CoChat
Sign Up
Cited by 8ECG Monitoring and AnalysisMachine Learning in HealthcareCardiac electrophysiology and arrhythmias

Diagnosis of cardiac conditions from 12-lead electrocardiogram through natural language supervision

Open access
Zhou, Xue, Li, Tianhui, Hayama, Hiromasa, Nakamura, Keijiro, Liu, Shing-Hong, Chen, Wenxi, Zhu, Xin·2025·npj Digital Medicine
DOI 10.1038/s41746-025-02074-3PMC PMC12630586
Open this paper in CoChat — free accountOpen PDF

Abstract

Current AI approaches for cardiac diagnosis require condition-specific supervised learning with extensive labeled datasets, leading to fundamental scalability barriers. We developed an ECG-CLIP model, applying contrastive multimodal learning to enable zero-shot cardiac diagnosis from 12-lead ECGs using natural language supervision. Trained on 800,034 ECG-text pairs from MIMIC-IV-ECG, ECG-CLIP evaluated 18 cardiac conditions without condition-specific training. The model achieved superior performance for rhythm abnormalities (AUROC > 0.90) compared to morphological conditions. External validation demonstrated robust AUROC rank consistency ( $${\rm{\rho }}$$ = 0.934), including remarkable zero-shot performance for pediatric patients despite no pediatric training cases. Direct comparison showed ECG-CLIP approached supervised models while providing broader diagnostic coverage. Demographic analysis revealed U-shaped age-dependent performance and condition-specific sex-age patterns. By eliminating dependence on labeled data, ECG-CLIP enables diagnosis of various cardiac conditions via text-based queries. This paradigm shift from rigid task-specific models to flexible unified systems addresses critical deployment barriers, potentially expanding global access to expert-level ECG interpretation.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41746-025-02074-3

Cites

4 of 4 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Fairness is not only a model-performance issue; it is also shaped by ECG acquisition quality, care setting, access to confirmatory echocardiography, and linkage to disease-specific treatment.

    “Fairness is not only a model-performance issue; it is also shaped by ECG acquisition quality, care setting, access to confirmatory echocardiography, and linkage to disease-specific treatment.”

    From AI-ECG for Echocardiography Triage in Structural Heart Disease: Evidence, Implementation, and Future Directions (2026) · cites this paper as [60]

    The Discussion (page unavailable) warns that institutional differences in ECG acquisition protocols may limit generalizability. The Methods and Results (page unavailable) assess performance across care settings, with small differences between them. But the paper’s fairness analysis concerns model performance by age and sex; the cited passages do not connect fairness to access to confirmatory echocardiography or disease-specific treatment. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • Uneven performance in gatekeeper use could translate into unequal access to imaging.

    “This issue is especially important for gatekeeper use, where uneven performance could translate into unequal access to imaging.”

    From AI-ECG for Echocardiography Triage in Structural Heart Disease: Evidence, Implementation, and Future Directions (2026) · cites this paper as [60]

    The Results (Demographic stratification analysis; page unavailable) report age-related differences in diagnostic performance, and the Discussion identifies condition-specific sex differences that warrant clinical attention. The paper does not connect those differences to gatekeeper use or unequal access to imaging, so that consequence is not supported by our copy. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Cites

