CoChat
Sign Up
Cited by 0ECG Monitoring and AnalysisMachine Learning in HealthcareCardiac electrophysiology and arrhythmias

Deep Learning Applied to 12-Lead ECGs for Detection of Structural, Metabolic and Systemic Disease

Open access
Shaulian, Shlomo, Zeltser, Roman, Makaryus, Amgad N.·2026·Diagnostics
DOI 10.3390/diagnostics16172834PMC PMC13565729
Open this paper in CoChat — free account

Abstract

The 12-lead electrocardiogram (ECG) is inexpensive, noninvasive, and widely available, but conventional interpretation may not capture subtle signals related to cardiac structure, systemic physiology, and future risk. This review examines the use of deep learning–enabled ECG analysis beyond conventional arrhythmia detection and evaluates the maturity, clinical relevance, and implementation challenges of these applications. This narrative review synthesizes landmark studies, external validation cohorts, pragmatic implementation trials, and recent investigations of artificial intelligence–enabled ECG (AI-ECG) models for structural, metabolic, systemic, and prognostic assessment. Particular attention is given to model performance, validation, clinical actionability, and barriers to translation. AI-ECG models have demonstrated the ability to detect reduced left ventricular ejection fraction, hypertrophic cardiomyopathy, valvular disease, cardiac amyloidosis, pulmonary hypertension, hyperkalemia and other dyskalemias, hyperthyroidism, anemia, and sepsis, and to estimate biologic age and mortality risk. Evidence is most mature for screening for reduced left ventricular ejection fraction, supported by large derivation cohorts, external validation, prognostic follow-up, and the EAGLE pragmatic trial. Evidence for many other applications remains retrospective or exploratory, with limitations related to reference standards, generalizability, calibration, disease prevalence, interpretability, and the absence of clearly defined clinical pathways. AI-ECG has the potential to expand the ECG from a conventional diagnostic test into a screening and decision-support platform. Its near-term role is to identify patients who may benefit from confirmatory imaging, laboratory testing, rhythm monitoring, or specialist evaluation. Broader adoption will require prospective validation, workflow integration, subgroup assessment, post-deployment monitoring, regulatory oversight, and evidence that AI-guided care improves outcomes.

AI snapshot

Written from the paper’s full text.

TL;DR

This narrative review surveys deep learning–enabled 12-lead ECG (AI-ECG) applications beyond arrhythmia detection, covering structural, metabolic, systemic, and prognostic targets. It concludes that evidence is most mature for screening reduced left ventricular ejection fraction, while most other applications remain retrospective or exploratory and lack prospective proof of clinical benefit.

Key findings

  • Reduced LVEF detection is the most mature application: Attia et al. trained a CNN on >44,000 patients and tested on >52,000, achieving AUC 0.93, sensitivity 86.3%, specificity 85.7%; the EAGLE pragmatic cluster-randomized trial showed AI-guided screening increased new low-ejection-fraction diagnoses within 90 days and targeted echocardiography in AI-positive patients without substantially increasing overall echocardiography use.
  • Structural heart disease models show high discrimination: HCM detection AUC 0.96 (sensitivity 87%, specificity 90%), with international external validation AUC 0.922; aortic stenosis screening sensitivity 78% and specificity 74%; cardiac amyloidosis post-development validation AUC 0.84 (lower in Hispanic patients and those with LVH or LBBB).
  • Metabolic/systemic models report strong performance: hyperkalemia AUC 0.853–0.883 across validation sites (trained on >1.5 million ECGs from >449,000 patients); point-of-care potassium AUC >0.95 for moderate-to-severe hyperkalemia and >0.85 for hypokalemia; hyperthyroidism AUC 0.926 internal/0.883 external; anemia 0.854/0.824; sepsis 0.901/0.863.
  • Prognostic models: one-year all-cause mortality AUC 0.88 overall and 0.85 among ECGs read as normal by physicians; ECG-age >8 years above chronological age was associated with higher mortality, and >8 years younger with lower mortality, even among apparently normal ECGs.

Methods

  • Narrative (not systematic) review; targeted literature search of PubMed/MEDLINE and Google Scholar for English-language articles from January 2015 through August 2026, with the final search on 21 August 2026.
  • Search combined terms for AI/deep learning/machine learning with ECG terms and condition-specific terms (ventricular dysfunction, heart failure, structural heart disease, HCM, valvular disease, amyloidosis, pulmonary hypertension, hyperkalemia, biologic age, mortality, implementation); reference lists were also screened.
  • Priority given to peer-reviewed original studies, external validation studies, prospective/pragmatic implementation studies, and consensus/reporting guidelines; arrhythmia-classification-only studies generally excluded. No formal risk-of-bias assessment or quantitative meta-analysis was performed.

Limitations

  • Stated by the paper: as a narrative review, it performed no formal risk-of-bias assessment or meta-analysis; evidence for most applications beyond reduced ejection fraction remains retrospective or exploratory, with limitations in reference standards, generalizability, calibration, disease prevalence, interpretability, and absent defined clinical pathways.
  • Stated by the paper: many studies use echocardiographic estimates rather than gold-standard right heart catheterization (pulmonary hypertension) or derive labels from EHR/imaging/billing data, risking selection and verification bias; high discrimination in uncommon conditions does not guarantee high positive predictive value.
  • Evident from the text: reported performance metrics are drawn from heterogeneous studies differing in target condition, dataset composition, sample size, preprocessing, and validation strategy, so the review notes performance cannot be directly compared across architectures or studies, and no universally superior architecture is established.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/diagnostics16172834

Cites

66 of 66 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

73 of 116 cited sources checked · 59 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    ECG waveforms contain latent signatures of metabolic abnormalities.

