Pith. sign in

REVIEW 5 major objections 4 minor 78 references

ECG-LLM: Foundation Model for ECG-Based Cardiac Reasoning

T0 review · 5 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ECG-LLM, an ECG-conditioned large language model trained on multimodal QA supervision, answers free-text cardiac questions and predicts imaging-derived phenotypes such as ventricular volumes, wall thickness, aortic stenosis, and RV dysfunction from the ECG alone.

desk verdict The unified language-supervision idea is worth engaging, but the synthetic QA pipeline is the hinge: without validation of those pairs, the imaging-phenotype claims stay conditional. read the letter →

arxiv 2607.16323 v1 pith:TLZEO27G submitted 2026-07-15 eess.SP cs.CVcs.LGeess.IV

Alexander Selivanov , Friederike Jungmann , Jan Kehrer , Karl-Ludwig Laugwitz , Eimo Martens , Daniel Rueckert This is my paper · ORCID
classification eess.SPcs.CVcs.LGeess.IV
keywords ecg-llmcardiacmodelphenotypesreasoningcardiovascularclinicalclinically
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

ECG-LLM is an AI model that takes a standard 12-lead ECG trace and a clinical question, then produces a plain-language answer. The model is built by connecting a self-supervised ECG encoder to a large language model (Llama-3.1-8B) and fine-tuning it on more than five million question-answer pairs. These pairs were generated by another AI (GPT-OSS-20B) from structured records that include ECG measurements, cardiologist reports, and imaging-derived phenotypes from cardiac MRI and echocardiography. The training questions are phrased as if a general practitioner is asking about the ECG, and the answers are grounded in the patient's full cardiac record, but at inference the model sees only the ECG and the question.

The key claim is that this training lets the model answer not just standard ECG questions (like rhythm or heart rate) but also questions about cardiac structure and function that are normally assessed by imaging, such as whether the left ventricle is thick, whether aortic stenosis is present, or whether right ventricular function is impaired. On the EchoNext test set, the model achieved F1 scores of 0.77 for increased LV wall thickness, 0.67 for moderate-or-greater aortic stenosis, and 0.61 for right-ventricular systolic dysfunction. On the ECG-QA benchmark it set new state-of-the-art exact-match scores for single-verification and single-choice questions. The paper also reports that the model can generate free-text ECG diagnostic summaries comparable to a strong image-based baseline (PULSE-7B), though evaluation used an AI judge rather than cardiologists.

The work is a step toward using language as a unified supervision space for multimodal cardiac data, potentially enabling ECG-only triage when imaging is unavailable. However, the QA pairs are synthetic and unv

Extended reading notes

Core claim

ECG-LLM successfully recovers conventional ECG measurements, such as heart rate, and strongly predicts complex CMR-derived phenotypes, including ventricular and atrial volumes and ventricular function. Crucially, it detects vital echocardiographic phenotypes, including increased LV wall thickness, aortic stenosis, and right-ventricular systolic dysfunction. (Abstract)

Load-bearing premise

The QA generator (GPT-OSS-20B) produces clinically faithful question-answer pairs from structured cardiac records, so that training on them teaches genuine ECG-to-phenotype associations. Section 4.4 states the generator is prompted to answer 'as if insights come purely from ECG interpretation' and 'never mention anything not present in the provided text,' but there is no validation of the generated pairs against independent expert review; if the generator hallucinates or introduces systematic biases, the model learns those artifacts.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper introduces ECG-LLM, an ECG-conditioned autoregressive LLM that processes raw 12-lead ECG waveforms and answers free-text clinical questions. The model is trained on 679,112 ECG studies from 186,409 patients across UK Biobank, MIMIC-IV-ECG, PTB-XL, and EchoNext, using a 'multimodal-to-language supervision' strategy: structured per-study cardiac records — including ECG measurements, CMR-derived phenotypes, ECHO-derived phenotypes, demographics, and symptoms — are converted by GPT-OSS-20B into over 5 million question–answer pairs. At inference the model receives only the ECG and the question. The authors report strong performance on conventional ECG interpretation, above-majority performance on several CMR- and ECHO-derived phenotypes, improved results on the ECG-QA benchmark relative to published baselines, and competitive free-text report generation as judged by an LLM judge.

