Artificial intelligence in cardiology in the current era: a narrative review

Introduction

Background

Cardiovascular disease remains a leading cause of morbidity and mortality worldwide, accounting for an estimated 17.9 million deaths annually, or about 31% of all global deaths (1). The increasing volume of cardiovascular encounters generates large repositories of structured and unstructured data within electronic health records (EHRs), imaging archives, and wearable/device streams, enabling data-driven model development and validation. In recent years, artificial intelligence (AI) has emerged as a transformative tool in cardiology, capable of analyzing vast amounts of data with speed and precision, often achieving similar rates of diagnosis as board‐certified cardiologists (2). AI algorithms, especially those based on machine learning (ML), can uncover subtle patterns in echocardiograms, electrocardiograms (ECGs), and radiographic images that might elude human experts.

In our manuscript, AI is an umbrella term describing computational systems that perform tasks typically requiring human intelligence, including perception, pattern recognition, prediction, and language understanding. ML is a subset of AI in which algorithms learn statistical patterns from data to make predictions or classifications without being explicitly rule-programmed. Deep learning (DL) is a subset of ML that uses multi-layer neural networks [e.g., convolutional neural networks (CNNs) and transformer architectures] to learn hierarchical representations directly from raw inputs such as waveforms, images, and text. In cardiovascular medicine, most “traditional” AI applications have been discriminative (e.g., detecting disease or predicting outcomes), whereas the current era increasingly includes generative and foundation-model approaches that can synthesize text or structured outputs, integrate multimodal inputs, and support clinical reasoning tasks, introducing new benefits and new safety risks.

By leveraging AI, cardiology benefits from earlier disease detection, more accurate diagnoses, and personalized treatment strategies that were previously unattainable with conventional methods. It is also important to distinguish between automation and AI. Automation refers to technologies that perform tasks based on explicit, pre-programmed rules, such as robotic systems or scripted workflows. AI, particularly ML and DL, describes systems that learn from data, adapt to new scenarios, and make predictions or decisions beyond fixed rules. Throughout this review, we will clarify which technologies are rule-based automation and which are true Al- driven interventions.

Rationale and knowledge gap

One of the most desirable benefits of using AI is that it decreases workload and increases efficiency in a fast-paced clinical environment. Advanced algorithms allow for automated execution of complex procedures, including magnetic resonance imaging (MRI)-based ejection fraction (EF) and volumetric measurements, as well as real-time hospital telemetry data screening. A recent meta-analysis demonstrated that AI assistance in medical image interpretation reduces reading time by 27%, according to the study (3). Integrating AI predictive models with various data sources, such as physical exam results, auscultation findings, multimodal imaging, lab results, and genomic information, improves patient risk stratification over current risk scores. Essentially, AI offers an efficient solution for cardiology to handle increasing healthcare data while supporting precision medicine initiatives.

Despite the enthusiasm, it is essential to acknowledge that most AI tools in cardiology are still in early stages of clinical adoption (4). The deployment of such methods still faces significant challenges because of inadequate data quality and biased algorithms, as well as the process of getting regulatory approval for use in the real world. Successful translation requires multidisciplinary collaboration among clinical cardiologists, data scientists, and engineers.

Prior reviews have summarized broad applications of AI across cardiology, including electrocardiography, imaging, and EHR-based prediction, and have emphasized recurring limitations such as dataset shift, limited external validation, and uncertain clinical impact (5-8). In contrast, our review centers on practical clinical translation: explicitly distinguishing automation from AI, mapping evidence tiers (internal vs. multisite validation and retrospective performance vs. prospective utility), and highlighting workflow-facing paradigms (generative AI, foundation and multimodal models, privacy-preserving learning, and evolving regulatory expectations) that are likely to determine real-world adoption.

For clarity, this review distinguishes between rule-based automation (deterministic “if-then” logic and traditional decision rules) and data-driven AI, including ML and DL models that learn patterns from data. These categories differ in how they are developed, validated, monitored, and regulated; we use “automation” for rule-based tools and “AI” for data-trained models unless otherwise specified.

Objective

The current literature on AI implementation continues to expand rapidly. The available evidence from real-world applications is limited, and external validation results show considerable variability. Our narrative review examines how AI technology functions in cardiology through its uses in multimodality imaging, ECG/waveforms, EHR risk modeling, and remote monitoring devices. We present this article in accordance with the Narrative Review reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-510/rc).

Methods

The authors analyzed AI applications in cardiology through five categories which included multimodality imaging techniques [cardiac computed tomography (CT), cardiac magnetic resonance (CMR), nuclear imaging], ECG/telemetry waveform analysis, and various risk prediction models from EHR data, wearable devices, coronary angiography with percutaneous coronary intervention (PCI) procedures and robotic systems.

The research team conducted database searches through MEDLINE/PubMed, Web of Science, Scopus, EMBASE for studies published between January 1, 2015, and December 21, 2025. The search strategy used both controlled vocabulary terms and free-text words to find studies on AI and ML/DL in cardiology fields [ECG, echocardiography, CMR, coronary computed tomography angiography (CCTA), fractional flow reserve computed tomography (FFRct)/quantitative flow ratio (QFR), and robotics] and their related outcomes (diagnosis, prognosis, and workflow). The researchers performed manual reference list searches for key reviews, major society statements [American College of Cardiology (ACC)/American Heart Association (AHA)/European Society of Cardiology (ESC)/Society for Cardiovascular Magnetic Resonance (SCMR)/Society of Cardiovascular Computed Tomography (SCCT)], and relevant regulatory communications (Table 1). Study selection is summarized in a PRISMA-style flow diagram (Figure 1).

Figure 1 PRISMA flow diagram of the literature search and study selection. Records were identified through database searching and supplemental hand-searching. After deduplication, titles/abstracts were screened and full texts assessed for relevance to cardiology-focused automation and AI applications. Studies were included in the narrative synthesis if they addressed clinically relevant cardiovascular use cases with sufficient methodological detail to support interpretation. AI, artificial intelligence.

The study involved human cardiology research that used AI, ML, and DL methods, including prospective and retrospective cohort designs, randomized and non-randomized trials, reader studies, implementation, decision-impact research, consensus, and guidance documents. It excluded reports with fewer than 50 participants unless they provided significant new information, as well as studies that relied solely on simulations, focused on technical aspects without clinical value, or involved animal or in vitro research. Non-English language content was also excluded.

Two researchers (A.M. and B.B.) independently evaluated titles, abstracts, and full texts before a third reviewer (H.T.) settled any disagreements. The researchers used a standardized form to extract information about study settings and participant demographics, data sources, reference standards, model types, validation approaches, performance metrics, workflow data, clinical results, fairness assessments, and funding sources.

We did not perform a formal risk-of-bias scoring or checklist-based reporting-quality appraisal, consistent with the scope of a structured narrative review. Instead, during extraction we recorded a focused set of methodological characteristics relevant to clinical translation, including study design, data source and reference standard, and whether studies evaluated workflow metrics or patient outcomes. We also noted recurrent limitations that affect real-world reliability, including data leakage, spectrum bias, missing data handling, limited external validation, and absent calibration or subgroup reporting. Ethics approval was not required because no individual-level patient data were used.

Clinical applications of AI in cardiology

In this section, we will explore in detail how AI is transforming various fields within cardiology, examining the diverse ways in which AI and ML are being applied to imaging, diagnostics, risk modeling, and patient monitoring. Table 2 summarizes the difference between rule-based automation and AI/ML approaches across cardiology domains.

AI-assisted auscultation

Auscultation is one of the oldest, most commonly used, and highly respected clinical skills in medical diagnosis, especially in cardiovascular medicine. However, despite being a fundamental part of clinical management, it is highly subject to inter-user variability and poor reproducibility (9). A study evaluating 400 residents in training from internal medicine and family practice revealed a disturbingly low identification rate for the most important and commonly encountered cardiac events (9). ML and DL techniques, such as CNN and classifier-trained methods like AdaBoost, are now being used to assist with auscultation, helping to reduce errors and improve accuracy and consistency. First, the stethoscope records the sounds, then undergoes preprocessing and segments sections of the cardiac cycle. It then uses feature extraction and combines classifier training techniques and CNNs to distinguish between normal and abnormal heart sounds (10). Thompson et al. showed that an AI-based murmur detection system achieved a sensitivity of 87% and specificity of 89% detecting heart murmurs almost as accurately as cardiologists, based on results confirmed by the echocardiograms, highlighting its use as a triage tool (11).

However, there are still some challenges. One disadvantage of AI systems stems from their need for a vast amount of audio data that spans multiple patient demographics for proper operation. The current datasets contain insufficient diversity regarding patient demographics and environmental settings, which restricts the models’ ability to generalize their results. The diagnostic performance of AI tools shows significant differences because researchers used different validation approaches and external validation methods according to a systematic review and meta-analysis (89).

Evidence quality and limitations: most AI-auscultation evidence remains derived from retrospective datasets with limited standardization of recording hardware and labeling, which can reduce generalizability across environments and patient groups. Prospective workflow studies and multi-site external validation are still limited; therefore, reported accuracy may not translate directly to bedside performance. Future work should prioritize diverse external validation, calibration, and integration with complementary signals (ECG/echo) to improve clinical actionability (8-10).