    “Emerging evidence suggests that ECG waveforms contain latent signatures of ventricular dysfunction [4], structural heart disease [5], metabolic abnormalities [6], and broader physiologic characteristics such as biological age [7] and future mortality risk [8].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [6] Galloway, 2019, Development and Validation of a Deep-Learning Model to Screen for Hyperkalemia From the Electrocardiogram, JAMA Cardiol. source
  • Premisecitation

    ECG waveforms contain latent signatures of structural heart disease.

    “Emerging evidence suggests that ECG waveforms contain latent signatures of ventricular dysfunction [4], structural heart disease [5], metabolic abnormalities [6], and broader physiologic characteristics such as biological age [7] and future mortality risk [8].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [5] Tison, 2019, Automated and Interpretable Patient ECG Profiles for Disease Detection, Tracking, and Discovery, Circ. Cardiovasc. Qual. Outcomes source
  • Premisecitation

    ECG waveforms contain latent signatures of biological age.

    “Emerging evidence suggests that ECG waveforms contain latent signatures of ventricular dysfunction [4], structural heart disease [5], metabolic abnormalities [6], and broader physiologic characteristics such as biological age [7] and future mortality risk [8].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [7] Lima, 2021, Deep Neural Network-Estimated Electrocardiographic Age as a Mortality Predictor, Nat. Commun. source
  • Premisecitation

    AI-ECG has been applied to point-of-care potassium estimation.

    “Later work carried the approach to 12-lead detection of both hypokalemia and hyperkalemia [11] and point-of-care potassium estimation [15].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [15] Lin, 2022, Point-of-Care Artificial Intelligence-Enabled ECG for Dyskalemia: A Retrospective Cohort Analysis for Accuracy and Outcome Prediction, npj Digit. Med. source
  • Premisecitation

    An AI-ECG model had stable performance across time and demographic subgroups.

    “Harmon et al. reported stable performance of an AI-ECG model across time and demographic subgroups, addressing concerns about temporal drift and subgroup bias [20].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [20] Harmon, 2022, Real-World Performance, Long-Term Efficacy, and Absence of Bias in the Artificial Intelligence Enhanced Electrocardiogram to Detect Left Ventricular Systolic Dysfunction, Eur. Heart J.-Digit. Health source
  • Classified as probably not a substantial claim. 45
  • May combine two claims. 38

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Deep learning models can identify intricate patterns within large datasets without requiring explicit feature selection.

    “Unlike traditional rule-based approaches, deep learning models can identify intricate patterns within large datasets without requiring explicit feature selection [3].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [3] Fahim, 2025, Artificial Intelligence in Healthcare and Medicine: Clinical Applications, Therapeutic Advances, and Future Perspectives, Eur. J. Med. Res. source

      The Foundational AI technologies sections (Deep learning and Machine learning; pages unavailable) say that deep learning represents complex, high-dimensional data and enhances feature extraction, while machine learning recognizes patterns in extensive, intricate datasets. They do not establish that deep learning requires no explicit feature selection or make the claimed comparison with traditional rule-based approaches. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The model for elevated estimated pulmonary artery pressure performed consistently across important clinical subgroups and was externally validated in an independent cohort.

    “The model also performed consistently across important clinical subgroups and was externally validated in an independent cohort [36].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [36] Liu, 2022, Artificial Intelligence-Enabled Electrocardiogram Improves the Diagnosis and Prediction of Mortality in Patients with Pulmonary Hypertension, JACC Asia source

  • May combine two claims.

    Premisecitation

    Deep learning models analyze the entire ECG waveform and can detect subtle patterns distributed across multiple leads.

    “Rather than focusing solely on established ECG features such as intervals, amplitudes, and segment morphology, deep learning models analyze the entire waveform and can detect subtle patterns distributed across multiple leads [9].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [9] Ribeiro, 2020, Automatic Diagnosis of the 12-Lead ECG Using a Deep Neural Network, Nat. Commun. source

      The Discussion says the end-to-end classifier takes the raw ECG signal and learns features rather than relying on features extracted by traditional methods. Methods > Neural network architecture and training says the input contains the sampled signal from each lead. The paper does not establish that the detected patterns are subtle or distributed across multiple leads. (LLM, conf=0.85)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Interpretationcitation

    HCM has shown consistently high discrimination, international external validation, and early real-world implementation, making it one of the more reproducible structural AI-ECG targets.

    “HCM has shown consistently high discrimination, international external validation, and early real-world implementation, making it one of the more reproducible structural AI-ECG targets.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [10] Ko, 2020, Detection of Hypertrophic Cardiomyopathy Using a Convolutional Neural Network-Enabled Electrocardiogram, J. Am. Coll. Cardiol. source

      The Discussion (page not provided) reports high discrimination across the training, validation, and testing cohorts, with AUCs of 0.97, 0.95, and 0.96. But the Conclusions and Study Limitations (pages not provided) say the model still requires external validation and describe screening as a future possibility; this paper does not support international external validation or real-world implementation. (LLM, conf=0.90)

      Experimental — accuracy not yet measured
    • cites [27] Siontis, 2024, Hypertrophic Cardiomyopathy Detection with Artificial Intelligence Electrocardiography in International Cohorts: An External Validation Study, Eur. Heart J.-Digit. Health source

      The Methods and results report high discrimination (AUC 0.922) across external cohorts in Switzerland, the UK, and South Korea; the Discussion also describes favourable performance across diverse cohorts. But the Conclusion says clinical implementation still needs prospective evaluation. An earlier internal validation in routine clinical practice is mentioned, not evidence that this algorithm had been implemented in real-world care. Page numbers are unavailable in our copy. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [28] Desai, 2025, Real-World Artificial Intelligence-Based Electrocardiographic Analysis to Diagnose Hypertrophic Cardiomyopathy, JACC Clin. Electrophysiol. source

      The Discussion reports an external-dataset AUC of 0.975, and the Conclusions describe real-world clinical implementation for HCM detection at one center. Our copy does not establish international external validation or compare HCM with other structural AI-ECG targets for reproducibility; the Discussion cautions that this single-center study may not generalize across geographic settings. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • Rests on an assumption: HCM may be underdiagnosed, particularly in patients with subtle or nonspecific ECG findings.