Flagged first, then by citations

Research works printed but not linked: 54

  1. [1]Maron, B. J. et al. Assessment of the 12-lead ECG as a screening test for detection of cardiovascular disease in healthy general populations of young people (12–25 Years of Age): a scientific statement from the American Heart Association and the American College of Cardiology. Circulation 130, 1303–1334 (2014). 25223981 · DOI
  2. [2]Ribeiro, A. H. et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat. Commun. 11, 1760 (2020). 32273514 PMC7145824 · DOI · Paper page
  3. [3]Yagi, R., Mori, Y., Goto, S., Iwami, T. & Inoue, K. Routine electrocardiogram screening and cardiovascular disease events in adults. JAMA Intern. Med. 184, 1035–1044 (2024). 38949831 PMC11217891 · DOI
  4. [4]Brosnan, M. et al. Modest agreement in ECG interpretation limits the application of ECG screening in young athletes. Heart Rhythm 12, 130–136 (2015). 25285648 · DOI
  5. [5]Cook, D. A., Oh, S. & Pusic, M. V. Accuracy of physicians’ electrocardiogram interpretations: a systematic review and meta-analysis. JAMA Intern. Med. 180, 1461–1471 (2020). 32986084 PMC7522782 · DOI
  6. [6]Fye, W. B. Introduction: the origins and implications of a growing shortage of cardiologists. J. Am. Coll. Cardiol. 44, 221–232 (2004). 15282879 · DOI
  7. [7]Attia, Z. I. 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 394, 861–867 (2019). 31378392 · DOI · Paper page
  8. [8]Lee, M. S. et al. Artificial intelligence applied to electrocardiogram to rule out acute myocardial infarction: the ROMIAE multicentre study. Eur. Heart J. ehaf004, 1917–1929 (2025). PMC12093146 39992309 · DOI
  9. [10]Lu, L. et al. Decoding 2.3 million ECGs: interpretable deep learning for advancing cardiovascular diagnosis and mortality risk stratification. Eur. Heart J. Digit. Health 5, 247–259 (2024). 38774384 PMC11104458 · DOI
  10. [11]Benjamens, S., Dhunnoo, P. & Meskó, B. The state of artificial intelligence-based FDA-approved medical devices and algorithms: an online database. NPJ Digit. Med. 3, 118 (2020). 32984550 PMC7486909 · DOI · Paper page
  11. [13]Attia, Z. I., Harmon, D. M., Behr, E. R. & Friedman, P. A. Application of artificial intelligence to the electrocardiogram. Eur. Heart J. 42, 4717–4730 (2021). 34534279 PMC8500024 · DOI
  12. [14]He, J. et al. The practical implementation of artificial intelligence technologies in medicine. Nat. Med. 25, 30–36 (2019). 30617336 PMC6995276 · DOI · Paper page
  13. [15]Vrudhula, A. et al. Impact of case and control selection on training artificial intelligence screening of cardiac amyloidosis. JACC Adv. 3, 100998 (2024). 39372462 PMC11450940 · DOI
  14. [16]Radford, A. et al. Learning transferable visual models from natural language supervision. Proceedings of the 38th International Conference on Machine Learning, PMLR 139, 8748–8763 (2021).
  15. [17]Tiu, E. et al. Expert-level detection of pathologies from unannotated chest X-ray images via self-supervised learning. Nat. Biomed. Eng. 6, 1399–1406 (2022). 36109605 PMC9792370 · DOI · Paper page
  16. [18]Christensen, M., Vukadinovic, M., Yuan, N. & Ouyang, D. Vision–language foundation model for echocardiogram interpretation. Nat. Med. 30, 1481–1488 (2024). 38689062 PMC11108770 · DOI · Paper page
  17. [19]Kligfield, P. Recommendations for the standardization and interpretation of the electrocardiogram: part I. Circulation 115, 1306–1324 (2007). 17322457 · DOI
  18. [20]Balady, G. J. et al. COCATS 4 Task Force 3: training in electrocardiography, ambulatory electrocardiography, and exercise testing. J. Am. Coll. Cardiol. 65, 1763–1777 (2015). 25777646 · DOI
  19. [21]de Hond, A. A. H. et al. Guidelines and quality criteria for artificial intelligence-based prediction models in healthcare: a scoping review. NPJ Digit. Med. 5, 2 (2022). 35013569 PMC8748878 · DOI · Paper page