AI in ECG

DL has transformed the ECG from a traditional, rule-based diagnostic tool into a comprehensive predictive biosignature. Early landmark studies showed that arrhythmia detection through single-lead ambulatory monitoring met cardiologist standards and revealed underlying signs of left ventricular systolic dysfunction (LVSD) that standard 12-lead ECGs failed to identify (12,90). ECG analysis now extends beyond simple disease detection because advanced models using sinus-rhythm 12-lead ECGs can predict atrial fibrillation (AF) before it occurs, offering better predictive results than existing clinical risk scores, and uncovering hidden electrophysiological patterns in waveforms (13,91).

The EAGLE trial used a cluster-randomized design to show that AI-ECG alerts in primary care settings resulted in higher LVSD detection rates during the first 90 days while reducing the number of echocardiography tests (14). Furthermore, the market expansion of consumer wearable technology has made cardiac monitoring devices accessible to all consumers. A prospective study demonstrated that smartwatch single-lead ECG recordings analyzed through an adjusted algorithm successfully identified EF ≤40% in non-clinical environments, which suggests their potential use for heart failure detection and opening venues for early preventative strategies (92).

Beyond offering insights into cardiac structure and anatomy, ECG can also act as a sensitive marker of metabolic abnormalities, particularly electrolyte imbalances like dyskalemias and systemic conditions such as anemia. This broadens the traditional role of ECGs from detecting cardiac issues to encompassing wider clinical monitoring, utilizing subtle waveform changes that indicate underlying physiological disturbances (93,94).

The methodological approach of CNNs continues to serve as the main framework for complete end-to-end analysis of unprocessed multi-lead EKGs because they show consistent performance in current comparative studies (95). Research has focused on transformer-based and hybrid CNN-transformer models to enhance the detection of extended temporal connections and rhythm changes. Applying unsupervised learning techniques to large unlabeled ECG datasets improves data efficiency and learning, which boosts the performance of subsequent tasks by predicting rhythm and clinical outcomes simultaneously (15).

Community challenges like the PhysioNet/Computing in Cardiology Challenges have been essential for standardizing how we evaluate ECG analysis algorithms. They offer large, multi-institutional datasets and open competitions that encourage reproducibility and model generalizability by requiring public code submissions (e.g., on GitHub) and testing on hidden external data. The field now concentrates on creating models that function across diverse populations and data sources because AUC-ROC results from one study do not guarantee accurate performance on different external datasets (96).

Clinical integration typically follows three models: (I) real-time triage in emergency or prehospital care [e.g., automated ST-elevation myocardial infarction (STEMI)/occlusion myocardial infarction (OMI) alerts from ambulance ECGs]; (II) batch screening of archived ECGs in primary care to identify risk for LVSD or future arrhythmias; (III) remote monitoring through wearables or patches for scalable detection and follow-up. Prehospital studies demonstrate the feasibility and high diagnostic accuracy of automated STEMI detection, with the potential to reduce treatment delays (16). International studies indicate that AIECG can detect occlusion MI across a wide spectrum of acute coronary syndrome (ACS) comprehensively than the STEMI criteria and with performance comparable to expert interpretation (97).

The deployment of this system faces restrictions because its performance becomes less consistent when using different vendors, sampling rates, and patient groups, thus requiring extensive multi-site external validation, continuous drift monitoring, and periodic system recalibration (96). Labeling strategies, such as echo-derived EF within a time window, can introduce noise and bias, affecting achievable performance. Safe and equitable integration requires clinician‑in‑the‑loop review, predefined action pathways, uncertainty‑aware thresholds, secure logging, and privacypreserving governance. Attention to equity is essential, as differential performance by sex, age, or race, and unequal access to connected devices, could exacerbate the digital divide unless explicitly addressed in development and trials (17,95).

Evidence quality and limitations: although many ECG-AI studies report strong discrimination in retrospective cohorts, generalizability depends on device vendor, lead configuration, sampling frequency, and population mix, and performance should be reported with calibration and clinically interpretable measures beyond AUC. The EAGLE trial provides higher-level evidence that ECG-AI can improve screening yield in pragmatic implementation, but broader prospective evaluations and monitoring for bias, alert fatigue, and workflow impact remain needed (12,16,90,91,93,95,96).

AI in cardiac imaging

Echocardiography

The echocardiography workflow has experienced rapid development through AI advancements, which now improve image acquisition, view identification, quality assurance, quantitative analysis, and disease detection processes. Developing video-based DL systems allows for left ventricular (LV) segmentation and EF estimation with each heartbeat. This new technology decreases inconsistencies in clinical readers, which has been possible because of extensive public databases, enhancing reliability (18). These achievements provide a solid foundation for end-to-end interpretation in routine clinical practice, accelerating the integration of automation into everyday echocardiography (19).

Real-time AI guidance enables early clinicians to obtain transthoracic echocardiograms with diagnostic quality. In a multicenter prospective study, nurses without prior ultrasound training successfully performed 10-view limited exams, accurately evaluating left ventricular size and function, right ventricular size, and pericardial effusion in more than 90–98% of patients (20). The first AI-based assessment of EF in a blinded randomized noninferiority trial achieved the same accuracy as experienced sonographers while reducing interpretation time compared to cardiologists. This shows AI can potentially improve cardiac measurement procedures in standard medical practice (21).

Automated view detection and chamber measurement may use rule-based algorithms, while DL models enable adaptive analysis and disease detection. A successful acquisition automation relies on image recognition accuracy, which has been shown to reach a 95% success rate when analyzing standard adult views from large, diverse datasets. Training advanced AI systems with tens of thousands of studies enables them to recognize over 20 different views and detect both off-axis and mislabeled images, while prompting users to reacquire suboptimal scans. These capabilities make the laboratory workflow more efficient by reducing non-diagnostic studies and improving overall exam quality (19,22). The use of DL techniques enables continuous segmentation of the left ventricular cavity and measurement of EF and volumes from apical views, with results that match or surpass human observer accuracy (18).

The AI pipelines developed by commercial organizations now perform left atrial volume measurements, right-sided chamber size assessments, and Doppler envelope analysis for left ventricular outflow tract (LVOT) velocity time integral (VTI) and mitral inflow, along with essential diastolic function parameters. The multicenter validation study of 23 parameters, including volumes, EF, and Doppler metrics, showed that the AI workflow produced results closer to core-laboratory expert readings than human expert discrepancies, supporting AI as a substitute for many quantitative assessments (23). AI models use standard echocardiographic clips to identify complex cardiac conditions, including hypertrophic and infiltrative cardiomyopathies, significant valvular diseases, pericardial effusions, and pulmonary hypertension risk (24).

The AI system analyzed a single apical four-chamber video to distinguish between HFpEF vs. non-HFpEF cases accurately. The research demonstrates that echocardiographic video phenotyping can be a powerful diagnostic tool for complex syndromic conditions requiring multiple parameter assessments (24). The vision language foundation model EchoCLIP achieved zero-shot generalization across various functions, including device identification, by training on more than one million echocardiographic videos and clinical reports (25). This system achieved retrieval and report-aware reasoning abilities, enabling the development of advanced AI assistants that generate structured diagnostic reports and access previous case data examples (25).

Evidence quality and limitations: AI performance can be affected by label noise, often due to inter-reader variability when measuring EF or strain, which highlights the need for quality filters and confidence thresholds. Establishing effective governance requires clinicians to review AI results, implement clear protocols for AI-human assessment and disagreements, conduct bias evaluations for sex, age, BMI, and other clinical parameters, and maintain model version control and secure data management systems. Deploying AI in point-of-care and remote settings demands representative training data and equal device and connectivity access to determine whether AI will improve or hinder healthcare delivery inequalities. Prospective studies demonstrating improved diagnostic accuracy, reduced time-to-diagnosis, or better outcomes remain fewer than retrospective performance reports, highlighting the need for multi-center, outcome-driven evaluations and transparent reporting of failure modes (17-23,97).

AI in nuclear cardiology

Nuclear cardiology is a subspecialty of cardiovascular imaging that uses non-invasive techniques such as SPECT and PET to evaluate the viability of myocardium, assess myocardial perfusion, and determine ventricular function (26). With the growth of DL in nuclear imaging within cardiology, this field will benefit from improvements in image acquisition and timing, attenuation correction, diagnostic advantages in coronary artery disease (CAD) through predictive modeling, and workflow enhancements.

Radiation exposure is a concern for myocardial perfusion imaging (MPI). AI-enhanced imaging using DL methods in MPI-single-photon emission computed tomography (MPI-SPECT) showed that reconstructed AI images were able to reduce dose time and acquire images faster without compromising imaging quality (27). The study was notable for a nearly 25–50% reduction in the time of the MPI-SPECT (27). Similar studies have been completed, such as Ramon et al., using up to one-eighth the dose of radiation (28). This is beneficial for reducing radiation exposure during testing and allows for more testing to be completed.

AI in image attenuation will play a vital role in SPECT imaging. Various facilities lack CT scans, where the approach by Yang et al. used AI to attenuate images using DL methods (29). CT scans are used for attenuation correction for soft tissue artifact in cardiac single-photon emission computed tomography (SPECT). Images obtained using the AI-enhanced features without CT scans will reduce radiation exposure and allow facilities without CT to have attenuation correction. This method’s downside was that female patients with larger body mass index (BMI) created bias; thus, feature AI-enhanced imaging will need more data for their DL network (29). In another study by Sohlberg et al., investigators used DL models to decrease noise artifact with several AI models. Their study summarized that all the DL models outperformed standard technique (30).