  • May combine two claims.

    Interpretationcitation

    AI-ECG has a dual role: opportunistic screening in asymptomatic patients and evaluation of symptomatic patients in whom ventricular dysfunction is part of the differential.

    “Read together, these studies point to a dual role: opportunistic screening in asymptomatic patients, and evaluation of symptomatic patients in whom ventricular dysfunction is part of the differential.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [17] Kashou, 2021, Artificial Intelligence-Augmented Electrocardiogram Detection of Left Ventricular Systolic Dysfunction in the General Population, Mayo Clin. Proc. source

      The Methods section models AI-ECG screening in a community population without heart-failure symptoms, and the Discussion describes its potential as a screening tool for preclinical LVSD. The paper does not establish the other proposed role—evaluating symptomatic patients when ventricular dysfunction is in the differential. Its Discussion mentions an earlier clinical ECG study, but does not identify that study’s patients as symptomatic or frame it around that diagnostic use. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [18] Adedinsewo, 2020, Artificial Intelligence-Enabled ECG Algorithm to Identify Patients With Left Ventricular Systolic Dysfunction Presenting to the Emergency Department with Dyspnea, Circ. Arrhythm. Electrophysiol. source

  • May combine two claims.

    Premisecitation

    AI-ECG-based HCM alerts can be deployed in routine care, although positive results still require clinical correlation and confirmatory imaging.

    “More recent real-world implementation studies suggest that AI-ECG-based HCM alerts can be deployed in routine care, although positive results still require clinical correlation and confirmatory imaging [28].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [28] Desai, 2025, Real-World Artificial Intelligence-Based Electrocardiographic Analysis to Diagnose Hypertrophic Cardiomyopathy, JACC Clin. Electrophysiol. source

  • May combine two claims.

    Interpretationcitation

    Routine ECGs may contain information reflecting pulmonary vascular load, right ventricular pressure overload, or secondary electrical changes caused by altered cardiopulmonary hemodynamics.

    “These findings support the possibility that routine ECGs may contain information reflecting pulmonary vascular load, right ventricular pressure overload, or secondary electrical changes caused by altered cardiopulmonary hemodynamics.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [36] Liu, 2022, Artificial Intelligence-Enabled Electrocardiogram Improves the Diagnosis and Prediction of Mortality in Patients with Pulmonary Hypertension, JACC Asia source

      The Results (page not provided) show that ECG leads V1–V3 helped the AI model predict elevated pulmonary artery pressure and that their signals were consistent with right ventricular hypertrophy. This supports the possibility that ECGs contain relevant information, but the paper does not establish that the signals reflect pulmonary vascular load or electrical changes caused by altered cardiopulmonary hemodynamics. (LLM, conf=0.78)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Interpretationcitation

    The pulmonary hypertension literature shows consistent discrimination across multiple cohorts and suggests that AI-ECG may identify disease before formal diagnosis.

    “Overall, the pulmonary hypertension literature shows consistent discrimination across multiple cohorts and suggests that AI-ECG may identify disease before formal diagnosis.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [35] Kwon, 2020, Artificial Intelligence for Early Prediction of Pulmonary Hypertension Using Electrocardiography, J. Heart Lung Transplant. source
    • cites [36] Liu, 2022, Artificial Intelligence-Enabled Electrocardiogram Improves the Diagnosis and Prediction of Mortality in Patients with Pulmonary Hypertension, JACC Asia source
    • cites [37] DuBrock, 2024, An Electrocardiogram-Based AI Algorithm for Early Detection of Pulmonary Hypertension, Eur. Respir. J. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The compressed ECG arrhythmia-classification model variants achieved accuracies of 97.2% to 98.2%.

    “reporting accuracies of 97.2% to 98.2%”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [66] Lee, K.-S., Park, H.-J., Kim, J.E., Kim, H.J., Chon, S., Kim, S., Jang, J., Kim, J.-K., Jang, S., and Gil, Y. (2022). Compressed Deep Learning to Classify Arrhythmia in an Embedded Wearable Device. Sensors, 22. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Recent work evaluating attribution methods for ECG classification has emphasized the need for sanity checks, rigorous validation, and caution when using post-hoc explanations in high-stakes clinical settings.

    “Recent work evaluating attribution methods for ECG classification has emphasized the need for sanity checks, rigorous validation, and caution when using post-hoc explanations in high-stakes clinical settings [50].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [50] Arends, 2026, Signal or Noise? Evaluating Commonly Used Attribution Methods for Explaining Deep Neural Networks in Electrocardiogram Classification, Eur. Heart J.-Digit. Health source

  • May combine two claims.

    Premisecitation

    Surveys of cardiologists suggest generally favorable views regarding the potential of AI to improve efficiency and patient care, but adoption remains uneven because of concerns regarding model validation, transparency, unfamiliarity with AI methods, and trust in automated recommendations.