  20. [24]Goldberger, A. et al. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation 101, e215–e220 (2000). 10851218 · DOI · Paper page
  21. [25]Zheng, J., Guo, H. & Chu, H. A large scale 12-lead electrocardiogram database for arrhythmia study (version 1.0.0). PhysioNet. 10.13026/wgex-er52 (2022). · DOI
  22. [26]Zheng, J. et al. Optimal multi-stage arrhythmia classification approach. Sci. Rep. 10, 2898 (2020). 32076033 PMC7031229 · DOI
  23. [27]Zhou, J., Duan, Y., Chang, Y., Wang, Y. & Lin, C. BELT: bootstrapped EEG-to-language training by natural language supervision. IEEE Trans. Neural Syst. Rehabil. Eng. 32, 3278–3288 (2024). 39190511 · DOI
  24. [28]Ito, A., Dohi, K. & Kawaguchi, Y. CLaSP: learning concepts for time-series signals from natural language supervision. Preprint at https://arxiv.org/abs/2411.08397 (2024). · arXiv
  25. [29]Bleijendaal, H. et al. Clinical applicability of artificial intelligence for patients with an inherited heart disease: a scoping review. Trends Cardiovasc. Med. 33, 274–282 (2023). 35101643 · DOI · Paper page
  26. [30]Katoh, T. et al. Expert consensus document on automated diagnosis of the electrocardiogram: The task force on automated diagnosis of the electrocardiogram in Japan part 3: Inappropriate internal processing in the automatic diagnosis of electrocardiograms and its problems. J. Arrhythmia 40, 17 (2023). PMC10848584 38333400 · DOI
  27. [31]Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A simple framework for contrastive learning of visual representations. ICML 149, 1597–1607 (2020).
  28. [32]Dickinson, D. F. The normal ECG in childhood and adolescence. Heart 91, 1626–1630 (2005). 16287757 PMC1769212 · DOI
  29. [33]Rijnbeek, P. R., Witsenburg, M., Schrama, E., Hess, J. & Kors, J. A. New normal limits for the paediatric electrocardiogram. Eur. Heart J. 22, 702–711 (2001). 11286528 · DOI
  30. [34]Schwartz, P. J. et al. Guidelines for the interpretation of the neonatal electrocardiogram. A task force of the European Society of Cardiology. Eur. Heart J. 23, 1329–1344 (2002). 12269267 · DOI
  31. [35]Somani, S. et al. Deep learning and the electrocardiogram: review of the current state-of-the-art. Europace 23, 1179–1191 (2021). 33564873 PMC8350862 · DOI
  32. [36]Mossie, A. et al. Prevalence and associated factors of preoperative abnormal electrocardiography among older surgical patients in southern Ethiopia: Multicenter cross-sectional study. BMC Geriatr. 24, 905 (2024). 39497046 PMC11533309 · DOI
  33. [37]Rautaharju, P. M. et al. AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram: part IV. J. Am. Coll. Cardiol. 53, 982–991 (2009). 19281931 · DOI
  34. [38]Sharma, S. et al. International recommendations for electrocardiographic interpretation in athletes. Eur. Heart J. 39, 1466–1480 (2018). 28329355 · DOI
  35. [39]Bai, W. A population-based phenome-wide association study of cardiac and aortic structure and function. Nat. Med. 26, 1654–1662 (2020). 32839619 PMC7613250 · DOI
  36. [40]Surawicz, B. & Parikh, S. R. Prevalence of male and female patterns of early ventricular repolarization in the normal ECG of males and females from childhood to old age. J. Am. Coll. Cardiol. 40, 1870–1876 (2002). 12446073 · DOI
  37. [41]Papadakis, M. et al. The prevalence, distribution, and clinical outcomes of electrocardiographic repolarization patterns in male athletes of African/Afro-Caribbean origin. Eur. Heart J. 32, 2304–2313 (2011). 21613263 · DOI
  38. [42]Sandau, K. E. et al. Update to practice standards for electrocardiographic monitoring in hospital settings: a scientific statement from the American Heart Association. Circulation 136, e273–e344 (2017). 28974521 · DOI
  39. [43]Fordyce, C. B. et al. Prevention of complications in the cardiac intensive care unit: a scientific statement from the American Heart Association. Circulation 142, 22 e379–e406 (2020). 33115261 · DOI · Paper page