AI-generated DL model in nuclear cardiology has also been used to assess CAD and has shown the ability to predict disease better than current models. In Betancur et al., investigators used invasive coronary angiography within 6 months of SPECT to compare standard methods to the DL AI (31). This study found that AI outperformed standard methods on a per-patient and per-coronary vessel basis, and the per-vessel prediction improved sensitivity from 64.4% to 69.8% (31). Another study used a ML AI method to predict early coronary revascularization following MPI (32). This study showed that the AI ML model had the highest area under the curve (AUC) (0.812) versus imaging only (0.786) versus imaging plus clinician variables (0.799); thus, ML had the highest prediction in early coronary revascularization, after incorporation of patient clinical and MPI data (32). This can be helpful for the future by finding early lesions and possibly saving patients from unnecessary procedures.

One of the most significant benefits to AI in nuclear cardiology is workflow optimization. One optimization is screening patients who should undergo studies, saving patients from undergoing invasive angiography (32). Another study, Apostolopoulos et al., shows AI models can use other patient diagnostic data, such as echocardiogram, patient risk factors, along with MPI, to detect disease on the perfusion scan better (33). Like other imaging modalities, AI models in nuclear cardiology have drawbacks. Algorithms may be vendor or site-specific, and some models may not be generalizable to assess all patients. Few tools have been cleared to be used by the FDA thus far; however, it’s reasonable to expect this to start advancing.

Evidence quality and limitations: nuclear cardiology AI evidence is dominated by retrospective imaging studies focused on attenuation correction and perfusion interpretation, with relatively fewer prospective evaluations of clinical utility. Generalizability may be limited by protocol heterogeneity (tracers, cameras, reconstruction settings) and variable ground-truth definitions; therefore, multi-site validation and clinically meaningful endpoints (beyond AUC) are important for translation (24-30).

CMR imaging

AI has also become essential to the CMR pipeline, from planning and acquisition to reconstruction, segmentation, quantification, and end-to-end diagnostic support. Initial advancements showed that using CNN allowed for quick and accurate segmentation of ventricles and atria on a large scale, such as in the UK Biobank study, achieving performance levels comparable to those of clinicians for routine chamber measurements, which are usually very time-consuming (34). Various quality control (QC) frameworks have also identified segmentations, in the process, before clinicians review them (35,36).

Vendor and independent studies have shown that DL-based cine reconstruction can more than halve acquisition time while maintaining image quality and with accurate volumetrics, advancing toward free-breathing, arrhythmia-corrected protocols and higher throughput (37).

In this prospective study, three cine CMR protocols—standard segmented cine, single-breath-hold DL-accelerated cine, and free-breathing motion-corrected DL cine—were implemented in patients undergoing 3T CMR. DL-accelerated protocols reduced total acquisition times by 73% and 62% compared to segmented cine, while preserving image quality and cardiac function assessment (98). In addition to releases of this technology in the market, available resources like CMRxRecon for k-space have played a role in standardizing evaluations and speeding up method development across cine and mapping techniques, resulting in enhanced transparency and comparability within the field (99).

The implementation of AI technology in medical imaging has automated time-consuming post-processing operations, strain analysis, and tissue mapping functions, which decreases inter-observer variability (34,36,100). The DL pipeline integrated into perfusion imaging enables automatic myocardium segmentation and quantitative blood-flow mapping of the myocardium during scanning through a single button operation, which both reduces analysis duration and creates uniform results between different imaging facilities (101). The automated quantitative perfusion results generated by these systems have proven useful as they show excellent predictive value for patient such as death [hazard ratio (HR) 1.93, 95% confidence interval (CI): 1.08–3.48 and HR 2.45, 95% CI: 1.42–4.24] and major adverse cardiovascular events (MACEs) (HR 2.14, 95% CI: 1.58–2.90 and HR 1.74, 95% CI: 1.36–2.22) using stress myocardial blood flow (MBF) and myocardial perfusion reserve (MPR) respectively (102).

The implementation of AI technology has improved the robustness of LGE scar segmentation and radiomics risk marker analysis through automated scar burden detection, which multiple studies have validated for outcome prediction (38,103). The VNE models, which combine cine/T1 maps to generate late gadolinium enhancement (LGE)-like data suggest potential applications for reducing gadolinium usage in particular imaging scenarios but need further prospective studies to confirm clinical equivalence (39).

The diagnostic capabilities of CMR extend beyond its traditional applications. A 2024 multicenter research study published in Nature Medicine (n=9,719) tested a two-stage video-based swin transformer (VS.T) system, which first detects cine abnormalities before using cine and LGE data to identify 11 CVDs with performance levels surpassing human cardiologists in specific tasks (AUC of 0.988±0.3% and 0.991±0.0%, in internal and external cohorts, respectively). The system demonstrates potential for complete CMR decision support integration, but needs further testing before deployment (40).

Future development of these techniques requires extensive testing across multiple sites to validate their performance, along with decision curve reporting and prospective studies that assess diagnostic effects, treatment choices, patient results, and ongoing QC and drift monitoring for safe and fair deployment at large scales (39).

Evidence quality and limitations: CMR AI has advanced rapidly in segmentation and reconstruction, and recent studies suggest that DL-accelerated cine protocols can shorten acquisition time while preserving functional assessment. Nevertheless, performance can vary by scanner vendor, field strength, and sequence parameters, and artifact sensitivity requires robust quality assurance and human oversight. Further multi-center validation and prospective demonstrations of workflow or outcome benefit remain key gaps (31-36).

AI in CCTA

Automated coronary artery calcium (CAC) scoring, detection, and plaque characterization and quantification of stenosis, noninvasive functional assessment like FFRct, and perfusion analysis are a few of the key applications in AI. AI can eliminate tedious manual segmentation of coronary arteries that serve as risk stratification tools for assessing future cardiovascular events. An early study from 2007 on ECG-gated cardiac CT achieved 73.8% sensitivity for detecting coronary calcifications and correctly stratified 93.4% of patients into the proper risk category, demonstrating the feasibility of fully automated CAC scoring (41). Recent DL models report near-perfect alignment with standard CAC scoring. Studies show excellent correlation (Spearman’s ρ ~0.93–0.99) and agreement [intraclass correlation coefficient (ICC) ~0.99] between AI and expert-derived scores, with approximately 89–98% concordance in assigning patients to the same risk categories (42).

AI tools have been developed to assist in quantifying plaque and classifying stenoses, as well as automated Coronary Artery Disease Reporting and Data System (CAD-RADS) scoring, while reducing interobserver variability compared to gold standard invasive measures. With the new dual energy CCTA studies, AI has been able to analyze scans more efficiently. In a sub-group of 232 patients of the CLARIFY trial, expert readers had only moderate agreement in total plaque volume (ρ ~0.7) and poor agreement in plaque composition (ρ ~0.4–0.6) compared to an AI-based quantitative analysis (43). DL models have performed automatic CAD-RADS risk category assessments from CCTA images. In a study by Sandstedt et al., a CNN model achieved a moderate level of diagnostic accuracy with sensitivity ranging from 47% to 82% and specificity ranging from 58% to 91% while providing faster results than clinicians (44). The subsequent models developed through time reached 93% sensitivity and 97% specificity [negative predictive value (NPV) 97%] for detecting significant disease while achieving expert-level performance. The ML model demonstrated 98% accuracy in CAD-RADS category classification, which matched the performance of board-certified imagers according to two separate studies, and has been validated against invasive standards (45,46). For example, an AI-generated contrast density metric demonstrated 75% specificity and 73% NPV when identifying hemodynamically significant lesions when compared to invasive FFR testing (47). The ML model in the CREDENCE trial substudy achieved 84–86% accuracy and showed strong correlation (ICC =0.73) with quantitative coronary angiography while reducing reading time to approximately 10 minutes, compared to hours (48). However, it produced 7.3% of false positive results for severe stenosis detection, which warrants careful evaluation of the results.

Machine-learning approaches have significantly enhanced the diagnostic utility of FFRct and myocardial perfusion analysis on CT. An ML-based FFRct algorithm surpassed CCTA anatomy-only assessment to achieve 78% per-vessel accuracy, higher than the 58% obtained through CCTA alone. The ML-based FFRct system demonstrated high sensitivity between 79–90% and specificity between 94–95% for detecting flow-limiting stenoses while producing results that matched both invasive FFR (r 0.73) and traditional computational FFRct (AUC 0.84) (49,50). In a prospective study, ML-based FFRct reduced the need for angiography as 19.8% of patients with FFRct guidance did not require angiograms compared to 33.3% with standard CCTA, resulting in fewer adverse cardiac events during the first year (51).

Evidence quality and limitations: many CCTA-AI tools are supported by large retrospective datasets and show promise for stenosis and plaque assessment, but performance is vulnerable to motion, noise, and heavy calcification, and may shift across scanners and reconstruction kernels. To support clinical adoption, studies should report calibration and clinically actionable endpoints, and emphasize external validation across diverse practice settings (38-46,103).