    “Surveys of cardiologists suggest generally favorable views regarding the potential of AI to improve efficiency and patient care, but adoption remains uneven because of concerns regarding model validation, transparency, unfamiliarity with AI methods, and trust in automated recommendations [45].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [45] Schepart, 2023, Artificial Intelligence-Enabled Tools in Cardiovascular Medicine: A Survey of Current Use, Perceptions, and Challenges, Cardiovasc. Digit. Health J. source

  • May combine two claims.

    Premisecitation

    Patients whose AI-estimated ECG age was more than eight years older than their chronological age had significantly higher mortality, while those whose ECG age was more than eight years younger had lower mortality.

    “Patients whose AI-estimated ECG age was more than eight years older than their chronological age had significantly higher mortality, while those whose ECG age was more than eight years younger had lower mortality.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [7] Lima, 2021, Deep Neural Network-Estimated Electrocardiographic Age as a Mortality Predictor, Nat. Commun. source

  • May combine two claims.

    Premisecitation

    A point-of-care AI-enabled ECG model estimated serum potassium and detected both hypokalemia and hyperkalemia in emergency department patients.

    “Lin et al. developed a point-of-care AI-enabled ECG model to estimate serum potassium and detect both hypokalemia and hyperkalemia in emergency department patients.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [15] Lin, 2022, Point-of-Care Artificial Intelligence-Enabled ECG for Dyskalemia: A Retrospective Cohort Analysis for Accuracy and Outcome Prediction, npj Digit. Med. source

  • Classified as probably not a substantial claim.

    Interpretationcitation

    ECG-age, mortality-prediction, and disease-specific models illustrate how AI-derived outputs may function as quantitative markers of physiologic risk rather than simple diagnostic labels.

    “As discussed above, ECG-age [7], mortality-prediction [8], and disease-specific models illustrate how AI-derived outputs may function as quantitative markers of physiologic risk rather than simple diagnostic labels.”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [7] Lima, 2021, Deep Neural Network-Estimated Electrocardiographic Age as a Mortality Predictor, Nat. Commun. source
    • cites [8] Raghunath, 2020, Prediction of Mortality from 12-Lead Electrocardiogram Voltage Data Using a Deep Neural Network, Nat. Med. source

  • May combine two claims.

    Premisecitation

    In a cohort including more than 34,000 emergency department visits with paired ECG and laboratory potassium measurements, the model achieved AUC values greater than 0.95 for moderate-to-severe hyperkalemia and greater than 0.85 for moderate-to-severe hypokalemia.

    “In a cohort including more than 34,000 emergency department visits with paired ECG and laboratory potassium measurements, the model achieved AUC values greater than 0.95 for moderate-to-severe hyperkalemia and greater than 0.85 for moderate-to-severe hypokalemia.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [15] Lin, 2022, Point-of-Care Artificial Intelligence-Enabled ECG for Dyskalemia: A Retrospective Cohort Analysis for Accuracy and Outcome Prediction, npj Digit. Med. source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Digital-biomarker frameworks emphasize analytical validity, clinical validity, reproducibility, and usefulness within a clearly defined clinical context.

    “More broadly, digital-biomarker frameworks emphasize analytical validity, clinical validity, reproducibility, and usefulness within a clearly defined clinical context [69].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [69] Lieberwirth, 2026, Challenges and Potential of Using Digital Biomarkers in Healthcare and Clinical Trials, Commun. Med. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The compressed ECG arrhythmia-classification model variants had inference times of approximately 150 to 600 ms.

    “inference times of approximately 150 to 600 ms across the model variants evaluated”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [66] Lee, K.-S., Park, H.-J., Kim, J.E., Kim, H.J., Chon, S., Kim, S., Jang, J., Kim, J.-K., Jang, S., and Gil, Y. (2022). Compressed Deep Learning to Classify Arrhythmia in an Embedded Wearable Device. Sensors, 22. source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Other work has used broader explainability frameworks to compare model attention with known cardiologist decision rules and to investigate whether deep learning models use clinically plausible ECG features.

    “Other work has used broader explainability frameworks to compare model attention with known cardiologist decision rules and to investigate whether deep learning models use clinically plausible ECG features [49].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [49] Wagner, 2024, Explaining Deep Learning for ECG Analysis: Building Blocks for Auditing and Knowledge Discovery, Comput. Biol. Med. source

  • May combine two claims.

    Premisecitation

    Conventional ECG interpretation relies on predefined waveform characteristics and rule-based pattern recognition, which may miss subtler physiologic and pathologic signals.

    “Conventional interpretation relies on predefined waveform characteristics and rule-based pattern recognition, which may miss subtler physiologic and pathologic signals [2].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [2] Kligfield, 2007, Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part I: The Electrocardiogram and Its Technology: A Scientific Statement from the American Heart Association Electrocardiography and Arrhythmias Committee, Council on Clinical Cardiology; the American College of Cardiology Foundation; and the Heart Rhythm Society: Endorsed by the International Society for Computerized Electrocardiology, Circulation source

  • May combine two claims.

    Interpretationreasoning

    Such cohorts may introduce verification and referral bias and may not reflect performance in unselected screening populations.

    “Such cohorts may introduce verification and referral bias and may not reflect performance in unselected screening populations.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    AI-ECG performance varies across populations and may be affected by rhythm, conduction abnormalities, disease prevalence, and the threshold selected for a positive result.

    “Nevertheless, performance varies across populations and may be affected by rhythm, conduction abnormalities, disease prevalence, and the threshold selected for a positive result.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    AI-ECG may not only identify existing pulmonary hypertension but may also detect early ECG signatures associated with future disease development.