Significance. If the central claim holds — that imaging-derived cardiac phenotypes can be inferred from ECG morphology through language supervision — the work would be a meaningful step toward flexible ECG-based triage and would extend ECG-language models beyond report generation. The architectural idea of using natural language as a unified supervision space for heterogeneous cardiac measurements is attractive, and the scale of the instruction-tuning corpus is substantial. The paper also shows concrete strengths: the ECG-QA results (Tables 3–4) are above published baselines on targeted question formats, the report-generation comparison includes several baselines, and the authors transparently discuss several limitations. However, the central claim currently rests on an unvalidated assumption about the fidelity of the synthetic QA supervision, and the evaluation pipeline shares a single LLM family for data generation, answer extraction, and report judging. The manuscript is not yet reproducible because code, model checkpoints, and derived data are not released. These issues are load-bearing for the paper's main conclusion, so I cannot recommend acceptance in the current form.

major comments (5)
  1. [§4.4, Supplementary prompts] The QA-generation prompt instructs the generator to 'use all provided information' from the structured cardiac record while answering 'as if the insights come purely from ECG interpretation,' and to 'never mention anything not present in the provided text.' Because the structured records contain CMR and ECHO values, the generator is effectively required to invent ECG-language explanations for imaging-derived findings. There is no human validation of the generated QA pairs, no assessment of hallucination or systematic bias, and no ablation separating ECG-grounded supervision from record-derived textual supervision. This is the load-bearing assumption for the central claim that ECG-LLM infers imaging phenotypes from ECG morphology. The qualitative EchoNext example in Figure 8 illustrates the risk: the model states that 'the ECG indicates that the pulmonary artery systolic pressure is at le
  2. [§2.1, §2.5, Supplementary §3, §15] There is a same-model circularity in the evaluation. GPT-OSS-20B is used to generate the training QA pairs (Section 4.4), to extract phenotype categories from ECG-LLM free-text answers in the UK Biobank evaluation (Supplementary Section 3), and to judge the quality of generated ECG reports (Section 2.5, Supplementary Section 15). No human validation, inter-rater agreement, or independent judge is reported. This can inflate apparent performance due to stylistic alignment between the generator and the judge. The authors acknowledge that the LLM judge 'should be viewed as a screening and comparison tool rather than a substitute for cardiologist evaluation,' but they still present report-generation scores as a main result. I request a human-annotated subset (e.g., 100–200 reports judged by a cardiologist), inter-rater statistics, and preferably an independent judge model for the report-gener
  3. [§2.3, Tables 1–2] The EchoNext comparison is not apples-to-apples. ECG-LLM is evaluated by parsing its free-text yes/no answers, yielding a single precision/recall/F1 operating point. The Columbia Mini-Model and the supervised ViT are evaluated with AUROC, AUPRC, and F1 at a threshold (presumably the default or an optimized one). Reporting F1 for ECG-LLM and AUROC/AUPRC for the baselines makes the 'highest F1 for 10 of 11 phenotypes' claim threshold-dependent. To support the claimed superiority, the authors should report AUROC/AUPRC for ECG-LLM (e.g., using the probability of the final 'yes' token or a prompt-ensemble score) and, if possible, threshold-matched F1 for all methods.
  4. [§2.2, Figure 2, Supplementary Table 8] The claim that ECG-LLM 'strongly predicts complex CMR-derived phenotypes' is stronger than the evidence. In Supplementary Table 8, several CMR-derived fields have macro-F1 only modestly above the majority baseline (e.g., LV end-diastolic volume 0.410 vs 0.322 majority; LV ejection fraction 0.371 vs 0.191 majority; many fields below 0.5). The authors correctly note that LV wall thickness may be inflated by overlap between CMR and ECHO training targets, but the abstract and introduction do not carry this qualification. I ask for a more measured interpretation, confidence intervals, and an explicit list of phenotypes that exceed a predefined clinically meaningful margin over the majority baseline.
  5. [Code availability, Data availability] The paper states that code, model checkpoints, and derived QA data 'will be available' after publication, with a placeholder GitHub URL. The reader cannot verify the 5M QA pipeline, the preprocessing, the splits, or the experimental results. Given that the central claim depends on a synthetic data-generation pipeline, and given the potential for leakage, the absence of any released artifacts or at least a detailed reproducible protocol is a substantive barrier. I strongly encourage the authors to release the QA-generation prompts, the structured-record definitions, the split files, and the evaluation scripts, or to provide a documented mechanism for accessing them under the data-use agreements.
minor comments (4)
  1. [Throughout] There are several typos and formatting artifacts: 'It’s' should be 'Its' in Section 2.5; 'combing' should be 'combining' in the EchoNext prompts; 'raped ventricular response' in Figure 10 should be 'rapid ventricular response'; and the spacing in 'L V' is inconsistent. The extracted text in Figure 2 appears garbled and should be regenerated.
  2. [§4.3, Table 5] Table 5 shows no validation split for UK Biobank, while Section 4.3 states that a validation split exists overall. Clarify whether UK Biobank contributes to validation and how hyperparameters were selected for that cohort.
  3. [Contributions] The contribution list describes ECG-LLM as 'open-source,' but the code availability section says code and checkpoints will be released 'upon publication' with a placeholder URL. Please reconcile these statements or remove 'open-source' until artifacts are actually available.
  4. [§2.4, Tables 3–4] The ECG-QA comparison is informative, but the baselines are supervised classifiers trained on the ECG-QA answer set, while ECG-LLM is a generative model with a parsing step. This is a meaningful difference; please state explicitly whether the baselines were retrained or taken from the original ECG-QA paper, and whether the parsing step was validated for the exact-match metric.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity in the derivation chain; central phenotype predictions are empirical supervised-learning outputs with held-out evaluation.