  40. [44]Magnani, J. W. et al. P wave duration and risk of longitudinal atrial fibrillation in persons ≥60 years old (from the Framingham Heart Study). Am. J. Cardiol. 107, 917–921 (2011). 21255761 PMC3049849 · DOI
  41. [45]Hancock, E. W. et al. AHA/ACCF/HRS recommendations for the standardization and interpretation of the electrocardiogram: Part V. J. Am. Coll. Cardiol. 53, 992–1002 (2009). 19281932 · DOI
  42. [46]Pewsner, D. et al. Accuracy of electrocardiography in diagnosis of left ventricular hypertrophy in arterial hypertension: systematic review. BMJ 335, 711 (2007). 17726091 PMC2001078 · DOI
  43. [47]Kelly, C. J., Karthikesalingam, A., Suleyman, M., Corrado, G. & King, D. Key challenges for delivering clinical impact with artificial intelligence. BMC Med. 17, 195 (2019). 31665002 PMC6821018 · DOI · Paper page
  44. [48]Topol, E. J. High-performance medicine: the convergence of human and artificial intelligence. Nat. Med. 25, 44–56 (2019). 30617339 · DOI · Paper page
  45. [49]Wei, J. et al. Finetuned language models are zero-shot learners. Preprint at10.48550/arXiv.2109.01652 (2022).
  46. [50]Brown, T. B. et al. Language models are few-shot learners. NeurIPS 33, 1877–1901 (2020).
  47. [51]Esteva, A. et al. A guide to deep learning in healthcare. Nat. Med. 25, 24–29 (2019). 30617335 · DOI · Paper page
  48. [52]Shah, N. H., Milstein, A. & Bagley, S. C. Making machine learning models clinically useful. JAMA 322, 1351–1352 (2019). 31393527 · DOI · Paper page
  49. [53]Lee, J. et al. Deep learning for rare disease: a scoping review. J. Biomed. Inform. 135, 104227 (2022). 36257483 · DOI
  50. [55]Strodthoff, N., Lopez Alcaraz, J. M., & Haverkamp, W. MIMIC-IV-ECG-Ext-ICD: Diagnostic labels for MIMIC-IV-ECG (version 1.0.0). PhysioNet. RRID:SCR_00734510.13026/hdyc-1h77 (2024). · DOI
  51. [56]Strodthoff, N., Alcaraz, J. M. L. & Haverkamp, W. Prospects for artificial intelligence-enhanced electrocardiogram as a unified screening tool for cardiac and non-cardiac conditions: an explorative study in emergency care. Eur. Heart J. Digit. Health 5, 454–460 (2024). 39081937 PMC11284007 · DOI · Paper page
  52. [58]Chen, Z. et al. Multi-modal Masked Autoencoders for Medical Vision-and-Language Pre-training. MICCAI, 679-689 (2022).
  53. [59]Vendrow, E. et al. INQUIRE: A natural world text-to-image retrieval benchmark. Adv. Neural Inf. Process. Syst. 37, 126500–126514 (2024).
  54. [60]Cummins, R. O. & Hazinski, M. F. Guidelines based on fear of type II (false-negative) errors: why we dropped the pulse check for lay rescuers. Circulation 102, I377–I379 (2000). 10966683 · DOI

Cited by

Sorted by citations

Authors
Zhou, X., Li, T., Hayama, H., Nakamura, K., Liu, S.-H., Chen, W., & Zhu, X.
Affiliations
  • University of Aizu (JP)
  • Toho University Ohashi Medical Center (JP)
  • Chaoyang University of Technology (TW)
  • Institute of Science Tokyo (JP)
Corresponding author
  • Chen, Wenxi
  • Zhu, Xin
ORCID
Venue
npj Digital Medicine
Publisher
Springer Science and Business Media LLC
Published
2025-11-19
Volume
8
Issue
1
Pages
697-697
ISSN
2398-6352 (ISSN-L)
DOI
10.1038/s41746-025-02074-3
PMC
PMC12630586
OpenAlex
W4416359119
Cited by
8
References
4
Citations by year
2026: 8
FWCI
5.59
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by-nc-nd)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • The University of Aizu (L502)
  • Japan Society for the Promotion of Science (JP21K10287, JP25K1597)
  • Japan Science and Technology Agency (JPMJKB231)
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, heart disease diagnosis, natural language supervision, zero-shot learning, external validation

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.

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