    “This suggested that AI-ECG may not only identify existing pulmonary hypertension but may also detect early ECG signatures associated with future disease development.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [35] Kwon, 2020, Artificial Intelligence for Early Prediction of Pulmonary Hypertension Using Electrocardiography, J. Heart Lung Transplant. source

  • May combine two claims.

    Interpretationreasoning

    Evidence for HFpEF and diastolic dysfunction remains less mature, with greater heterogeneity in disease definitions and limited prospective evidence that AI-guided detection improves clinical outcomes.

    “Evidence for HFpEF and diastolic dysfunction remains less mature, with greater heterogeneity in disease definitions and limited prospective evidence that AI-guided detection improves clinical outcomes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Reduced-ejection-fraction detection is the most clinically mature AI-ECG application and the closest to routine use.

    “Of the applications reviewed here, reduced-ejection-fraction detection is the most clinically mature and the closest to routine use: the evidence spans large derivation cohorts, external validation, prognostic follow-up, and a pragmatic implementation trial.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    In the EAGLE trial, the intervention increased targeted echocardiography in the AI-positive subgroup without substantially increasing echocardiography use across the entire population.

    “The intervention increased new diagnoses of low ejection fraction within 90 days, particularly among patients with a positive AI screen, and increased targeted echocardiography in the AI-positive subgroup without substantially increasing echocardiography use across the entire population [19].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [19] Yao, 2021, Artificial Intelligence–Enabled Electrocardiograms for Identification of Patients with Low Ejection Fraction: A Pragmatic, Randomized Clinical Trial, Nat. Med. source

  • May combine two claims.

    Premisecitation

    Attia et al.'s CNN identified patients with LVEF ≤ 35% with an AUC of 0.93, sensitivity of 86.3%, and specificity of 85.7%.

    “The model achieved an area under the receiver operating characteristic curve (AUC) of 0.93, with a sensitivity of 86.3% and specificity of 85.7%, a strong performance for a test this cheap and widely available [4].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [4] Attia, 2019, Screening for Cardiac Contractile Dysfunction Using an Artificial Intelligence–Enabled Electrocardiogram, Nat. Med. source

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Variations in the interval between ECG acquisition and blood sampling, treatment administered before testing, renal function, and medication use may affect the relationship between the ECG signal and measured electrolyte concentration.

    “variations in the interval between ECG acquisition and blood sampling, treatment administered before testing, renal function, and medication use may affect the relationship between the ECG signal and measured electrolyte concentration.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    Cost-effectiveness of AI-ECG depends on disease prevalence, alert thresholds, implementation costs, and downstream testing.

    “cost-effectiveness depends on disease prevalence, alert thresholds, implementation costs, and downstream testing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Alert fatigue is a related and underappreciated risk.

    “Alert fatigue is a related and underappreciated risk.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Interpretability creates practical challenges for clinical adoption, since physicians may be hesitant to act on a prediction if they cannot determine whether the model is relying on physiologically meaningful ECG features or on spurious correlations, artifacts, or dataset-specific patterns.

    “This creates practical challenges for clinical adoption, since physicians may be hesitant to act on a prediction if they cannot determine whether the model is relying on physiologically meaningful ECG features or on spurious correlations, artifacts, or dataset-specific patterns.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The FDA qualifies biomarkers for a specific context of use, particularly within medical-product and drug-development programs.

    “In regulatory settings, the FDA similarly qualifies biomarkers for a specific context of use, particularly within medical-product and drug-development programs [70].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [70] U.S. Food and Drug Administration (2026, July 17). About Biomarkers and Qualification, Available online: https://www.fda.gov/drugs/biomarker-qualification-program/about-biomarkers-and-qualification.

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    The EAGLE trial delivered AI results through the electronic health record while clinicians retained responsibility for ordering confirmatory echocardiography.

    “The EAGLE trial, discussed above, provides an early example of this model: AI results were delivered through the electronic health record (EHR) while clinicians retained responsibility for ordering confirmatory echocardiography [19].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] Yao, 2021, Artificial Intelligence–Enabled Electrocardiograms for Identification of Patients with Low Ejection Fraction: A Pragmatic, Randomized Clinical Trial, Nat. Med. source

  • May combine two claims.

    Interpretationnone given

    Scalability of AI-ECG is not automatic and depends on digital waveform storage, compatible ECG systems, software integration, technical support, and an available pathway for responding to positive results.

    “scalability is not automatic and depends on digital waveform storage, compatible ECG systems, software integration, technical support, and an available pathway for responding to positive results.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    A highlighted waveform segment does not necessarily prove that the model is using a clinically valid feature, and different explainability methods may produce different results for the same ECG.

    “A highlighted waveform segment does not necessarily prove that the model is using a clinically valid feature, and different explainability methods may produce different results for the same ECG.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    False positives from AI-ECG screening may lead to additional testing, patient anxiety, and increased resource use.

    “False positives may lead to additional testing, patient anxiety, and increased resource use”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationcitation

    Physician oversight remains essential because symptoms, comorbidities, ECG quality, rhythm, and conduction abnormalities may affect the meaning of a model output.

    “Physician oversight remains essential because symptoms, comorbidities, ECG quality, rhythm, and conduction abnormalities may affect the meaning of a model output.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    Many AI-ECG studies in pulmonary hypertension use echocardiographic estimates of pulmonary artery pressure as the reference standard, even though right heart catheterization remains the diagnostic gold standard.

