients with severe aortic stenosis undergoing transcatheter aortic valve replacement. Eur Heart J Digit Health 2024;5:295-302.
34. Scalia IG, Farina JM, Pietri MP, et al. Artificial Intelligence for Identification of Patients with Increased Risk of Severe Cancer Therapy-Related Cardiac Dysfunction Following Anthracycline Therapy. Am J Med 2025;138:1561-1568.e1.
35. Palermi S, Vecchiato M, Saglietto A, et al. Unlocking the potential of artificial intelligence in sports cardiology: does it have a role in evaluating athlete’s heart? Eur J Prev Cardiol 2024;31:470-82. [Crossref] [PubMed]
36. Perez Alday EA, Gu A. Classification of 12-lead ECGs: the PhysioNet/Computing in Cardiology Challenge 2020. Physiol Meas 2021;41:124003. [Crossref] [PubMed]
37. Cheema B, Hsieh C, Adams D, et al. Abstract 15694: Automated Guidance and Image Capture of Echocardiographic Views Using a Deep Learning-Derived Technology. Circulation 2019;140:A15694.
38. Mor-Avi V, Khandheria B, Klempfner R, et al. Real-Time Artificial Intelligence-Based Guidance of Echocardiographic Imaging by Novices: Image Quality and Suitability for Diagnostic Interpretation and Quantitative Analysis. Circ Cardiovasc Imaging 2023;16:e015569. [Crossref] [PubMed]
39. Chao CJ, Banerjee I, Arsanjani R, et al. Evaluating large language models in echocardiography reporting: opportunities and challenges. Eur Heart J Digit Health 2025;6:326-39. [Crossref] [PubMed]
40. Chao CJ, Jeong J, Arsanjani R, et al. Echocardiography-Based Deep Learning Model to Differentiate Constrictive Pericarditis and Restrictive Cardiomyopathy. JACC Cardiovasc Imaging 2024;17:349-60.
41. Ghorbani A, Ouyang D, Abid A, et al. Deep learning interpretation of echocardiograms. NPJ Digit Med 2020;3:10.
42. Rawlani M, Ieki H, Binder C, et al. Artificial intelligence prediction of age from echocardiography as a marker for cardiovascular disease. NPJ Digit Med 2025;8:688.
43. Yu J, Taskén AA, Berg EAR, et al. Continuous three-dimensional transesophageal echocardiography and deep learning for perioperative monitoring of left ventricular longitudinal function. Eur Heart J Imaging Methods Pract 2025;3:qyaf052.
44. Hirata Y, Kusunose K. AI in Echocardiography: State-of-the-art Automated Measurement Techniques and Clinical Applications. JMA J 2025;8:141-50.
45. O’Driscoll JM, Tuttolomondo D, Gaibazzi N. Artificial intelligence calculated global longitudinal strain and left ventricular ejection fraction predicts cardiac events and all-cause mortality in patients with chest pain. Echocardiography 2023;40:1356-64.
46. Goto S, Mahara K, Beussink-Nelson L, et al. Artificial intelligence-enabled fully automated detection of cardiac amyloidosis using electrocardiograms and echocardiograms. Nat Commun 2021;12:2726.
47. Ihdayhid AR, Tzimas G, Peterson K, et al. Diagnostic Performance of AI-enabled Plaque Quantification from Coronary CT Angiography Compared with Intravascular Ultrasound. Radiol Cardiothorac Imaging 2024;6:e230312.
48. Dundas J, Leipsic JA, Sellers S, et al. Artificial Intelligence-based Coronary Stenosis Quantification at Coronary CT Angiography versus Quantitative Coronary Angiography. Radiol Cardiothorac Imaging 2023;5:e230124.
49. Lin A, Manral N, McElhinney P, et al. Deep learning-enabled coronary CT angiography for plaque and stenosis quantification and cardiac risk prediction: an international multicentre study. Lancet Digit Health 2022;4:e256-65.
50. Kim S, Park EA, Ahn C, et al. Performance of fully automated deep-learning-based coronary artery calcium scoring in ECG-gated calcium CT and non-gated low-dose chest CT. Eur Radiol 2025;35:7084-95.