full rationale

The claimed derivation chain is: structured ECG/CMR/ECHO records are turned into QA pairs by GPT-OSS-20B (Section 4.4); ECG-LLM is trained with next-token prediction on the ECG waveform paired with these QA targets (Eq. 1); at inference the model receives only the ECG and the question; predictions are compared with held-out imaging labels on subject-disjoint splits (UK Biobank subject-wise split; EchoNext official split). No step in this chain defines the prediction in terms of the target label by construction: imaging phenotype labels appear only in the training answer text, not in the inference input, and the test labels are not used for fitting. The report-generation evaluation uses a GPT-OSS-20B LLM judge (Section 2.5) and the same model family generates the QA supervision; this is a methodological style-matching confound, not a derivation-level circularity, and the paper explicitly states the LLM judge 'should be viewed as a screening and comparison tool rather than a substitute for cardiologist evaluation.' The only self-citation ([26]) is background support for ECG-imaging transfer and is corroborated by external references [13,17-22], so it is not load-bearing. Independent anchors include the ECG-QA benchmark and EchoNext comparison against the external Columbia Mini-Model. The qualitative example in Figure 8 where the model asserts PASP>=45 despite no RVH pattern is a correctness/safety concern, not evidence of circular derivation. Per the scoring rubric, the paper is self-contained against external benchmarks, so a low score is appropriate.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the quality of synthetic QA supervision and the shared-state hypothesis. No new physical entities are postulated. The free parameters are primarily hand-set thresholds and hyperparameters that define the evaluation and training setup.

free parameters (2)
  • Clinical categorization thresholds = multiple (Table 14)
    Hand-defined bin boundaries convert continuous CMR/ECHO measurements into categorical labels that define the evaluation targets; different thresholds would change reported F1.
  • LoRA rank and learning rates = rank 128, lr projector 1e-4, lr LoRA 1e-5
    Chosen by hand; affect what the model learns but are not fitted scientific constants.
assumptions (3)
  • domain assumption ECG waveforms, ECHO, CMR, and clinical labels are different observations of a shared underlying cardiovascular state.
    Stated as central hypothesis in the Introduction; motivates using imaging labels as supervision for ECG-only inference.
  • ad hoc to paper The LLM QA generator accurately converts structured records into faithful QA pairs without hallucination or bias.
    Section 4.4 states the generator is prompted to be faithful, but there is no validation of generated pairs against expert review.
  • domain assumption The MAE-pretrained ECG encoder captures clinically relevant waveform features.
    Section 4.2; the encoder is pretrained on reconstructing masked ECG patches, and its representations are only later validated through downstream tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECG-LLM: Foundation Model for ECG-Based Cardiac Reasoning." pith.science (2026). https://pith.science/paper/TLZEO27G

@misc{pith2026260716323,
  author       = {Pith},
  title        = {Pith review of: ECG-LLM: Foundation Model for ECG-Based Cardiac Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLZEO27G}},
  note         = {Machine review of arXiv:2607.16323}
}
read the original abstract