    “Many AI-ECG studies in pulmonary hypertension use echocardiographic estimates of pulmonary artery pressure as the reference standard, even though right heart catheterization remains the diagnostic gold standard.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Right heart catheterization remains the diagnostic gold standard for pulmonary hypertension.

  • Classified as probably not a substantial claim.

    Interpretationnone given

    Current evidence regarding AI-ECG workflow integration remains limited.

    “Current evidence regarding AI-ECG workflow integration remains limited.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    Reduced spatial information, variable device placement, motion artifact, and hardware differences may alter performance of portable AI-ECG.

    “Reduced spatial information, variable device placement, motion artifact, and hardware differences may alter performance”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    A clinically plausible saliency map does not establish that a model is well calibrated, unbiased, generalizable, or using a causal disease signal.

    “A clinically plausible saliency map does not establish that a model is well calibrated, unbiased, generalizable, or using a causal disease signal.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    The practical value of AI-ECG depends on how the result is delivered, who receives it, how urgent the alert is, and what diagnostic or referral pathway follows.

    “The practical value of AI-ECG therefore depends on how the result is delivered, who receives it, how urgent the alert is, and what diagnostic or referral pathway follows.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    Validation in Kenyan outpatient facilities illustrates that AI-ECG could help prioritize scarce echocardiography and specialist resources and may extend cardiovascular screening beyond tertiary centers in settings where treatable disease is frequently recognized late.

    “Validation in Kenyan outpatient facilities illustrates how AI-ECG could help prioritize scarce echocardiography and specialist resources, and it may extend cardiovascular screening beyond tertiary centers in settings where treatable disease is frequently recognized late [23].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [23] Pandey, 2026, Artificial Intelligence Electrocardiogram and Left Ventricular Systolic Dysfunction in Kenya, JAMA Cardiol. source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Saliency mapping in AI-ECG detection of hypertrophic cardiomyopathy has highlighted regions concentrated primarily in ventricular repolarization and ST-T segments.

    “saliency mapping in AI-ECG detection of hypertrophic cardiomyopathy has highlighted regions concentrated primarily in ventricular repolarization and ST–T segments [48].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] Siontis, 2023, Saliency Maps Provide Insights into Artificial Intelligence-Based Electrocardiography Models for Detecting Hypertrophic Cardiomyopathy, J. Electrocardiol. source

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    The value of AI-ECG screening depends on disease prevalence, the availability of an effective confirmatory test and treatment, the capacity of the health system to evaluate positive results, and the consequences of false-positive and false-negative classifications.

    “This benefit depends on disease prevalence, the availability of an effective confirmatory test and treatment, the capacity of the health system to evaluate positive results, and the consequences of false-positive and false-negative classifications.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Frequent alerts may increase false-positive testing in portable AI-ECG.

    “frequent alerts may increase false-positive testing”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    AI-ECG analysis generates quantitative outputs that may reflect broader physiologic states, including ventricular dysfunction, structural heart disease, metabolic abnormalities, biologic aging, and future mortality risk.

    “AI-ECG analysis expands this role by generating quantitative outputs that may reflect broader physiologic states, including ventricular dysfunction, structural heart disease, metabolic abnormalities, biologic aging, and future mortality risk.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation

    Evidence from reduced-ejection-fraction models and the EAGLE trial suggests that AI-ECG screening can identify subclinical disease and influence diagnostic evaluation.

    “Evidence from reduced-ejection-fraction models and the EAGLE trial suggests that this approach can identify subclinical disease and influence diagnostic evaluation [4,19].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [4] Attia, 2019, Screening for Cardiac Contractile Dysfunction Using an Artificial Intelligence–Enabled Electrocardiogram, Nat. Med. source
    • cites [19] Yao, 2021, Artificial Intelligence–Enabled Electrocardiograms for Identification of Patients with Low Ejection Fraction: A Pragmatic, Randomized Clinical Trial, Nat. Med. source

  • May combine two claims.

    Premisecitation

    The EAGLE trial showed that results delivered through the EHR can increase detection and targeted echocardiography, but it did not establish the optimal alert format, recipient, timing, threshold, or follow-up pathway.

    “The EAGLE trial showed that results delivered through the EHR can increase detection and targeted echocardiography, but it did not establish the optimal alert format, recipient, timing, threshold, or follow-up pathway.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [19] Yao, 2021, Artificial Intelligence–Enabled Electrocardiograms for Identification of Patients with Low Ejection Fraction: A Pragmatic, Randomized Clinical Trial, Nat. Med. source

  • Classified as probably not a substantial claim.

    Interpretationcitation

    Portable and limited-lead technologies may extend non-arrhythmic AI-ECG applications beyond conventional 12-lead acquisition.

    “Portable and limited-lead technologies may extend non-arrhythmic AI-ECG applications beyond conventional 12-lead acquisition.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Successful integration of AI-ECG into clinical practice requires more than strong model performance.

    “Successful integration of AI-ECG analysis into clinical practice requires more than strong model performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationnone given

    Evidence for hyperthyroidism, anemia, and sepsis remains more exploratory, with fewer independent validations and no prospective evidence that AI-ECG screening improves management or outcomes.

    “Evidence for hyperthyroidism, anemia, and sepsis remains more exploratory, with fewer independent validations and no prospective evidence that AI-ECG screening improves management or outcomes.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Interpretability should not be equated with trustworthiness.

    “Interpretability should not be equated with trustworthiness.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    Interpretability is particularly important when AI-ECG is used for conditions that are not traditionally diagnosed by ECG, such as reduced ejection fraction, pulmonary hypertension, hyperkalemia, or mortality risk.

    “Interpretability is particularly important when AI-ECG is used for conditions that are not traditionally diagnosed by ECG, such as reduced ejection fraction, pulmonary hypertension, hyperkalemia, or mortality risk.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based).

    Interpretationnone given

    AI-ECG has the potential to expand the ECG from a conventional diagnostic test into a screening and decision-support platform.

    “AI-ECG has the potential to expand the ECG from a conventional diagnostic test into a screening and decision-support platform.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Compression may introduce trade-offs between model size, inference speed, energy consumption, and diagnostic performance.

    “compression may introduce trade-offs between model size, inference speed, energy consumption, and diagnostic performance”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Attribution maps can be unstable across models, sensitive to preprocessing choices, and may not faithfully represent the true causal basis of a model's prediction.

    “Attribution maps can be unstable across models, sensitive to preprocessing choices, and may not faithfully represent the true causal basis of a model’s prediction.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    False negatives from AI-ECG screening could create inappropriate reassurance if clinicians rely on AI output over clinical judgment.

    “false negatives could create inappropriate reassurance if clinicians rely on AI output over clinical judgment.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Accessibility of AI-ECG depends on more than the availability of the algorithm; a health system may be able to generate an AI-ECG risk score yet lack timely echocardiography, laboratory testing, specialist referral, or treatment capacity to act on the result.

    “A health system may be able to generate an AI-ECG risk score yet lack timely echocardiography, laboratory testing, specialist referral, or treatment capacity to act on the result.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    For deployment decisions, external validation, subgroup performance, calibration, uncertainty estimates, failure-mode testing, and post-deployment monitoring may be more important than a visually convincing explanation.

    “For deployment decisions, external validation, subgroup performance, calibration, uncertainty estimates, failure-mode testing, and post-deployment monitoring may be more important than a visually convincing explanation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Many high-performing AI-ECG models function as black boxes, meaning that the relationship between the input ECG waveform and the final prediction is not easily understood by clinicians.

    “Many high-performing AI-ECG models function as “black boxes,” meaning that the relationship between the input ECG waveform and the final prediction is not easily understood by clinicians.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Models developed in tertiary centers may require local validation or recalibration before use in populations with different disease prevalence, ECG hardware, and referral patterns.

    “Models developed in tertiary centers may also require local validation or recalibration before use in populations with different disease prevalence, ECG hardware, and referral patterns.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    If AI-ECG alerts are too frequent, poorly targeted, or disconnected from actionable next steps, clinicians may ignore them or view them as burdensome.

    “If AI-ECG alerts are too frequent, poorly targeted, or disconnected from actionable next steps, clinicians may ignore them or view them as burdensome.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Explainability methods remain useful for detecting artifacts, comparing model attention with established ECG knowledge, and investigating unexpected errors, but they cannot substitute for prospective clinical validation.

    “Explainability methods remain useful for detecting artifacts, comparing model attention with established ECG knowledge, and investigating unexpected errors, but they cannot substitute for prospective clinical validation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Without clearly assigned responsibility and measurable follow-up, even an accurate model may produce an ignored or clinically unactionable result.

    “Without clearly assigned responsibility and measurable follow-up, even an accurate model may produce an ignored or clinically unactionable result.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Automation bias can occur when clinicians accept an automated recommendation without sufficient independent evaluation, particularly when the system is perceived as highly accurate or simplifies a complex decision.

    “Automation bias can occur when clinicians accept an automated recommendation without sufficient independent evaluation, particularly when the system is perceived as highly accurate or simplifies a complex decision [46].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [46] Abdelwanis, 2024, Exploring the Risks of Automation Bias in Healthcare Artificial Intelligence Applications: A Bowtie Analysis, J. Saf. Sci. Resil. source

  • May combine two claims.

    Interpretationreasoning

    Evidence for AI-ECG in pulmonary hypertension remains less clinically mature than that for reduced ejection fraction because most studies are retrospective, reference standards are inconsistent, and no prospective trial has demonstrated that AI-guided screening reduces diagnostic delay or improves patient outcomes without producing excessive downstream testing.

    “However, the evidence remains less clinically mature than that for reduced ejection fraction because most studies are retrospective, reference standards are inconsistent, and no prospective trial has demonstrated that AI-guided screening reduces diagnostic delay or improves patient outcomes without producing excessive downstream testing.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    ECG patterns associated with PH may overlap with those caused by left-sided heart disease, chronic lung disease, conduction abnormalities, atrial fibrillation, or other causes of right ventricular strain.

    “In addition, ECG patterns associated with PH may overlap with those caused by left-sided heart disease, chronic lung disease, conduction abnormalities, atrial fibrillation, or other causes of right ventricular strain.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Evidence is most persuasive when AI-ECG targets a condition such as reduced ejection fraction, for which a positive screen can lead to echocardiography and established treatment.

    “The evidence is therefore most persuasive when AI-ECG targets a condition such as reduced ejection fraction, for which a positive screen can lead to echocardiography and established treatment.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Federated learning may provide another pathway for expanding AI-ECG development across multiple institutions while limiting the need to centralize sensitive patient data.

    “Federated learning may provide another pathway for expanding AI-ECG development across multiple institutions while limiting the need to centralize sensitive patient data.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    High discrimination does not by itself establish clinical benefit for AI-ECG.

    “However, high discrimination does not by itself establish clinical benefit.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given

    The most mature evidence supports AI-ECG screening for reduced left ventricular ejection fraction, while applications involving structural heart disease, pulmonary hypertension, metabolic abnormalities, systemic disease, and broad prognostic assessment remain at different stages of validation.

    “The most mature evidence supports AI-ECG screening for reduced left ventricular ejection fraction, while applications involving structural heart disease, pulmonary hypertension, metabolic abnormalities, systemic disease, and broad prognostic assessment remain at different stages of validation.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    AI analysis may support earlier detection, triage, and selection for confirmatory testing because ECGs are inexpensive and widely available.

    “Because ECGs are inexpensive and widely available, AI analysis may support earlier detection, triage, and selection for confirmatory testing.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: ECGs are inexpensive and widely available.

  • Classified as probably not a substantial claim.

    Interpretationnone given

    Deep learning has expanded the potential of the 12-lead ECG beyond conventional rhythm and waveform interpretation.

    “Deep learning has expanded the potential of the 12-lead ECG beyond conventional rhythm and waveform interpretation.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Prognostic association alone is insufficient for AI-ECG outputs to be considered clinically useful digital biomarkers.

    “AI-ECG outputs may ultimately function as digital biomarkers, but prognostic association alone is insufficient.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    This approach could allow AI-ECG models to learn from larger and more diverse populations while reducing barriers associated with sharing identifiable health data.

    “This approach could allow AI-ECG models to learn from larger and more diverse populations while reducing barriers associated with sharing identifiable health data.”
    Quote in the paperSubstantial claim
75 of 75

  • The paper relies on this without support.

    Conventional ECG changes in hyperkalemia can be inconsistent and may be absent even when serum potassium is dangerously elevated.

    “Hyperkalemia is known to affect cardiac conduction and repolarization, but conventional ECG changes can be inconsistent and may be absent even when serum potassium is dangerously elevated.”

  • The paper relies on this without support.

    AF may remain intermittent and undetected until complications such as stroke occur.

    “Because AF may remain intermittent and undetected until complications such as stroke occur, AI-ECG could help select patients for ambulatory monitoring or closer rhythm surveillance.”

  • The paper relies on this without support.

    HCM may be underdiagnosed, particularly in patients with subtle or nonspecific ECG findings.

    “This was clinically important because HCM may be underdiagnosed, particularly in patients with subtle or nonspecific ECG findings, and earlier detection can prompt confirmatory imaging, family screening, risk stratification, and disease-specific management.”

  • The paper relies on this without support.

    Cardiac amyloidosis diagnosis is often delayed.

    “Cardiac amyloidosis is another clinically relevant application because diagnosis is often delayed, while earlier recognition has become increasingly important with the availability of disease-modifying therapy for transthyretin amyloid cardiomyopathy.”

  • The paper relies on this without support.

    Delayed recognition of pulmonary hypertension is common.

    “This finding is clinically important because delayed recognition of PH is common, and earlier identification may allow more timely echocardiography, specialist referral, and confirmatory hemodynamic testing when appropriate [37].”

  • The paper relies on this without support.

    Structural heart disorders are relatively uncommon in general populations.

    “Moreover, because these disorders are relatively uncommon in general populations, high discrimination does not necessarily translate into a high positive predictive value.”

  • The paper relies on this without support.

    Right heart catheterization remains the diagnostic gold standard for pulmonary hypertension.

    “Many AI-ECG studies in pulmonary hypertension use echocardiographic estimates of pulmonary artery pressure as the reference standard, even though right heart catheterization remains the diagnostic gold standard.”

  • The paper relies on this without support.

    A risk estimate has limited clinical value unless it leads to a defined and beneficial action.

    “a risk estimate has limited clinical value unless it leads to a defined and beneficial action.”

  • The paper relies on this without support.

    Disease prevalence strongly influences positive predictive value.

    “disease prevalence strongly influences positive predictive value”

  • The paper relies on this without support.

    ECGs are inexpensive and widely available.

    “Because ECGs are inexpensive and widely available”

  • The paper relies on this without support.

    Wearable and portable devices have limited memory, processing capacity, and battery power, making deployment of large deep-learning models challenging.

    “Wearable and portable devices have limited memory, processing capacity, and battery power, making deployment of large deep-learning models challenging.”
11 of 11

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. Sattar, Y., and Chhabra, L. (2026, July 17). Electrocardiogram, StatPearls, Available online: https://www.ncbi.nlm.nih.gov/books/NBK549803/.

Authors
Shaulian, S., Zeltser, R., & Makaryus, A. N.
Affiliations
  • New York Institute of Technology (US)
  • Northwell Health (US)
Corresponding author
Makaryus, Amgad N.
ORCID
Venue
Diagnostics
Publisher
MDPI AG
Published
2026-09-03
Volume
16
Issue
17
Pages
2834
ISSN
2075-4418 (ISSN-L)
DOI
10.3390/diagnostics16172834
PMC
PMC13565729
OpenAlex
W7207520242
Semantic Scholar
36c31abbff96f977ead36f1660b11b1f39969b9a
Cited by
0
References
65
FWCI
0.00
Type
journal article
Language
en
Open access
Licence
Locations
  • Diagnostics (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (submittedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Topics
  • Health Sciences › Medicine › Cardiology and Cardiovascular Medicine › ECG Monitoring and Analysis
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Medicine › Cardiology and Cardiovascular Medicine › Cardiac electrophysiology and arrhythmias
Keywords
12-lead ECG, deep learning, left ventricular ejection fraction, hypertrophic cardiomyopathy, valvular heart disease, cardiac amyloidosis, pulmonary hypertension, hyperkalemia, hyperthyroidism, anemia, sepsis, biological age estimation, mortality prediction, disease screening, clinical decision support, external validation, implementation barriers

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-27.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-27.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review