AI in wearables & remote monitoring

Wearable technology systems now track patient health data through continuous monitoring of ECG, PPG signals, accelerometry, and sleep metrics, which expands the scope of medical surveillance compared to scheduled outpatient visits (52-57). The application of ML models, particularly CNNs and RNNs has significantly enhanced disease-specific signal detection from real-world datasets. This finds applications in treating multiple medical conditions, which include cardiovascular disease (CVD), sleep apnea, diabetes, mental health disorders, and rehabilitation (52-57). The evidence supporting the broad clinical impact of these technologies in clinical practice is still developing, since researchers need to conduct extended studies and establish systematic integration methods for healthcare systems (52,53,57,58).

The signal processing workflow begins with preprocessing and denoising operations, followed by DL architecture-based feature extraction through CNNs, RNNs, and transformers (59,60). The predictive accuracy of multi-sensor data improves through temporal fusion methods, while researchers now focus on developing methods to quantify uncertainty in their results. The deployment of AI systems follows two main strategies, which include on-device and far-edge inference for fast processing and privacy protection, and cloud-based aggregation for extensive population modelling and model improvement (55,58). The current AI field depends on supervised DL for most applications, but researchers now focus on developing federated learning and explainable AI systems to handle data sharing problems and improve model interpretability and user trust (53,59).

A systematic review of 26 studies demonstrated that AI models used with wearables achieved a combined sensitivity of 94.8% and specificity of 97.0% in their analysis. The performance of deep neural networks surpassed conventional ML models in terms of area under the receiver operating characteristic curve (AUROC) values at 0.981 compared to 0.961. The performance results showed better outcomes when using public and hospital-based datasets instead of proprietary and wearable-based datasets, indicating the need to address dataset shift problems for practical deployment (52). Detecting arrhythmias, sleep apnea, and stress through wearable ECG/PPG data processing now relies on AI systems that use DL models instead of traditional ML approaches. The use of ML for CVD prediction and diagnosis outperforms statistical models, but real-world testing and device precision issues act as limitations for practical applications (60,61). The integration of Edge AI systems allows for real-time analytics that detect decompensation risks and operate independently from constant cloud connections, but requires architectural assessment before deployment to prove additional value and match clinical operational procedures (55).

The verification of accurate data requires sensor error and characterization of motion artifact for device accuracy assessment and model training with diverse calibrated datasets to reduce bias levels (56,59). The implementation of privacy and security frameworks requires organizations to establish data ownership systems and encryption protocols, and obtain consent from users while providing equal access to data resources (54,55). The development of models for better generalization requires researchers to use multicentre, diverse cohorts instead of proprietary datasets for validation and training (52,53,56).

Implementing AI in wearable devices enables healthcare providers to detect diseases earlier and monitor treatment effects objectively while sending specific alerts for preventive care instead of reactive treatment. The successful deployment of wearable-based AI monitoring depends on team-based operational systems, predefined alert limits, medical staff supervision, and EHR system integration for maintaining clinical value and minimizing alert fatigue (54-56,59).

Evidence quality and limitations: wearable AI studies benefit from scale and real-world data streams, but clinical confirmation of detected events and downstream impact are variable across products and populations. False positives can drive unnecessary testing and anxiety, and performance may differ by skin tone, age, comorbidity, and signal quality. Pragmatic evaluations that measure decision impact and equity-focused validation are important for responsible deployment (47-56).

Heart failure

AI can now monitor the entire spectrum of heart failure progression, from detecting LV systolic dysfunction in asymptomatic patients to predicting decompensation and readmissions. One of the most successful uses of AI-ECG technology is its ability to detect LVSD with a 12-lead CNN, which achieved an AUC of 0.93 with 86% sensitivity and 86% specificity for identifying EF ≤35%. It also identified patients who are likely to develop future LV dysfunction at a fourfold higher rate as false positives. Implementing this AI-ECG model improved the detection of new LVSD cases without causing unnecessary tests or follow-ups, demonstrating its practical value in real-world settings (62).

The application of AI technology extends its capabilities to diastolic heart function assessment and systolic function assessment. A ResNet-based AI-ECG system achieved AUC 0.911 (sensitivity 83%, specificity 83%) for predicting elevated LV filling pressure. It demonstrated similar mortality risk stratification performance to echocardiogram in a patient population of approximately 100,000, which makes it possible to use a 10-second ECG as a scalable screening tool for heart failure with preserved ejection fraction (HFpEF) diagnostic pathways (63). Combining structured EHR data with gradient-boosted trees and ensemble algorithms enables risk prediction across entire populations. An extensive study involving multiple hospitals demonstrated that ML models that integrated social determinants of health data achieved C-statistics between 0.79 and 0.80 for in-hospital mortality prediction across different racial groups while outperforming traditional risk scores and providing explainable feature contributions for clinical staff acceptance (64).

Preventing hospital readmissions relies on identifying signs of decompensation early. Supervised time-series ML analysis of wearable multi-sensor data, including ECG, accelerometry, and temperature measurements, enables the prediction of heart failure events before they happen. The LINK-HF study showed that its personalized analytics system could forecast worsening heart failure 7–10 days in advance, with prediction accuracy reaching 0.84–0.89 AUC, sensitivity between 0.76–0.88, and specificity between 0.85–0.86 through continuous signal processing for low-latency alert generation (65).

Evidence quality and limitations: heart-failure AI models often draw on observational EHR data and can be affected by confounding, label leakage, and documentation variability, which can inflate retrospective performance. External validation across institutions and prospective testing of workflow integration and clinical outcomes remain inconsistent, emphasizing the need for transparent reporting and calibration assessment when models are used for risk-guided care (57-60).

Cardiac arrhythmia

Implementing AI in cardiac arrhythmia has led to improved screening methods, predictive tools, monitoring antiarrhythmic drug safety, and analyzing continuous telemetry. The CNN developed by Mayo Clinic used a sinus-rhythm 12-lead ECG to detect latent AF substrate with an AUROC of 0.87 for a single ECG, which reached 0.90 when analyzing multiple ECGs, proving the method can identify cases before they occur (66). A DL model from a single research center demonstrated the ability to identify two-thirds of patients without AF history who developed stroke due to AF before the stroke occurred (13). A 2023 deep-learning classifier achieved 84.1% sensitivity, 89.3% specificity, 55.2% positive predictive value, and 97.3% negative predictive value when detecting AF events from single-lead signals, making it suitable for triage and telemetry applications (67).

The “QTNet” model was trained from more than 7 million ECGs to determine QT intervals and detect QTc prolongation after initiation with dofetilide with 87% sensitivity, 77% specificity, and more than 95% negative predictive value when the pre-test probability is below 25% thus enabling safe outpatient and loading surveillance with wearable devices (68). This will be effective in reducing hospital stays and costs. A multicenter JAMA Cardiology study demonstrated that the ECG-only model achieved 0.93 external AUC with 0.90 sensitivity for detecting (LQTS) cases, which helps with genotype-based treatment decisions and family screening (69). An ECG-based model for sudden cardiac death (SCD) risk assessment achieved 0.889 AUROC in independent testing while providing noninvasive substrate information beyond left ventricular ejection fraction (LVEF) measurements. The “DEEP-RISK” model which combines ECG data with LGE-CMR results and clinical information achieved an AUROC of 0.84 for predicting malignant ventricular tachycardia/ventricular fibrillation (VT/VF) with 0.98 sensitivity and 0.73 specificity while outperforming individual testing methods for implantable cardioverter-defibrillator (ICD) selection (70).

AI can identify AF phenotypes, drug-induced QT risks, and VT/VF occurrences while producing useful results for triage decisions and showing measurable benefits in telemetry systems. Future research needs to focus on conducting prospective trials integrating AI into clinical workflows to determine action thresholds, decision curves, and bias assessments for sex, age groups, and device types, and establish post-deployment monitoring systems to maintain reliable performance.

Evidence quality and limitations: arrhythmia AI has strong retrospective evidence in large ECG datasets, including models for QT risk and AF phenotyping; however, prospective decision-impact evidence is limited, and model performance can vary across device types and recording conditions. Clear action thresholds, interpretability, and monitoring for subgroup performance are needed to reduce unintended harm and improve clinician trust (61-66).

AI in angiography and PCI/interventional procedures, including robotic guided procedures

In the last decade, multiple studies have helped validate the use of AI in coronary angiography. FLASH trial aimed to evaluate the efficacy of AI-QCA (artificial intelligence-based coronary angiography)-assisted PCI compared to optical coherence tomography (OCT)-guided PCI in terms of post-PCI results (71). This was an investigator-initiated, multicenter, randomized, prospective trial conducted at 13 sites in South Korea (n=400), who received either AI or OCT-guided PCI. The primary endpoint was the post-PCI minimal stent area and concluded noninferiority of AI-QCA with a margin of 0.8 mm². The AI-ENCODE study used a library of more than 20,000 angiograms performed at Mayo Clinic, developed and validated AI algorithms to extract left ventricular EF, diastolic dysfunction, right ventricular dysfunction and cardiac index from 1–2 angiographic views thus validating the ability of AI to broaden diagnostic scope of coronary angiography by automatically extracting physiologic and functional data which can positively impact patient outcomes (72). While robotic PCl systems automate procedural steps through pre- programmed instructions, true Al-driven interventions in interventional cardiology involve ML models that analyze imaging data or predict outcomes to guide clinical decisions.

The combination of robotic systems and AI technologies has transformed the field of interventional cardiology, beginning with guidewire use to improve precision, decrease radiation exposure, and standardize treatment protocols. The coronary robotics market features two leading platforms: Corindus (Siemens Healthineers) CorPath GRX and Robocath R-OneTM. Current research shows that robotic PCI (R-PCI) achieves high success rates in both technical and clinical outcomes, provides significant radiation protection for operators, and maintains procedural speed. However, the superiority of hard clinical endpoints over manual PCI has yet to be proven (73,74).

The R-One (R-EVOLUTION) and GRX registries from multicenter studies demonstrate the feasibility and safety of robotic PCI while showing substantial reductions in operator radiation exposure, which aligns with the expected benefits of cockpit-based control (73). The PARTY trial is a single-center, randomized, prospective clinical trial comparing R-PCI to conventional PCI, which aims to provide answers towards the potential benefits of R-PCI (75). The first human telerobotic PCI procedure from a distant location succeeded in 2019 when an operator located 20 miles from the cath lab controlled coronary devices through GRX. The procedure established a crucial precedent for remote interventions and disaster response, yet exposed operational challenges related to network stability, operator certification, and system delays (76). The deployment of robotic systems on a large scale for patient treatment remains untested and requires validation for widespread use.

Intravascular imaging has experienced rapid AI development, enabling users to pick optimal frames and to help characterize tissue and perform automated measurements of essential structures. The commercial software Abbott’s Ultreon applies ML to OCT imaging for calcium detection, EEL/vessel size estimation, and malapposition/under-expansion identification, which results in more precise, faster, and less variable PCI planning according to comparative studies (77,78).

Angiography-derived physiology, such as QFR, blends a three-dimensional vascular model of the coronary artery from two angiographic views. Then, it uses computational fluid dynamics to simulate blood flow and calculate the pressure gradient to determine the FFR. In FAVOR III China, QFR guidance improved 1-year and sustained 2-year outcomes versus angiography alone (79). But in FAVOR III Europe, QFR was not a match for FFR when a wire-based gold standard was readily available (80). QFR is a credible option where wire physiology is complex or underused, but it does not supplant FFR in settings where FFR is feasible.

OFR technology is a wire-free FFR technology that uses the high resolution and accuracy of OCT images, and reconstructs three-dimensional vascular models through AI technology, simulating the calculation of FFR indicators (81). This provides detailed coronary microstructure imaging through OCT, enhancing diagnostic precision. The system delivers fast and efficient patient care through its non-invasive procedure, which requires only multiple cross-sectional images for hemodynamic assessment. The clinical trials have proven OFR’s reliability through validation, while showing strong agreement with traditional wire-based methods, according to studies (61,82,83).

Evidence quality and limitations: evidence in interventional cardiology spans both ML-driven image analysis and robotics, but it is important to distinguish rule-based automation from adaptive ML guidance. Research studies must evaluate robotic PCI performance against manual PCI results while assessing how AI-based imaging affects stent failure rates. The effectiveness of QFR needs to be assessed in medical facilities that do not implement FFR technology. The safety and reliability of robotic procedures with tele-robotics require additional clinical trials for validation. Current data include registries and early prospective trials; broader evidence on long-term outcomes, operator learning curves, cost-effectiveness, and real-world safety monitoring remains limited. Multi-center trials and transparent reporting of failure modes will be crucial for translation (13,67-78).

Risk prediction & stratification

Risk tools throughout the cardiac continuum now depend on AI technology, which starts in emergency department triage and extends to long-term event prediction. AI models achieve better results than traditional rules-of-thumb, only when implemented correctly, and their performance is validated.

The CoDE-ACS model represents a multicenter machine-learning system that analyzes 10,038 patient data to personalize high-sensitivity troponin results based on clinical factors, including age and sex, comorbidities, and blood sample timing. The CoDE-ACS method demonstrated excellent performance in external testing by correctly eliminating myocardial infarction in 61% of patients with a 99.6% negative predictive value. The model showed a 75.5% positive predictive value and 97.1% specificity when identifying 10% of patients needing immediate myocardial infarction evaluation. The deep-learning ECG models now identify occlusion MI without ST-elevation (OMI), which traditional STEMI criteria fail to detect, while providing immediate reperfusion benefits to these patients, according to observational cohort studies (84).

Wearable devices have demonstrated their ability to perform population-scale AF screening through the Apple Heart Study, which enrolled 419,000 participants to validate an irregular-pulse detection algorithm for AF identification and subsequent monitoring purposes (85). AI technology enables 12-lead EKG analysis to identify future and paroxysmal AF occurrences from sinus-rhythm EKG recordings, which researchers have confirmed through extensive health-system data analysis and clinical trials. A study by Attia et al. demonstrated how their model detects hidden AF risk factors by analyzing EKGs that appear normal (66).

The stroke prediction performance of CHA₂DS₂-VASc and HAS-BLED risk scores remains limited for death prediction, but advanced ML models, including Multi-Label Gradient Boosting Decision Tree (ML-GBDT) and biomarker-based ABC-AF scores, show better predictive capabilities (86,87). The ML-GBDT model demonstrated better stroke prediction performance than CHA₂DS₂-VASc in a global registry study, with an AUC of 0.691 compared to 0.613, and it outperformed CHA₂DS₂-VASc in death and major bleeding risk prediction, with AUC values of 0.785 and 0.698 respectively (86). The ABC-stroke score achieved a c-index of 0.65, which outperformed the CHA₂DS₂-VASc score with a c-index of 0.60 (87). The AI models provide enhanced risk assessment through their ability to process diverse clinical and biomarker information, which leads to better stroke prevention decisions.

ML models have performed better than conventional stroke risk assessment methods in every analysis. The analysis of 7 million patients through ML models using Extreme Gradient Boosting (XGB) and Random Forest (RF) demonstrated superior performance in stroke prediction and prognosis for AF patients with an average AUROC of 0.75 (88). In contrast, studies on traditional scores like CHA₂DS₂-VASc often reveal more modest performance, with a c-index as low as 0.55. On the other hand, the ABC-stroke score, which incorporates biomarkers, also showed superior performance with a c-index of 0.64 (104).

AI now predicts lethal arrhythmias and SCD with greater precision than conventional methods. The DL model processed ECG data to achieve a 0.889 (95% CI: 0.861–0.917) AUROC for SCD prediction, which exceeded the conventional model’s 0.712 AUROC (105). The combination of ECG data with LGE MRI, clinical information through a multimodal AI model produced an AUROC of 0.84 (95% CI: 0.71–0.96) and sensitivity of 0.98 (95% CI: 0.75–1.00) and specificity of 0.73 (95% CI: 0.58–0.97) in non-ischemic cardiomyopathy patients (80). The dynamic ML model processed ECG data over time through VAEs to achieve an AUROC of 0.738. The method outperformed the 0.639 result obtained by the static baseline-only model (70). The results demonstrate how AI performs better at risk prediction because it processes intricate and time-dependent data sets.

AI models based on ECG technology have successfully identified asymptomatic LV systolic dysfunction cases. The AI system using ECG data achieved an AUC of 0.839 and 73% sensitivity and 83% specificity for LVSD detection in Chagas disease patients (106). The CNN model processed ECG data to predict EF values below or equal to 35% with a 0.93 AUC and 85.7% total accuracy in general population studies. The model successfully predicted ventricular dysfunction development with a hazard ratio of 4.1 (13). The AI system for LVSD detection in emergency departments achieved a detection accuracy of 0.89, which exceeded the performance of NT-proBNP testing at 0.80 (107). The AI system demonstrated success in detecting diastolic dysfunction and elevated filling pressure by achieving a detection accuracy of 0.911 for increased filling pressure (63). The AI-ECG system demonstrated equivalent prognostic capabilities to echocardiography while providing early detection capabilities.

Evidence quality and limitations: risk prediction studies frequently emphasize AUC while underreporting calibration, absolute risk, confidence intervals, and decision-impact measures, which limits clinical interpretability. Prospective evaluation and monitoring for drift are particularly important when models are deployed across evolving populations and practice patterns. Future work should prioritize calibration, subgroup analyses, and outcome-oriented validation rather than performance metrics alone (79-87).

Emerging paradigms and future directions

Generative AI and large language models (LLMs) in cardiology

Generative AI and LLMs extend cardiovascular AI beyond classification to workflow-facing tasks such as clinical summarization, drafting structured reports (e.g., echocardiography or catheterization narratives from measurements), patient-facing education materials, and guideline-aware decision support. Because these functions address documentation burden and information overload, they may drive adoption faster than incremental gains in retrospective diagnostic performance. However, LLMs pose distinct safety risks, including hallucinations, fabricated citations, and failures under distribution shift when prompts, templates, or workflows differ from training conditions. Deployment therefore requires guardrails, including constrained outputs, retrieval-augmented generation, clinician verification, and monitoring for drift and recurrent error patterns (108).

Foundation models and self-supervised learning in ECG and imaging

Foundation models trained with self-supervised or weakly supervised objectives can be adapted to multiple downstream tasks with limited labeled data, which is attractive in cardiology where labels are costly and heterogeneous. In ECG, learned representations may enable broader diagnostic coverage and improved generalization, though external validation and prospective evaluation remain essential (109). In imaging, transformer-based backbones and multimodal encoders support transfer across tasks such as segmentation, detection, and report generation. These approaches may reduce “data scarcity” but increase requirements for calibration, transparency, and governance across heterogeneous settings (4).

Multimodal models and the end of “Single-Silo” cardiology AI

Multimodal learning integrates waveforms, imaging, structured EHR variables, and clinical text, more closely reflecting clinical reasoning than single-modality models. Potential advantages include robustness to missing inputs, cross-modal retrieval (e.g., linking ECG patterns to imaging phenotypes), and interactive synthesis of longitudinal records. Key risks include label leakage, temporality violations, and evaluation using proxy endpoints; studies should therefore prespecify time ordering and emphasize clinically meaningful outcomes.

Privacy-preserving learning, federated approaches, and synthetic data

Federated learning and privacy-preserving training can enable multi-institutional model development without centralizing raw data, potentially improving generalizability while reducing governance barriers. Synthetic data may support development and testing, but reliability depends on preserving rare, safety-critical edge cases and maintaining performance under real-world artifacts and documentation variability. These methods should be viewed as complementary tools rather than universal solutions, with transparent reporting of data provenance, validation, and failure modes.

Translation, evidence quality, and regulation in the current era

Despite impressive retrospective performance, adoption is constrained by workflow integration, alert fatigue, reimbursement limitations, and calibration drift under dataset shift. Clinical value depends on actionability: outputs must be routed appropriately, paired with thresholds and confirmatory testing, and embedded in follow-up pathways. Evidence should be interpreted by tier, distinguishing internal from multisite external validation and retrospective accuracy from prospective trials demonstrating improved decision-making, time-to-diagnosis, or patient outcomes (110). In addition, the FDA has described mechanisms such as a Predetermined Change Control Plan (PCCP) to support iterative improvement while maintaining safety and effectiveness (111).

Limitations

This narrative review summarizes a rapidly evolving literature with substantial heterogeneity in wearable devices, signals, endpoints, and analytic methods, limiting direct cross-study comparability. Because many studies emphasize feasibility and diagnostic performance rather than clinical outcomes, the strength of evidence supporting downstream outcome improvement remains variable. Publication bias and selective reporting are possible, and the pace of device and software iteration may outdate specific technical implementations over time. Accordingly, the findings should be interpreted as an evidence map of current capabilities and gaps rather than definitive comparative effectiveness conclusions.

Conclusions

Wearable and remote monitoring technologies, increasingly augmented by AI methods, can enable scalable cardiovascular screening and longitudinal risk assessment outside traditional care settings. Current evidence supports feasibility and promising diagnostic performance for select applications, but consistent demonstration of improved hard clinical outcomes remains limited. Future work should prioritize external validation, transparent reporting, bias assessment, and pragmatic trials that measure clinically meaningful endpoints within real-world workflows.

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-2025-510/rc

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-510/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-2025-510/coif). K.G. serves as an unpaid editorial board member of Cardiovascular Diagnosis and Therapy from April 2025 to March 2027. 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. World Health Organization. Cardiovascular diseases (CVDs). Available online: https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds). Accessed Jan 22, 2026.
2. Schwab K, Nguyen D, Ungab G, et al. Artificial intelligence MacHIne learning for the detection and treatment of atrial fibrillation guidelines in the emergency department setting (AIM HIGHER): Assessing a machine learning clinical decision support tool to detect and treat non-valvular atrial fibrillation in the emergency department. J Am Coll Emerg Physicians Open 2021;2:e12534. [Crossref] [PubMed]
3. Chen M, Wang Y, Wang Q, et al. Impact of human and artificial intelligence collaboration on workload reduction in medical image interpretation. NPJ Digit Med 2024;7:349. [Crossref] [PubMed]
4. Vandenberk B, Chew DS, Prasana D, et al. Successes and challenges of artificial intelligence in cardiology. Front Digit Health 2023;5:1201392. [Crossref] [PubMed]
5. Rudnicka Z, Pręgowska A, Glądys K, et al. Advancements in artificial intelligence-driven techniques for interventional cardiology. Cardiol J 2024;31:321-41. [Crossref] [PubMed]
6. Gala D, Behl H, Shah M, et al. The Role of Artificial Intelligence in Improving Patient Outcomes and Future of Healthcare Delivery in Cardiology: A Narrative Review of the Literature. Healthcare (Basel) 2024;12:481. [Crossref] [PubMed]
7. Shuja MH, Shakil F, Shuja SH, et al. Harnessing Artificial Intelligence in Cardiology: Advancements in Diagnosis, Treatment, and Patient Care. Heart Views 2024;25:241-8. [Crossref] [PubMed]
8. Zippel-Schultz B, Schenk C, Müller-Wieland D, et al. Künstliche Intelligenz in der Kardiologie. Herzschrittmacherther Elektrophysiol 2021;32:89-98. [Crossref] [PubMed]
9. Mangione S, Nieman LZ. Cardiac auscultatory skills of internal medicine and family practice trainees: a comparison of diagnostic proficiency. JAMA 1997;278:717-22. [Crossref] [PubMed]
10. Potes C, Parvaneh S, Rahman A, et al. Ensemble of feature-based and deep learning-based classifiers for detection of abnormal heart sounds. Comput Cardiol 2016;43:182-399.
11. 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]
12. Attia ZI, Kapa S, Lopez-Jimenez F, et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med 2019;25:70-4. [Crossref] [PubMed]
13. Raghunath S, Pfeifer JM, Ulloa-Cerna AE, et al. Deep Neural Networks Can Predict New-Onset Atrial Fibrillation From the 12-Lead ECG and Help Identify Those at Risk of Atrial Fibrillation-Related Stroke. Circulation 2021;143:1287-98. [Crossref] [PubMed]
14. Yao X, Rush RD, Inselman JW, et al. AI-enabled ECGs for identification of low ejection fraction: a pragmatic randomized trial. Nat Med 2021;27:815-23. [Crossref] [PubMed]
15. Mehari T, Strodthoff N. Self-supervised representation learning from 12-lead ECG data. Comput Biol Med 2022;141:105114. [Crossref] [PubMed]
16. Chen KW, Wu YC, Liu MH, et al. Artificial intelligence-assisted remote detection of ST-elevation myocardial infarction using a mini-12-lead electrocardiogram device in prehospital ambulance care. Front Cardiovasc Med 2022;9:1001982. [Crossref] [PubMed]
17. Adedinsewo D, Eberly L, Sokumbi O, et al. Health Disparities, Clinical Trials, and the Digital Divide. Mayo Clin Proc 2023;98:1875-87. [Crossref] [PubMed]
18. 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]
19. Zhang J, Gajjala S, Agrawal P, et al. Fully Automated Echocardiogram Interpretation in Clinical Practice. Circulation 2018;138:1623-35. [Crossref] [PubMed]
20. 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]
21. He B, Kwan AC, Cho JH, et al. Blinded, randomized trial of sonographer versus AI cardiac function assessment. Nature 2023;616:520-4. [Crossref] [PubMed]
22. Kusunose K, Haga A, Inoue M, et al. Clinically Feasible and Accurate View Classification of Echocardiographic Images Using Deep Learning. Biomolecules 2020;10:665. [Crossref] [PubMed]
23. Tromp J, Bauer D, Claggett BL, et al. A formal validation of a deep learning-based automated workflow for the interpretation of the echocardiogram. Nat Commun 2022;13:6776. [Crossref] [PubMed]
24. Akerman AP, Porumb M, Scott CG, et al. Automated Echocardiographic Detection of Heart Failure With Preserved Ejection Fraction Using Artificial Intelligence. JACC Adv 2023;2:100452. [Crossref] [PubMed]
25. Christensen M, Vukadinovic M, Yuan N, et al. Vision-language foundation model for echocardiogram interpretation. Nat Med 2024;30:1481-8. [Crossref] [PubMed]
26. Hachamovitch R, Berman DS, Shaw LJ, et al. Incremental prognostic value of myocardial perfusion single photon emission computed tomography for the prediction of cardiac death: differential stratification for risk of cardiac death and myocardial infarction. Circulation 1998;97:535-43. [Crossref] [PubMed]
27. Shiri I. Standard SPECT myocardial perfusion estimation from half-time acquisitions using deep convolutional residual neural networks. J Nucl Cardiol 2021;28:2761-79. [Crossref] [PubMed]
28. Ramon AJ, Yuan Y, Pretorius PH, et al. Initial investigation of low-dose SPECT-MPI via deep learning. IEEE Nucl Sci Symp Med Imaging Conf Proc 2018. doi: 10.1109/NSSMIC.2018.8824548.10.1109/NSSMIC.2018.8824548
29. Yang P, Zhang Z, Wei J, et al. Deep learning-based CT-free attenuation correction for cardiac SPECT: a new approach. BMC Med Imaging 2025;25:38. [Crossref] [PubMed]
30. Sohlberg A, Kangasmaa T, Constable C, et al. Comparison of deep learning-based denoising methods in cardiac SPECT. EJNMMI Phys 2023;10:9. [Crossref] [PubMed]
31. 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]
32. Hu LH, Betancur J, Sharir T, et al. Machine learning predicts per-vessel early coronary revascularization after fast myocardial perfusion SPECT: results from multicentre REFINE SPECT registry. Eur Heart J Cardiovasc Imaging 2020;21:549-59. [Crossref] [PubMed]
33. Apostolopoulos ID, Apostolopoulos DI, Spyridonidis TI, et al. Multi-input deep learning approach for Cardiovascular Disease diagnosis using Myocardial Perfusion Imaging and clinical data. Phys Med 2021;84:168-77. [Crossref] [PubMed]
34. Bai W, Sinclair M, Tarroni G, et al. Automated cardiovascular magnetic resonance image analysis with fully convolutional networks. J Cardiovasc Magn Reson 2018;20:65. [Crossref] [PubMed]
35. Robinson R, Valindria VV, Bai W, et al. Automated quality control in image segmentation: application to the UK Biobank cardiovascular magnetic resonance imaging study. J Cardiovasc Magn Reson 2019;21:18. [Crossref] [PubMed]
36. Ruijsink B, Puyol-Antón E, Oksuz I, et al. Fully Automated, Quality-Controlled Cardiac Analysis From CMR: Validation and Large-Scale Application to Characterize Cardiac Function. JACC Cardiovasc Imaging 2020;13:684-95. [Crossref] [PubMed]
37. Klemenz AC, Reichardt L, Gorodezky M, et al. Accelerated Cardiac MRI with Deep Learning-based Image Reconstruction for Cine Imaging. Radiol Cardiothorac Imaging 2024;6:e230419. [Crossref] [PubMed]
38. Ghanbari F, Joyce T, Lorenzoni V, et al. AI Cardiac MRI Scar Analysis Aids Prediction of Major Arrhythmic Events in the Multicenter DERIVATE Registry. Radiology 2023;307:e222239. [Crossref] [PubMed]
39. Zhang Q, Burrage MK, Shanmuganathan M, et al. Artificial Intelligence for Contrast-Free MRI: Scar Assessment in Myocardial Infarction Using Deep Learning-Based Virtual Native Enhancement. Circulation 2022;146:1492-503. [Crossref] [PubMed]
40. 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. [Crossref] [PubMed]
41. Isgum I, Rutten A, Prokop M, et al. Detection of coronary calcifications from computed tomography scans for automated risk assessment of coronary artery disease. Med Phys 2007;34:1450-61. [Crossref] [PubMed]
42. Sandstedt M, Henriksson L, Janzon M, et al. Evaluation of an AI-based, automatic coronary artery calcium scoring software. Eur Radiol 2020;30:1671-8. [Crossref] [PubMed]
43. Jonas RA, Weerakoon S, Fisher R, et al. Interobserver variability among expert readers quantifying plaque volume and plaque characteristics on coronary CT angiography: a CLARIFY trial sub-study. Clin Imaging 2022;91:19-25. [Crossref] [PubMed]
44. Sandstedt M, Henriksson L, Janzon M, et al. Evaluation of an AI-based, automatic coronary artery calcium scoring software. Eur Radiol 2020;30:1671-8. [Crossref] [PubMed]
45. Choi AD, Marques H, Kumar V, et al. CT evaluation by artificial intelligence for atherosclerosis, stenosis and vascular morphology (CLARIFY): a multi-center, international study. J Cardiovasc Comput Tomogr 2021;15:470-6. [Crossref] [PubMed]
46. Paul JF, Rohnean A, Giroussens H, et al. Evaluation of a deep learning model on coronary CT angiography for automatic stenosis detection. Diagn Interv Imaging 2022;103:316-23. [Crossref] [PubMed]
47. Hell MM, Dey D, Marwan M, et al. Non-invasive prediction of hemodynamically significant coronary artery stenoses by contrast density difference in coronary CT angiography. Eur J Radiol 2015;84:1502-8. [Crossref] [PubMed]
48. Griffin WF, Choi AD, Riess JS, et al. AI Evaluation of Stenosis on Coronary CTA, Comparison With Quantitative Coronary Angiography and Fractional Flow Reserve: A CREDENCE Trial Substudy. JACC Cardiovasc Imaging 2023;16:193-205. [Crossref] [PubMed]
49. Tang CX, Guo BJ, Schoepf JU, et al. Feasibility and prognostic role of machine learning-based FFR(CT) in patients with stent implantation. Eur Radiol 2021;31:6592-604. [Crossref] [PubMed]
50. Tesche C, De Cecco CN, Baumann S, et al. Coronary CT Angiography-derived Fractional Flow Reserve: Machine Learning Algorithm versus Computational Fluid Dynamics Modeling. Radiology 2018;288:64-72. [Crossref] [PubMed]
51. Qiao HY, Tang CX, Schoepf UJ, et al. One-year outcomes of CCTA alone versus machine learning-based FFR(CT) for coronary artery disease: a single-center, prospective study. Eur Radiol 2022;32:5179-88. [Crossref] [PubMed]
52. Lee S, Chu Y, Ryu J, et al. Artificial Intelligence for Detection of Cardiovascular-Related Diseases from Wearable Devices: A Systematic Review and Meta-Analysis. Yonsei Med J 2022;63:S93-S107. [Crossref] [PubMed]
53. Bhatt P, Liu J, Gong Y, et al. Emerging Artificial Intelligence-Empowered mHealth: Scoping Review. JMIR Mhealth Uhealth 2022;10:e35053. [Crossref] [PubMed]
54. Ginsburg GS, Picard RW, Friend SH. Key Issues as Wearable Digital Health Technologies Enter Clinical Care. N Engl J Med 2024;390:1118-27. [Crossref] [PubMed]
55. Shumba AT, Montanaro T, Sergi I, et al. Wearable Technologies and AI at the Far Edge for Chronic Heart Failure Prevention and Management: A Systematic Review and Prospects. Sensors (Basel) 2023;23:6896. [Crossref] [PubMed]
56. Bayoumy K, Gaber M, Elshafeey A, et al. Smart wearable devices in cardiovascular care: where we are and how to move forward. Nat Rev Cardiol 2021;18:581-99. [Crossref] [PubMed]
57. Leclercq C, Witt H, Hindricks G, et al. Wearables, telemedicine, and artificial intelligence in arrhythmias and heart failure: Proceedings of the European Society of Cardiology Cardiovascular Round Table. Europace 2022;24:1372-83. [Crossref] [PubMed]
58. Motwani A, Shukla PK, Pawar M. Ubiquitous and smart healthcare monitoring frameworks based on machine learning: A comprehensive review. Artif Intell Med 2022;134:102431. [Crossref] [PubMed]
59. Marvasti TB, Gao Y, Murray KR, et al. Unlocking Tomorrow's Health Care: Expanding the Clinical Scope of Wearables by Applying Artificial Intelligence. Can J Cardiol 2024;40:1934-45. [Crossref] [PubMed]
60. Neri L, Oberdier MT, van Abeelen KCJ, et al. Electrocardiogram Monitoring Wearable Devices and Artificial-Intelligence-Enabled Diagnostic Capabilities: A Review. Sensors (Basel) 2023;23:4805. [Crossref] [PubMed]
61. Huang J, Emori H, Ding D, et al. Diagnostic performance of intracoronary optical coherence tomography-based versus angiography-based fractional flow reserve for the evaluation of coronary lesions. EuroIntervention 2020;16:568-76. [Crossref] [PubMed]
62. Tsai DJ, Lin C, Liu WT, et al. Artificial intelligence-assisted diagnosis and prognostication in low ejection fraction using electrocardiograms in inpatient department: a pragmatic randomized controlled trial. BMC Med 2025;23:342. [Crossref] [PubMed]
63. Lee E, Ito S, Miranda WR, et al. Artificial intelligence-enabled ECG for left ventricular diastolic function and filling pressure. NPJ Digit Med 2024;7:4. [Crossref] [PubMed]
64. Segar MW, Hall JL, Jhund PS, et al. Machine Learning-Based Models Incorporating Social Determinants of Health vs Traditional Models for Predicting In-Hospital Mortality in Patients With Heart Failure. JAMA Cardiol 2022;7:844-54. [Crossref] [PubMed]
65. Stehlik J, Schmalfuss C, Bozkurt B, et al. Continuous Wearable Monitoring Analytics Predict Heart Failure Hospitalization: The LINK-HF Multicenter Study. Circ Heart Fail 2020;13:e006513. [Crossref] [PubMed]
66. 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]
67. Chen B, Maslove DM, Curran JD, et al. A deep learning model for the classification of atrial fibrillation in critically ill patients. Intensive Care Med Exp 2023;11:2. [Crossref] [PubMed]
68. Alam R, Aguirre A, Stultz CM. Detecting QT prolongation from a single-lead ECG with deep learning. PLOS Digit Health 2024;3:e0000539. [Crossref] [PubMed]
69. Jiang R, Cheung CC, Garcia-Montero M, et al. Deep Learning-Augmented ECG Analysis for Screening and Genotype Prediction of Congenital Long QT Syndrome. JAMA Cardiol 2024;9:377-84. [Crossref] [PubMed]
70. Kolk MZH, Ruipérez-Campillo S, Allaart CP, et al. Multimodal explainable artificial intelligence identifies patients with non-ischaemic cardiomyopathy at risk of lethal ventricular arrhythmias. Sci Rep 2024;14:14889. [Crossref] [PubMed]
71. Kim Y, Yoon HJ, Suh J, et al. Artificial Intelligence-Based Fully Automated Quantitative Coronary Angiography vs Optical Coherence Tomography-Guided PCI: The FLASH Trial. JACC Cardiovasc Interv 2025;18:187-97. [Crossref] [PubMed]
72. Alkhouli M, Rostami B, Attia Z, et al. Artificial intelligence for extracting noncoronary data from angiography: the AI-ENCODE study. J Soc Cardiovasc Angiogr Interv 2024;3:101870. [Crossref]
73. Durand E, Sabatier R, Smits PC, et al. Evaluation of the R-One robotic system for percutaneous coronary intervention: the R-EVOLUTION study. EuroIntervention 2023;18:e1339-47. [Crossref] [PubMed]
74. Łajczak P, Ayesha A, Sahin OK, et al. Comparison between robot-assisted and manual percutaneous coronary intervention - an updated systematic review, meta-analysis, propensity-matched investigation, and trial sequential analysis. Cardiovasc Interv Ther 2025;40:474-89. [Crossref] [PubMed]
75. Leung J, Xu J, French JK, et al. Rationale and design of the randomized, controlled percutaneous coronary intervention using assisted robotic technology (PARTY) trial. Open Heart 2024;11:e002950. [Crossref]
76. Patel TM, Shah SC, Pancholy SB. Long Distance Tele-Robotic-Assisted Percutaneous Coronary Intervention: A Report of First-in-Human Experience. EClinicalMedicine 2019;14:53-8. [Crossref] [PubMed]
77. Chandramohan N, Hinton J, O'Kane P, et al. Artificial Intelligence for the Interventional Cardiologist: Powering and Enabling OCT Image Interpretation. Interv Cardiol 2024;19:e03. [Crossref] [PubMed]
78. Sibbald M, Mitchell HR, Buccola J, et al. Impact of Artificial Intelligence-Enhanced Optical Coherence Tomography Software on Percutaneous Coronary Intervention Decisions. J Soc Cardiovasc Angiogr Interv 2025;4:102438. [Crossref] [PubMed]
79. Xu B, Tu S, Song L, et al. Angiographic quantitative flow ratio-guided coronary intervention (FAVOR III China): a multicentre, randomised, sham-controlled trial. Lancet 2021;398:2149-59. [Crossref] [PubMed]
80. Andersen BK, Sejr-Hansen M, Maillard L, et al. Quantitative flow ratio versus fractional flow reserve for coronary revascularisation guidance (FAVOR III Europe): a multicentre, randomised, non-inferiority trial. Lancet 2024;404:1835-46. [Crossref] [PubMed]
81. Yu W, Huang J, Jia D, et al. Diagnostic accuracy of intracoronary optical coherence tomography-derived fractional flow reserve for assessment of coronary stenosis severity. EuroIntervention 2019;15:189-97. [Crossref] [PubMed]
82. Hu F, Ding D, Westra J, et al. Diagnostic accuracy of optical flow ratio: an individual patient-data meta-analysis. EuroIntervention 2023;19:e145-54. [Crossref] [PubMed]
83. Tu S, Westra J, Adjedj J, et al. Fractional flow reserve in clinical practice: from wire-based invasive measurement to image-based computation. Eur Heart J 2020;41:3271-9. [Crossref] [PubMed]
84. Doudesis D, Lee KK, Boeddinghaus J, et al. Machine learning for diagnosis of myocardial infarction using cardiac troponin concentrations. Nat Med 2023;29:1201-10. [Crossref] [PubMed]
85. 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]
86. Lu J, Bisson A, Bennamoun M, et al. Predicting multifaceted risks using machine learning in atrial fibrillation: insights from GLORIA-AF study. Eur Heart J Digit Health 2024;5:235-46. [Crossref] [PubMed]
87. Niederdöckl J, Oppenauer J, Schnaubelt S, et al. The ABC-Stroke Score Refines Stroke Risk Stratification in Patients With Atrial Fibrillation at the Emergency Department. Front Med (Lausanne) 2022;9:830580. [Crossref] [PubMed]
88. Goh B, Bhaskar SMM. Evaluating Machine Learning Models for Stroke Prognosis and Prediction in Atrial Fibrillation Patients: A Comprehensive Meta-Analysis. Diagnostics (Basel) 2024;14:2391. [Crossref] [PubMed]
89. Zhang L, Cheng Z, Xu D, et al. Developing an AI-assisted digital auscultation tool for automatic assessment of the severity of mitral regurgitation: protocol for a cross-sectional, non-interventional study. BMJ Open 2024;14:e074288. [Crossref] [PubMed]
90. Hannun AY, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nat Med 2019;25:65-9. [Crossref] [PubMed]
91. Khurshid S, Friedman S, Reeder C, et al. ECG-Based Deep Learning and Clinical Risk Factors to Predict Atrial Fibrillation. Circulation 2022;145:122-33. [Crossref] [PubMed]
92. Attia ZI, Harmon DM, Dugan J, et al. Prospective evaluation of smartwatch-enabled detection of left ventricular dysfunction. Nat Med 2022;28:2497-503. [Crossref] [PubMed]
93. Lin C, Lin YH, Chen SW, et al. Point-of-care AI-enabled ECG for dyskalemia recognition. NPJ Digit Med 2022;5:60.
94. Kwon JM, Choi Y, Jeon KH, et al. A deep learning algorithm to detect anaemia with ECGs: a retrospective, multicentre study. Lancet Digit Health 2020;2:e358-e367. [Crossref] [PubMed]
95. Petmezas G, Stefanopoulos L, Tachos N, et al. State-of-the-Art Deep Learning Methods on Electrocardiogram Data: Systematic Review. JMIR Med Inform 2022;10:e38454. [Crossref] [PubMed]
96. Reyna MA, Sadr N, Perez Alday EA, et al. PhysioNet/CinC Challenge 2021: generalization across leads and institutions. Comput Cardiol 2021;48:1-4.
97. Herman R, Mohareb H, Smith SW, et al. AI-ECG for acute coronary occlusion MI: international evaluation. Eur Heart J Digit Health 2024;5:123-34. [Crossref] [PubMed]
98. Kong H, Wang Z, Zhou Z, et al. Clinical feasibility of two cardiac deep learning cine magnetic resonance imaging sequences: Single-breath-hold and free-breathing motion-corrected approaches. J Cardiovasc Magn Reson 2025;27:101983. [Crossref] [PubMed]
99. Wang C, Lyu J, Wang S, et al. CMRxRecon: A publicly available k-space dataset and benchmark to advance deep learning for cardiac MRI. Sci Data 2024;11:687. [Crossref] [PubMed]
100. Morales MA, Manning WJ, Nezafat R. Present and Future Innovations in AI and Cardiac MRI. Radiology 2024;310:e231269. [Crossref] [PubMed]
101. Xue H, Davies RH, Brown LAE, et al. Automated Inline Analysis of Myocardial Perfusion MRI with Deep Learning. Radiol Artif Intell 2020;2:e200009. [Crossref] [PubMed]
102. Knott KD, Seraphim A, Augusto JB, et al. The Prognostic Significance of Quantitative Myocardial Perfusion: An Artificial Intelligence-Based Approach Using Perfusion Mapping. Circulation 2020;141:1282-91. [Crossref] [PubMed]
103. Fahmy AS, Rowin EJ, Jaafar N, et al. Radiomics of Late Gadolinium Enhancement Reveals Prognostic Value of Myocardial Scar Heterogeneity in Hypertrophic Cardiomyopathy. JACC Cardiovasc Imaging 2024;17:16-27. [Crossref] [PubMed]
104. Oldgren J, Hijazi Z, Lindbäck J, et al. Performance and Validation of a Novel Biomarker-Based Stroke Risk Score for Atrial Fibrillation. Circulation 2016;134:1697-707. [Crossref] [PubMed]
105. Holmstrom L, Chugh H, Nakamura K, et al. An ECG-based artificial intelligence model for assessment of sudden cardiac death risk. Commun Med (Lond) 2024;4:17. [Crossref] [PubMed]
106. Brito BOF, Attia ZI, Martins LNA, et al. Left ventricular systolic dysfunction predicted by artificial intelligence using the electrocardiogram in Chagas disease patients-The SaMi-Trop cohort. PLoS Negl Trop Dis 2021;15:e0009974. [Crossref] [PubMed]
107. Adedinsewo D, Carter RE, Attia Z, et al. Artificial Intelligence-Enabled ECG Algorithm to Identify Patients With Left Ventricular Systolic Dysfunction Presenting to the Emergency Department With Dyspnea. Circ Arrhythm Electrophysiol 2020;13:e008437. [Crossref] [PubMed]
108. U.S. Food and Drug Administration. Artificial intelligence-enabled device software functions: lifecycle management and marketing. Available online: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing. Accessed Jan 22, 2026.
109. Tian Y, Li Z, Jin Y, et al. Foundation model of ECG diagnosis: Diagnostics and explanations of any form and rhythm on ECG. Cell Rep Med 2024;5:101875. [Crossref] [PubMed]
110. U.S. Food and Drug Administration. Artificial intelligence-enabled device software functions: lifecycle management and marketing submission recommendations (draft guidance). Available online: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing. Published Jan 7, 2025. Accessed Jan 22, 2026.
111. U.S. Food and Drug Administration. Marketing submission recommendations for a predetermined change control plan for artificial intelligence-enabled device software functions: guidance for industry and Food and Drug Administration staff. Available online: https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enabled-device-software-functions-lifecycle-management-and-marketing. Published Aug 18, 2025. Accessed Jan 22, 2026.