Electrocardiography (ECG) is an inexpensive, standard-of-care test for cardiac symptoms, but front-line triage often lacks immediate access to definitive imaging such as echocardiography (ECHO) or cardiac magnetic resonance (CMR). Furthermore, most existing ECGAI systems are limited to fixed diagnostic labels or automated reports, constraining their use for patient-specific clinical reasoning. To address this gap, we introduce ECG-LLM, an ECG-conditioned large language model trained across four cohorts comprising 679,112 ECG studies from 186,409 patients. Using a novel multimodal-to-language supervision strategy, ECG-LLM is trained on clinically structured question-answer pairs derived from ECG signals, clinical context, CMR, and ECHO. This unified approach enables the model to answer diverse cardiovascular questions from a 12-lead ECG alone, spanning both conventional interpretation and phenotypes not directly visible on standard ECGs. ECG-LLM successfully recovers conventional ECG measurements, such as heart rate, and strongly predicts complex CMR-derived phenotypes, including ventricular and atrial volumes and ventricular function. Crucially, it detects vital echocardiographic phenotypes, including increased LV wall thickness, aortic stenosis, and right-ventricular systolic dysfunction. On standard ECG understanding tasks, ECG-LLM matches or exceeds existing baselines for diagnostic report generation and the ECG-QA benchmark. By moving beyond fixed-label prediction, this multimodal framework provides clinically valuable, question-driven cardiovascular reasoning to support general practitioner and front-line triage decisions when specialist review is delayed.

Figures

Figures reproduced from arXiv: 2607.16323 by the authors.

Figure 1
Figure 1. Overview of language-mediated multimodal cardiac reasoning. Upper panels: existing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. ECG-only question-answering performance for ECG-derived and CMR-derived phe [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Confusion matrices for the top six ECG-derived and top six CMR-derived phenotype [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Single-Verify ECG-QA exact-match (EM) performance across ECG concepts on PTB [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Single-Verify ECG-QA exact-match (EM) performance across ECG concepts on [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: LLM-judge evaluation of free-text ECG diagnostic summaries across UK Biobank, [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Qualitative ECG-LLM examples from UK Biobank for ECG-derived findings and [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Qualitative ECG-LLM examples from EchoNext for ECG-derived findings and [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Qualitative ECG-LLM examples from MIMIC-IV-ECG and PTB-XL for direct ECG [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]
Figure 10
Figure 10. Figure 10: Clinical QA supervision from cardiac records. (A) ECG, CMR, echocardiography, [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 1 linked inside Pith

  1. [1]

    WHO Cardiovascular-Diseases, 2025

    World Health Organization. WHO Cardiovascular-Diseases, 2025

    2025

  2. [2]

    Global Burden of Cardiovascular Diseases and Risks 2023 Collaborators. Global, regional, and national burden of cardiovascular diseases and risk factors in 204 countries and terri- tories, 1990-2023.Journal of the American College of Cardiology, 86(22):2167–2243, 2025

    2023

  3. [3]

    Sacha Bhatia et al

    R. Sacha Bhatia et al. Electrocardiograms in low-risk patients undergoing an annual health examination.JAMA Internal Medicine, 177(9):1326–1333, Sep 2017

    2017

  4. [4]

    Role of echocardiography in diagnosis and risk stratifi- cation in heart failure with left ventricular systolic dysfunction.Cardiovascular Ultrasound, 5(1):34, Oct 2007

    Quirino Ciampi and Bruno Villari. Role of echocardiography in diagnosis and risk stratifi- cation in heart failure with left ventricular systolic dysfunction.Cardiovascular Ultrasound, 5(1):34, Oct 2007

    2007

  5. [5]

    The role of echocardiography in the diagnosis and management of patients with pulmonary hypertension.Eur

    G Habib and A Torbicki. The role of echocardiography in the diagnosis and management of patients with pulmonary hypertension.Eur. Respir. Rev., 19(118):288–299, Dec 2010

    2010

  6. [6]

    A narrative review of the clinical applications of echocardiography in right heart failure.J

    North J Noelck, Heather A Perry, Phyllis L Talley, and D Elizabeth Le. A narrative review of the clinical applications of echocardiography in right heart failure.J. Clin. Med., 14(15):5505, Aug 2025

    2025

  7. [7]

    Recent advances in cardiovascular magnetic resonance: Techniques and applications.Circ

    Michael Salerno et al. Recent advances in cardiovascular magnetic resonance: Techniques and applications.Circ. Cardiovasc. Imaging, 10(6), Jun 2017

    2017

  8. [8]

    The role of Cardiac Magnetic Resonance (CMR) in the diagnosis of cardiomyopathy: A systematic review

    Henry Anselmo Mayala, Khamis Hassan Bakari, and Zhaohui Wang. The role of Cardiac Magnetic Resonance (CMR) in the diagnosis of cardiomyopathy: A systematic review. Malawi Med. J., 31(3):241–245, Sep 2019

    2019

Show all 78 references
  1. [9]

    Complementary role of echocar- diography and cardiac magnetic resonance in hypertrophic cardiomyopathy.Curr

    Waseem Hindieh, Raymond Chan, and Harry Rakowski. Complementary role of echocar- diography and cardiac magnetic resonance in hypertrophic cardiomyopathy.Curr. Cardiol. Rep., 19(9):81, Sep 2017

    2017

  2. [10]

    Sudlow et al

    C. Sudlow et al. UK Biobank: An open access resource for identifying the causes of a wide range of complex diseases of middle and old age.PLoS Med., 12(3):e1001779, 2015

    2015

  3. [11]

    MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset.Phys- ioNet, Sep 2023

    Brian Gow et al. MIMIC-IV-ECG: Diagnostic Electrocardiogram Matched Subset.Phys- ioNet, Sep 2023. Version 1.0

    2023

  4. [12]

    PTB-XL, a large publicly available electrocardiography dataset.PhysioNet, Nov

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Wojciech Samek, and Tobias Scha- effter. PTB-XL, a large publicly available electrocardiography dataset.PhysioNet, Nov

  5. [13]

    EchoNext: A Dataset for Detecting Echocardiogram- Confirmed Structural Heart Disease from ECGs.PhysioNet, Sep 2025

    Pierre Elias and Joshua Finer. EchoNext: A Dataset for Detecting Echocardiogram- Confirmed Structural Heart Disease from ECGs.PhysioNet, Sep 2025. Version 1.1.0

    2025

  6. [14]

    Hannun et al

    Awni Y. Hannun et al. Cardiologist-level arrhythmia detection and classification in ambu- latory electrocardiograms using a deep neural network.Nature Medicine, 25(1):65–69, Jan 2019. 25

    2019

  7. [15]

    Ribeiro et al

    Antˆ onio H. Ribeiro et al. Automatic diagnosis of the 12-lead ECG using a deep neural network.Nature Communications, 11(1):1760, Apr 2020

    2020

  8. [16]

    Enhancing ECG classification with con- tinuous wavelet transform and multi-branch transformer.Heliyon, 10(5):e26147, 2024

    Chenyang Qiu, Hao Li, Chaoqun Qi, and Bo Li. Enhancing ECG classification with con- tinuous wavelet transform and multi-branch transformer.Heliyon, 10(5):e26147, 2024

    2024

  9. [17]

    Attia et al

    Zachi I. Attia et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram.Nature Medicine, 25(1):70–74, Jan 2019

    2019

  10. [18]

    Sushravya Raghunath et al. Deep Neural Networks Can Predict New-Onset Atrial Fibrilla- tion From the 12-Lead ECG and Help Identify Those at Risk of Atrial Fibrillation–Related Stroke.Circulation, 143(13):1287–1298, 2021

    2021

  11. [19]

    Electrocardiogram screening for aortic valve stenosis using artificial intelligence.Eur

    Michal Cohen-Shelly et al. Electrocardiogram screening for aortic valve stenosis using artificial intelligence.Eur. Heart J., 42(30):2885–2896, Aug 2021

    2021

  12. [20]

    Artificial intelligence-enabled ECG algorithm to identify pa- tients with left ventricular systolic dysfunction presenting to the emergency department with dyspnea.Circ

    Demilade Adedinsewo et al. Artificial intelligence-enabled ECG algorithm to identify pa- tients with left ventricular systolic dysfunction presenting to the emergency department with dyspnea.Circ. Arrhythm. Electrophysiol., 13(8):e008437, Aug 2020

    2020

  13. [21]

    Detection of hypertrophic cardiomyopathy using a convolutional neural network-enabled electrocardiogram.J

    Wei-Yin Ko et al. Detection of hypertrophic cardiomyopathy using a convolutional neural network-enabled electrocardiogram.J. Am. Coll. Cardiol., 75(7):722–733, Feb 2020

    2020

  14. [22]

    Tison et al

    Geoffrey H. Tison et al. Assessment of disease status and treatment response with artificial intelligence-enhanced electrocardiography in obstructive hypertrophic cardiomyopathy.J. Am. Coll. Cardiol., 79(10):1032–1034, Mar 2022

    2022

  15. [23]

    Artificial intelligence-enabled ECG for left ventricular diastolic function and filling pressure.npj Digital Medicine, 7(1):4, Jan 2024

    Eunjung Lee et al. Artificial intelligence-enabled ECG for left ventricular diastolic function and filling pressure.npj Digital Medicine, 7(1):4, Jan 2024

    2024

  16. [24]

    Friedman et al

    Sam F. Friedman et al. Unsupervised deep learning of electrocardiograms enables scalable human disease profiling.npj Digital Medicine, 8(1):23, Jan 2025

    2025

  17. [25]

    Al-Alusi et al

    Mostafa A. Al-Alusi et al. A deep learning digital biomarker to detect hypertension and stratify cardiovascular risk from the electrocardiogram.npj Digital Medicine, 8(1):120, Feb 2025

    2025

  18. [26]

    Global and Local Contrastive Learning for Joint Representations from Cardiac MRI and ECG

    Alexander Selivanov, Philip M¨ uller, ¨Ozg¨ un Turgut, Nil Stolt-Ans´ o, and Daniel Rueckert. Global and Local Contrastive Learning for Joint Representations from Cardiac MRI and ECG . Inproceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, v...

    2025

  19. [27]

    EchoingECG: An Electrocardiogram Cross-Modal Model for Echocardiogram Tasks

    Yuan Gao, Sangwook Kim, and Chris McIntosh. EchoingECG: An Electrocardiogram Cross-Modal Model for Echocardiogram Tasks . Inproceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025, volume 15960 ofLecture Notes in Computer Science, pages 175–185....

    2025

  20. [28]

    Specialized curricula for training vision language models in retinal image analysis.npj Digital Medicine, 8(1):532, Aug 2025

    Robbie Holland et al. Specialized curricula for training vision language models in retinal image analysis.npj Digital Medicine, 8(1):532, Aug 2025

    2025

  21. [29]

    Chexagent: Towards a foundation model for chest x-ray interpretation

    Zhihong Chen et al. Chexagent: Towards a foundation model for chest x-ray interpretation. arXiv preprint arXiv:2401.12208, 2024

    2024 arXiv

  22. [30]

    LLaV A-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day

    Chunyuan Li et al. LLaV A-Med: Training a Large Language-and-Vision Assistant for Biomedicine in One Day. InAdvances in Neural Information Processing Systems, vol- ume 36, pages 28541–28564, 2023. 26

    2023

  23. [31]

    Med-Flamingo: a Multimodal Medical Few-shot Learner

    Michael Moor et al. Med-Flamingo: a Multimodal Medical Few-shot Learner. In Ste- fan Hegselmann, Antonio Parziale, Divya Shanmugam, Shengpu Tang, Mercy Nyamewaa Asiedu, Serina Chang, Tom Hartvigsen, and Harvineet Singh, editors,Proceedings of the 3rd Machine Learning for Healt...

    2023

  24. [32]

    Xiao Gu et al. Cardiac health assessment across scenarios and devices using a multimodal foundation model pretrained on data from 1.7 million individuals.Nature Machine Intel- ligence, 8(2):220–233, Feb 2026

    2026

  25. [33]

    ECG-Chat: A Large ECG-Language Model for Cardiac Disease Diagnosis

    Yubao Zhao, Jiaju Kang, Tian Zhang, Puyu Han, and Tong Chen. ECG-Chat: A Large ECG-Language Model for Cardiac Disease Diagnosis. In2025 IEEE International Confer- ence on Multimedia and Expo (ICME), pages 1–6, 2025

    2025

  26. [34]

    Teaching multimodal LLMs to com- prehend 12-lead electrocardiographic images.npj Digital Medicine, 9(1):349, Mar 2026

    Ruoqi Liu, Yuelin Bai, Xiang Yue, and Ping Zhang. Teaching multimodal LLMs to com- prehend 12-lead electrocardiographic images.npj Digital Medicine, 9(1):349, Mar 2026

    2026

  27. [35]

    ECG-Doctor: An Interpretable Multimodal ECG Diagnosis Frame- work Based on Large Language Models

    Dongsheng Tian et al. ECG-Doctor: An Interpretable Multimodal ECG Diagnosis Frame- work Based on Large Language Models. InProceedings of the 34th ACM International Conference on Information and Knowledge Management, CIKM ’25, pages 2863–2873, New York, NY, USA, 2025. Associati...

    2025

  28. [36]

    GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images

    Xiang Lan et al. GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  29. [37]

    ECG-QA: A comprehensive question answering dataset combined with electrocardiogram.Advances in Neural Information Processing Systems, 36:66277–66288, 2023

    Jungwoo Oh, Gyubok Lee, Seongsu Bae, Joon-myoung Kwon, and Edward Choi. ECG-QA: A comprehensive question answering dataset combined with electrocardiogram.Advances in Neural Information Processing Systems, 36:66277–66288, 2023

    2023

  30. [38]

    Multi-modal understanding and generation for medical images and text via vision-language pre-training.IEEE Journal of Biomedical and Health Informatics, 26(12):6070–6080, 2022

    Jong Hak Moon, Hyungyung Lee, Woncheol Shin, Young-Hak Kim, and Edward Choi. Multi-modal understanding and generation for medical images and text via vision-language pre-training.IEEE Journal of Biomedical and Health Informatics, 26(12):6070–6080, 2022

    2022

  31. [39]

    Multi-modal Masked Autoencoders for Medical Vision-and-Language Pre-training

    Zhihong Chen et al. Multi-modal Masked Autoencoders for Medical Vision-and-Language Pre-training. InMedical Image Computing and Computer Assisted Intervention – MICCAI 2022, volume 13435 ofLecture Notes in Computer Science, pages 679–689, Cham, 2022. Springer Nature Switzerland

    2022

  32. [40]

    Sellergren et al

    Andrew B. Sellergren et al. MedGemma Technical Report, 2025

    2025

  33. [41]

    MedGemma.https://deepmind.google/models/gemma/medgemma/,

    Google DeepMind. MedGemma.https://deepmind.google/models/gemma/medgemma/,

  34. [42]

    google/medgemma-4b-it.https://huggingface.co/google/medgemma-4b-it,

    Google. google/medgemma-4b-it.https://huggingface.co/google/medgemma-4b-it,

  35. [43]

    MedGemma.https://developers.google

    Google Health AI Developer Foundations. MedGemma.https://developers.google. com/health-ai-developer-foundations/medgemma, 2025. Developer documentation

    2025

  36. [44]

    PULSE-7B.https://huggingface.co/PULSE-ECG/PULSE-7B, 2024

    PULSE-ECG. PULSE-7B.https://huggingface.co/PULSE-ECG/PULSE-7B, 2024. Hug- ging Face model card

    2024

  37. [45]

    From local to global: A graph RAG approach to query-focused sum- marization, 2024

    Darren Edge et al. From local to global: A graph RAG approach to query-focused sum- marization, 2024. 27

    2024

  38. [46]

    Fine-tuning and prompt optimization: Two great steps that work better together

    Dilara Soylu, Christopher Potts, and Omar Khattab. Fine-tuning and prompt optimization: Two great steps that work better together. In Yaser Al-Onaizan, Mohit Bansal, and Yun- Nung Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Process...

    2024

  39. [47]

    ECGInstruct.https://huggingface.co/datasets/PULSE-ECG/ ECGInstruct, 2024

    PULSE-ECG. ECGInstruct.https://huggingface.co/datasets/PULSE-ECG/ ECGInstruct, 2024. Hugging Face dataset card

    2024

  40. [48]

    ECG-Image-Kit: a synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization.Physiological Mea- surement, 45(5):055019, May 2024

    Kshama Kodthalu Shivashankara et al. ECG-Image-Kit: a synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization.Physiological Mea- surement, 45(5):055019, May 2024

    2024

  41. [49]

    ECG-Image-Kit: A toolkit for synthesis, analysis, and digitization of electrocardiogram images.https://github.com/alphanumericslab/ecg-image-kit,

    AlphaNumerics Lab. ECG-Image-Kit: A toolkit for synthesis, analysis, and digitization of electrocardiogram images.https://github.com/alphanumericslab/ecg-image-kit,

  42. [50]

    Introducing gpt-oss.https://openai.com/index/introducing-gpt-oss/, 2025

    OpenAI. Introducing gpt-oss.https://openai.com/index/introducing-gpt-oss/, 2025

    2025

  43. [51]

    gpt-oss-20b model card.https://huggingface.co/openai/gpt-oss-20b, 2025

    OpenAI. gpt-oss-20b model card.https://huggingface.co/openai/gpt-oss-20b, 2025

    2025

  44. [52]

    gpt-oss-120b and gpt-oss-20b model card, 2025

    OpenAI. gpt-oss-120b and gpt-oss-20b model card, 2025

    2025

  45. [53]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

    Lianmin Zheng et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. In Advances in Neural Information Processing Systems, volume 36, pages 46595–46623, 2023

    2023

  46. [54]

    LLM-RadJudge: Achieving Radiologist-Level Evaluation for X-Ray Report Generation, 2024

    Zilong Wang, Xufang Luo, Xinyang Jiang, Dongsheng Li, and Lili Qiu. LLM-RadJudge: Achieving Radiologist-Level Evaluation for X-Ray Report Generation, 2024

    2024

  47. [55]

    Evaluation and mitigation of the limitations of large language models in clinical decision-making.Nature Medicine, 30(9):2613–2622, Sep 2024

    Paul Hager et al. Evaluation and mitigation of the limitations of large language models in clinical decision-making.Nature Medicine, 30(9):2613–2622, Sep 2024

    2024

  48. [56]

    MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models. In The Twelfth International Conference on Learning Representations, 2024

    2024

  49. [57]

    Visual Instruction Tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Information Processing Systems, volume 36, pages 34892–34916, 2023

    2023

  50. [58]

    Improved Baselines with Visual Instruction Tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26286–26296, 2024

    2024

  51. [59]

    Masked Autoencoders Are Scalable Vision Learners

    Kaiming He et al. Masked Autoencoders Are Scalable Vision Learners. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15979–15988. IEEE, Jun 2022

    2022

  52. [60]

    Attention is All you Need

    Ashish Vaswani et al. Attention is All you Need. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30, pages 5998–6008. Curran Associates, Inc., 2017

    2017

  53. [61]

    Introducing Llama 3.1: Our most capable models to date.https://ai.meta

    AI at Meta. Introducing Llama 3.1: Our most capable models to date.https://ai.meta. com/blog/meta-llama-3-1/, 2024. 28

    2024

  54. [62]

    Llama 3.1 8B Instruct model card.https://huggingface.co/meta-llama/ Llama-3.1-8B-Instruct, 2024

    Meta AI. Llama 3.1 8B Instruct model card.https://huggingface.co/meta-llama/ Llama-3.1-8B-Instruct, 2024. Model card

    2024

  55. [63]

    Hu et al

    Edward J. Hu et al. LoRA: Low-Rank Adaptation of Large Language Models. InInterna- tional Conference on Learning Representations, 2022

    2022

  56. [64]

    Electrocardiogram baseline removal using wavelet approximations

    David Cuesta Frau, Daniel Nov´ ak, Vladimir Eck, Juan C P´ erez Cort´ es, and Gabriela Andreu Garc ´ ıa. Electrocardiogram baseline removal using wavelet approximations. In Proceeding of the 15th Biennial Eurasip Conference BIOSIGNAL, pages 136–138, 2000

    2000

  57. [65]

    A population-based phenome-wide association study of cardiac and aortic structure and function.Nat

    Wenjia Bai et al. A population-based phenome-wide association study of cardiac and aortic structure and function.Nat. Med., 26(10):1654–1662, Oct 2020

    2020

  58. [66]

    When Can Transformers Count ton?, 2024

    Gilad Yehudai, Haim Kaplan, Asma Ghandeharioun, Mor Geva, and Amir Globerson. When Can Transformers Count ton?, 2024

    2024

  59. [67]

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

    Jason Wei et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. InAdvances in Neural Information Processing Systems, volume 35, pages 24824–24837, 2022

    2022

  60. [68]

    Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes

    Cheng-Yu Hsieh et al. Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. In Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki, editors,Findings of the Association for Computational Linguistics: ACL 2023, pages 8003–...

    2023

  61. [69]

    8-bit Optimizers via Block-wise Quantization

    Tim Dettmers, Mike Lewis, Sam Shleifer, and Luke Zettlemoyer. 8-bit Optimizers via Block-wise Quantization. InInternational Conference on Learning Representations, 2022

    2022

  62. [70]

    Langextract v1.4.0, 2026

    Akshay Goel. Langextract v1.4.0, 2026. Version 1.4.0

    2026

  63. [71]

    Give the final answer to the original question as yes or no

    Google. Langextract.https://github.com/google/langextract, 2025. GitHub reposi- tory. 29 Supplementary Information 1 Dataset statistics Table 5 shows the number of ECG recordings and patients in each dataset split. Table 6 summarizes the final ECG-LLM instruction-tuning corpus...

    2025

  64. [73]

    a challenging ECG-focused clinical question,

  65. [74]

    a concise reasoning field that links the available findings step by step,

  66. [75]

    question

    a final answer that directly addresses the question. - The reasoning must be clinically grounded and directly supported by the provided fields. - The final answer must be concise, direct, and unambiguous. - All reasoning and answers must be directly supported by the provided f...

  67. [76]

    A challenging question that requires step-by-step reasoning

  68. [77]

    Detailed reasoning steps that break down the problem

  69. [78]

    question

    A concise final answer Return JSON format only: 50 [ {{ "question": "Complex question about the text?", "reasoning": "Step 1: First, I need to consider...\nStep 2: Then, I analyze...\nStep 3: Finally, I can conclude...",,→ "answer": "Final answer based on the reasoning." }}, {...

  70. [2025]

    Hugging Face model card

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.