51. Sandhu AT, Rodriguez F, Ngo S, et al. Incidental Coronary Artery Calcium: Opportunistic Screening of Previous Nongated Chest Computed Tomography Scans to Improve Statin Rates (NOTIFY-1 Project). Circulation 2023;147:703-14.
52. Christensen JL, Tan S, Chung HE, et al. Aortic valve calcification predicts all-cause mortality independent of coronary calcification and severe stenosis. Atherosclerosis 2020;307:16-20.
53. Bunkerhill H. Bunkerhill Health Secures the First FDA Clearance Enabling Automated Aortic Valve Calcification Quantification on Non-Cardiac Chest CT. 2025. Available online: https://www.bunkerhillhealth.com/resources/bunkerhill-health-secures-the-first-fda-clearance-enabling-automated-aortic-valve-calcification-quantification-on-non-cardiac-ct
54. Farina JM, Pereyra M, Mahmoud AK, et al. Artificial Intelligence-Based Prediction of Cardiovascular Diseases from Chest Radiography. J Imaging 2023;9:236.
55. Urooj A, Dapamede T, Patel B, et al. A Multi-Task Learning Approach for Segmentation of Breast Arterial Calcifications in Screening Mammograms. AMIA Annu Symp Proc 2024;2024:1139-48.
56. Kusumoto D, Akiyama T, Hashimoto M, et al. A deep learning-based automated diagnosis system for SPECT myocardial perfusion imaging. Sci Rep 2024;14:13583. [Crossref] [PubMed]
57. Betancur J, Commandeur F, Motlagh M, et al. Deep Learning for Prediction of Obstructive Disease From Fast Myocardial Perfusion SPECT: A Multicenter Study. JACC Cardiovasc Imaging 2018;11:1654-63. [Crossref] [PubMed]
58. Hashimoto F, Onishi Y, Ote K, et al. Deep learning-based PET image denoising and reconstruction: a review. Radiol Phys Technol 2024;17:24-46. [Crossref] [PubMed]
59. Bors S, Abler D, Dietz M, et al. Comparing various AI approaches to traditional quantitative assessment of the myocardial perfusion in [82Rb] PET for MACE prediction. Sci Rep 2024;14:9644.
60. Food US, Drug A. K241350: Clarify DL — Deep learning image reconstruction software for SPECT; 510(k) Summary. U.S. Food and Drug Administration. 2025. Available online: https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpmn/pmn.cfm?ID=K241350
61. Edalati M, Zheng Y, Watkins MP, et al. Implementation and prospective clinical validation of AI-based planning and shimming techniques in cardiac MRI. Med Phys 2022;49:129-43.
62. Wood G, Pedersen AU, Kunze KP, et al. Automated detection of cardiac rest period for trigger delay calculation for image-based navigator coronary magnetic resonance angiography. J Cardiovasc Magn Reson 2023;25:52.
63. Wang YJ, 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.
64. Cho H, Lee JG, Kang SJ, et al. Angiography-Based Machine Learning for Predicting Fractional Flow Reserve in Intermediate Coronary Artery Lesions. J Am Heart Assoc 2019;8:e011685.
65. Sandhu AT, Rodriguez F, Ngo S, et al. Incidental Coronary Artery Calcium: Opportunistic Screening of Previous Nongated Chest Computed Tomography Scans to Improve Statin Rates (NOTIFY-1 Project). Circulation 2023;147:703-14.
66. Puyol-Antón E, Ruijsink B, Mariscal Harana J, et al. Fairness in Cardiac Magnetic Resonance Imaging: Assessing Sex and Racial Bias in Deep Learning-Based Segmentation. Front Cardiovasc Med 2022;9:859310.
67. Ong Ly C, Unnikrishnan B, Tadic T, et al. Shortcut learning in medical AI hinders generalization: method for estimating AI model generalization without external data. NPJ Digit Med 2024;7:124.
68. Goddard K, Roudsari A, Wyatt JC. Automation bias: a systematic review of frequency, effect mediators, and mitigators. J Am Med Inform Assoc 2012;19:121-7.
69. U.S. Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence/Machine Learning (AI/ML)-Enabled Device Software Functions: Draft Guidance for Industry and Food and Drug Administration Staff. April 2023. Available online: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing