Pith. sign in

REVIEW 5 major objections 5 minor 39 references

ECG-LENS: Lead-Aware Clinical Context Enriched ECG Report Generation and Evaluation

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ECG-LENS claims a lead-aware encoder plus a clinical prompt writes ECG reports that beat all compared systems on PTB-XL and MIMIC-IV-ECG.

desk verdict A sensible, well-ablated ECG-to-report system whose headline SOTA claim is not yet supported: the baseline comparison is uncontrolled and the abstract's numbers don't match the tables. read the letter →

arxiv 2608.05893 v1 pith:UGXGLEU5 submitted 2026-08-06 cs.AI

Akanta Das , Tasinul Islam Ahon , Ahmed Mahir Sultan Rumi , Md Mahbubur Rahman , Tausif Amim Shadly , Tanzima Hashem This is my paper · ORCID
classification cs.AI
keywords ECGreportgenerationmulti-leadelectrocardiographylead-awareencodingclinicalpromptconditioningF1-ECGBERTmetricPTB-XLMIMIC-IV-ECGGPT-2decoder
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-LENS is an end-to-end system for turning a 12-lead ECG recording into a written clinical report, and the paper's aim is to show that report quality improves when the model never collapses the leads into a single embedding. The architecture keeps one representation per lead plus a global representation of the whole recording, feeds the decoder a short clinical paragraph built from high-confidence diagnostic labels, and trains on reports that were rewritten by GPT-5.5 and checked by a cardiologist. The paper further introduces F1-ECGBERT, a BERT-based scorer that measures whether generated and reference reports contain the same diagnostic labels, on the grounds that lexical metrics miss clinically meaningful paraphrases. The reported result is that ECG-LENS beats every compared baseline on every lexical and diagnostic metric, with absolute gains of 4.0% in METEOR, 6.3% in ROUGE-L, and 11.5% in F1-ECGBERT over the strongest baselines.

What carries the argument

Two mechanisms carry the argument. The generation mechanism is the $(L+1)$-token encoder stack: $L$ lead-wise 1D ResNet-18 encoders produce a per-lead embedding $h_\ell$, a global 1D ResNet-18 processing all 12 leads as channels produces $h_{\mathrm{global}}$, and the stack $H_{\mathrm{ECG}}=(h_1,\ldots,h_L,h_{\mathrm{global}})$ is concatenated with a clinical prompt $p$ built by a frozen ECG classifier's high-confidence labels. The decoder is a GPT-2-style transformer trained from scratch on a domain-specific tokenizer. The evaluation mechanism is F1-ECGBERT: four BERT classifiers, one per diagnostic category, extract binary label vectors from reference and generated reports, and the F1 score between those vectors measures diagnostic preservation rather than surface wording.

What would settle it

Recompute Tables 1-3 with raw PTB-XL reference reports in place of the refined ones, or retrain MEIT, BiECG-LLM, and HeartLLM on the same GPT-5.5-refined targets and rescore them; if the 4.0% METEOR, 6.3% ROUGE-L, and 11.5% F1-ECGBERT advantages shrink to near zero, the claimed lead is an artifact of reference-report mismatch rather than the architecture.

Watch

Extended reading notes

Core claim

In the paper's own terms, the central discovery is that four components compound: a 1D ResNet global encoder beats a Transformer global encoder; report preprocessing alone lifts all metrics; adding per-lead encoders adds more; and adding the clinical prompt produces the final gains (Table 4). The framework's encoder output is a stack of $L=12$ lead embeddings plus one global embedding, which keeps lead-specific findings such as inferior-wall ischemia in leads II, III, and aVF from being diluted before decoding. A frozen MERL classifier supplies diagnostic label confidences, and category-specific thresholds convert only high-confidence predictions into a standardized clinical prompt that conditions a from-scratch GPT-2 decoder. The paper also claims that F1-ECGBERT, trained as four binary BERT classifiers over superclass, subclass, rhythm, and form, captures diagnostic agreement better than lexical metrics, and that ECG-LENS reaches 67% fully clinically correct reports in a cardiologist review.

Load-bearing premise

The claimed superiority over prior systems rests on comparing ECG-LENS, trained on GPT-5.5-refined and cardiologist-checked reports, with published baseline scores from models trained on the raw PTB-XL reports, and the paper never states whether all systems are evaluated against the same reference reports.

Editorial extensions

If this is right

  • If the reported results are correct, ECG-LENS is the strongest published ECG-to-report generator on PTB-XL and MIMIC-IV-ECG across both lexical and diagnostic metrics.
  • A clinically usable report generator can run at about 30 ms per recording on a single consumer GPU, which makes automated drafting feasible in real-time and resource-constrained settings.
  • The F1-ECGBERT recipe, four category-wise BERT label extractors plus F1 agreement, can serve as a model for evaluating other kinds of clinical report generation where paraphrases are clinically equivalent.
  • Because the clinical prompt comes from a frozen classifier that the paper says can be swapped, the framework's gains are tied to having reliable label confidences, not to one specific classifier.
  • The ablation ordering implies that target-report standardization is the largest single contributor among the added components, since preprocessing alone accounts for a large step from row (a) to row (c).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that its headline gains compare a model trained on GPT-5.5-refined targets against published baseline numbers for models trained on raw PTB-XL reports; retraining the baselines on the same refined targets would reveal how much of the 4.0/6.3/11.5 point gap is architectural versus a shared-target effect.
  • A testable extension is to score ECG-LENS against the raw PTB-XL reports rather than the refined ones; if the gains shrink, the clinical prompt and per-lead encoders matter less than target standardization.
  • The 67% fully-correct cardiologist judgment was run on a small random subset; a blinded head-to-head between ECG-LENS, a strong baseline, and human-written reports would give a more direct estimate of practical clinical value.
  • The same lead-aware architecture transfers naturally to other multichannel physiological signals such as EEG or polysomnography, where per-channel morphology and cross-channel dependencies both carry diagnostic meaning.
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 / 5 minor

Summary. The paper proposes ECG-LENS, an end-to-end ECG-to-report generation framework combining lead-wise 1D ResNet-18 encoders with a global encoder, a clinical prompt built from frozen MERL classifier predictions, a GPT-2 decoder trained from scratch, and a GPT-5.5-based report preprocessing pipeline. It also introduces F1-ECGBERT, a BERT-based metric that scores diagnostic-label agreement between generated and reference reports. On PTB-XL in-domain and MIMIC-IV-ECG cross-domain evaluations, the authors claim consistent state-of-the-art results, with absolute gains of 4.0%, 6.3%, and 11.5% in METEOR, ROUGE-L, and F1-ECGBERT over the strongest baselines. A five-configuration ablation study attributes gains to report preprocessing, lead-specific encoders, and clinical prompting. The central claim is an architecture-level improvement over existing ECG report generators.

Significance. If the claimed results were established, ECG-LENS would be a useful contribution: the lead-aware encoding design is well motivated, the use of a frozen classifier to construct diagnostic prompts is a practical way to inject clinical context, and the report preprocessing strategy addresses a real problem of noisy, multilingual training targets. The ablation study is clean and shows consistent incremental gains. The F1-ECGBERT metric is also a reasonable direction for ECG-specific evaluation, extending CheXbert-style ideas to ECG reports. However, the headline state-of-the-art claim is not currently supported because the comparison against baselines is uncontrolled, the abstract's gain magnitudes do not match the tables, and the proposed metric lacks the required validation. The paper's value is therefore conditional on substantial additional evaluation work.

major comments (5)
  1. [Experiments, Baselines and Evaluation Metrics; Tables 1-3] The headline comparison is not controlled. ECG-LENS is trained on GPT-5.5-refined, cardiologist-verified reports (see 'Clinically Focused Report Preprocessing'), while the baseline scores in Tables 1-3 are taken from models trained on raw PTB-XL and MIMIC-IV-ECG reports by their original authors. The manuscript never states whether the reference reports used in Tables 1-3 are raw or refined. If they are refined, every baseline is scored against a distribution it never trained on; if they are raw, ECG-LENS is scored against a distribution different from its training target. In either case, the reported gains conflate the effect of report preprocessing with the effect of the ECG-LENS architecture. To support the state-of-the-art claim, the authors must either retrain all baselines on the identical refined targets and evaluate under the identical reference protocol, or report both raw-reference and refined-reference results for all methods.
  2. [Abstract and Conclusion] The abstract's claimed gains of 4.0%, 6.3%, and 11.5% in METEOR, ROUGE-L, and F1-ECGBERT are not reproduced by the tables. In Table 1, ECG-LENS exceeds BiECG-LLM by 3.2 percentage points in METEOR and 8.5 points in ROUGE-L. In Table 2, the largest F1-ECGBERT gain is 10.9 points on the superclass category. In Table 3, the METEOR gain is 4.0 points but the ROUGE-L gain is only 2.3 points. No table entry yields 6.3% or 11.5%. The authors should correct the abstract and conclusion to report numbers that actually appear in the tables, or specify exactly which table and comparison each claimed gain refers to.
  3. [ECG-Specific Clinical Evaluation and F1-ECGBERT] The F1-ECGBERT metric is trained on the same GPT-5.5-refined report style that ECG-LENS is trained to reproduce. This creates a circularity risk: the metric may reward stylistic alignment with the refined reports rather than clinical truth, and it will systematically favor ECG-LENS over baselines trained on raw reports. The paper states that 'rigorous validation' appears in Supplementary Section 1, but that supplementary material is not included with the manuscript. The authors should report labeler accuracy, agreement with cardiologist judgments, and sensitivity of F1-ECGBERT to report preprocessing and paraphrasing; without such validation, Table 2 cannot be interpreted as evidence of clinical superiority.
  4. [Cross-Domain Evaluation on MIMIC-IV-ECG (Table 3)] The MIMIC-IV-ECG cross-domain protocol is underspecified. The manuscript does not state whether the baseline scores in Table 3 were reproduced by the authors on the same MIMIC-IV-ECG subset and with the same reference reports, or whether they were copied from the original publications. It also does not state whether ECG-LENS was fine-tuned on MIMIC-IV-ECG or evaluated zero-shot, nor how the MIMIC-IV-ECG reference reports were processed. These details are necessary because MIMIC-IV-ECG has no standard report-generation split, and different preprocessing of reference reports can materially change lexical metric scores. The cross-domain claim of 'comparable scores' is therefore not yet supported.
  5. [Cardiologist Evaluation] The cardiologist evaluation reported as '67% fully clinically correct' lacks the information needed to assess its validity: the number of reports evaluated, how they were sampled, whether the cardiologist was blinded to model identity, whether the reference was the raw or refined report, and the criteria for 'fully correct' and 'partially correct' are all omitted. This evaluation is mentioned only as a summary with details deferred to Supplementary Section 5, which is not provided. Given that the paper's central claim is clinical reliability, this evidence should appear in the main text or be made available in full.
minor comments (5)
  1. [Clinical Context-Enriched Prompt Generation] In the sentence describing MERL confidence scores, the phrase 'for each candidate conditioncfor each candidate condition c' is duplicated and contains a typographical artifact; it should be cleaned up.
  2. [Introduction] The sentence beginning 'It suffers fromweak clinical grounding and supervision' is grammatically incomplete and reads as two fragments; it should be rewritten as a complete sentence.
  3. [Experiments, Implementation Details] The category-specific confidence thresholds for the clinical prompt are tuned by grid search on the validation set, but the search ranges, selected thresholds, and their effect on the final results are not reported; please include these details to rule out overfitting.
  4. [Figure 2] The t-SNE visualization is described as showing that ECG-LENS produces 'the most distinguishable clusters,' but no quantitative cluster-quality metric is reported; please add a numerical evaluation such as silhouette score or adjusted Rand index.
  5. [References] Several references are duplicated with different years or venues (e.g., Jin et al. 2025a/2025b, Liu et al. 2024a/2024b, Qiang et al. 2025a/2025b, Yang et al. 2026a/2026b); these should be consolidated into single entries.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ECG-LENS's architecture and training are self-contained; the F1-ECGBERT metric and baseline comparison raise validity questions but do not reduce the derivation to its inputs.

full rationale

ECG-LENS's derivation chain is explicit. Equations (1)-(5) define the lead-wise encoders, global encoder, clinical prompt construction, and the autoregressive training objective. The model minimizes negative log-likelihood over target reports, which are independently produced by GPT-5.5 refinement and cardiologist verification; this is a data-transformation step, not a circular parameter. MERL is introduced as an external, frozen classifier: the paper states 'MERL serves as an external, frozen classifier, its parameters are not updated, and no classification or contrastive objective is included in our training process.' The clinical prompt is therefore not a self-referential component of the derivation. The F1-ECGBERT metric is a separate report-to-label classifier, inspired by CheXbert; the paper does not state that its training data are the GPT-5.5-refined reports, and no equation in the paper makes the metric equivalent to the model's training target. The reported gains in METEOR, ROUGE-L, and F1-ECGBERT are empirical measurements, not quantities derived by definition. The main threats to the state-of-the-art claim are experimental-protocol issues: the paper never states whether Tables 1-3 use raw or refined reference reports, and baseline scores are taken from publications that trained on raw reports. These are correctness and validity concerns, not circularity. No load-bearing self-citations are present. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The evaluation is built on several premises that are plausible but not independently verified in the main text: the fidelity of GPT-5.5 refinement, the reliability of MERL predictions and thresholds, the validity of the new F1-ECGBERT metric, and the comparability of published baseline scores. The first and fourth premises are especially load-bearing because they directly affect the size and meaning of the reported gains.

free parameters (2)
  • MERL category-specific confidence thresholds = Not reported; tuned by grid search on validation set
    Thresholds for superclass, subclass, rhythm, and form decide which predicted conditions enter the clinical prompt; they are fitted to validation data and applied at test time (Prompt Construction).
  • GPT-5.5 few-shot prompt examples = Not quantified; hand-selected pairs of original and refined reports
    The representative examples in the few-shot prompt shape the report-refinement output and therefore the training targets, but they are not described or released (Clinically Focused Report Preprocessing).
assumptions (5)
  • domain assumption PTB-XL diagnostic labels (superclass, subclass, rhythm, form) are correct and complete ground truth.
    These labels train the F1-ECGBERT evaluators and define the diagnostic categories used in prompts; label noise would bias both training and evaluation (Experiments).
  • domain assumption GPT-5.5 report refinement preserves all clinically relevant diagnoses and introduces no new ones.
    Refined reports are used as generation targets; the paper states this is instructed and checked by a cardiologist, but the verification details are not in the main text (Clinically Focused Report Preprocessing).
  • domain assumption MERL's frozen zero-shot predictions are accurate enough, and validation-tuned thresholds generalize to test and MIMIC-IV-ECG.
    The decoder's clinical prompt is built entirely from these predictions; a wrong prompt can push the generated report toward incorrect diagnoses (Clinical Context-Enriched Prompt Generation).
  • ad hoc to paper F1-ECGBERT label extractors provide a valid measure of diagnostic agreement beyond lexical similarity.
    This metric is introduced in the same paper and is used as the headline diagnostic-consistency result, but its validation and human correlation are deferred to a supplementary that is not available here (ECG-Specific Clinical Evaluation).
  • ad hoc to paper Published baseline scores are comparable to ECG-LENS despite different report-preprocessing and training targets.
    Baselines were trained on raw reports, while ECG-LENS was trained on refined reports; the paper does not describe a controlled re-run, so the comparisons assume equivalence (Experiments, Baselines).

how reviews work

0 comments
Cite this review

Pith. "Pith review of ECG-LENS: Lead-Aware Clinical Context Enriched ECG Report Generation and Evaluation." pith.science (2026). https://pith.science/paper/UGXGLEU5

@misc{pith2026260805893,
  author       = {Pith},
  title        = {Pith review of: ECG-LENS: Lead-Aware Clinical Context Enriched ECG Report Generation and Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UGXGLEU5}},
  note         = {Machine review of arXiv:2608.05893}
}
read the original abstract

Electrocardiography (ECG) is one of the most widely used non-invasive tools for diagnosing cardiovascular disease, but transforming multi-lead ECG recordings into reliable clinical reports remains challenging. Automating ECG report generation could reduce clinicians' interpretive workload, improve diagnostic efficiency, and expand access to cardiac assessment in underserved communities. Unlike image-based report-generation tasks, ECG interpretation requires the analysis of subtle temporal morphologies, followed by coherent diagnostic reasoning expressed in dense clinical terminology. Existing systems predominantly focus on classification, while current report-generation methods often produce outputs that remain inadequate for practical clinical use. To address these challenges, we propose ECG-LENS, an end-to-end ECG report-generation framework that jointly integrates multi-lead signal modeling, diagnosis-aware representations, and clinically grounded text generation. ECG-LENS combines lead-wise encoders that preserve localized waveform morphology with a global encoder that captures inter-lead dependencies. To guide report generation, we fuse signal representations with clinically enriched textual prompts that condition a GPT-2 decoder. We further introduce an ECG-specific report-preprocessing strategy that helps the model focus on clinically meaningful findings. Finally, because lexical metrics may under- or overestimate report quality, we propose F1-ECGBERT, a BERT-based, ECG-specific metric that measures agreement between diagnostic labels extracted from generated and reference reports. In-domain experiments on PTB-XL and cross-domain evaluation on MIMIC-IV-ECG show that ECG-LENS consistently outperforms state-of-the-art methods, with absolute gains of 4.0%, 6.3%, and 11.5% in METEOR, ROUGE-L, and F1-ECGBERT, respectively, over the strongest baselines.

Figures

Figures reproduced from arXiv: 2608.05893 by the authors.

Figure 1
Figure 1. Overview of ECG-LENS. Lead-specific and global encoders capture intra and inter-lead context. Predicted diagnostic labels form a structured clinical prompt that guides decoding, while LLM-refined reference reports provide clinically focused training targets. for addressing these challenges because ECGs are multi￾variate time-series signals whose interpretation depends on temporal morphology and relationships across … view at source ↗
Figure 2
Figure 2. t-SNE plots of ECG representations by diagnos￾tic superclasses. ECG-LENS produces the most distinguish￾able clusters. PTB-XL report (ECG id-10453). Ground-truth report Generated report (ECG-LENS) Generated report (HeartLLM) sinus rhythm waves lead III aVF flat inverted aVL consistent ischemia heart disease old inferior myocardial infarction sinus rhythm waves leads III aVF suggesting old inferior myocardial infarcti… view at source ↗
Figure 3
Figure 3. Examples of reports demonstrating accurate lead [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: F1-ECGBERT evaluation. The reference report R and the generated report Rˆ are passed through the BERT evaluator (one per diagnostic category) to extract the appro￾priate labels for the reports.F1 score between the classifica￾tion results measures whether the generated …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 38 canonical work pages

  1. [1]

    and Lungren, Matthew P

    Smit, Akshay and Jain, Saahil and Rajpurkar, Pranav and Pareek, Anuj and Ng, Andrew Y. and Lungren, Matthew P. CheXbert: Combining Automatic Labelers and Expert Annotations for Accurate Radiology Report Labeling Using BERT. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    work page 2020

  2. [2]

    International Conference on Learning Representations , year =

    Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram , author =. International Conference on Learning Representations , year =

    work page

  3. [3]

    Reading Your Heart: Learning

    Jin, Jiarui and Wang, Haoyu and Li, Hongyan and Li, Jun and Pan, Jiahui and Hong, Shenda , booktitle =. Reading Your Heart: Learning. 2025 , url =

    work page 2025

  4. [4]

    Zero-Shot

    Liu, Che and Wan, Zhongwei and Ouyang, Cheng and Shah, Anand and Bai, Wenjia and Arcucci, Rossella , booktitle =. Zero-Shot. 2024 , url =

    work page 2024

  5. [5]

    2025 , month = jul, doi =

    Wan, Zhongwei and Liu, Che and Wang, Xin and Tao, Chaofan and Shen, Hui and Xiong, Jing and Arcucci, Rossella and Yao, Huaxiu and Zhang, Mi , booktitle =. 2025 , month = jul, doi =

    work page 2025

  6. [6]

    2025 , doi =

    Qiang, Yupeng and Dong, Xunde and Liu, Xiuling and Hu, Fei and Wang, Rongjia , booktitle =. 2025 , doi =

    work page 2025

  7. [7]

    2026 , doi =

    Xia, Yong and Li, Jingxuan and Sun, Yeteng and Bu, Jiarui and Wang, Kuanquan , journal =. 2026 , doi =

    work page 2026

  8. [8]

    2026 , doi =

    Yang, Jinning and Sun, Wenjie and Shi, Wen , journal =. 2026 , doi =

    work page 2026

Show all 39 references
  1. [9]

    Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram

    Na, Yeongyeon and Park, Minje and Tae, Yunwon and Joo, Sunghoon. Guiding Masked Representation Learning to Capture Spatio-Temporal Relationship of Electrocardiogram. International Conference on Learning Representations (ICLR)

  2. [10]

    Zero-Shot ECG Classification with Multimodal Learning and Test-Time Clinical Knowledge Enhancement

    Liu, Che and Wan, Zhongwei and Ouyang, Cheng and Shah, Anand and Bai, Wenjia and Arcucci, Rossella. Zero-Shot ECG Classification with Multimodal Learning and Test-Time Clinical Knowledge Enhancement. International Conference on Machine Learning (ICML)

  3. [11]

    Boosting Masked ECG-Text Auto-Encoders as Discriminative Learners

    Pham, Hung Manh and Saeed, Aaqib and Ma, Dong. Boosting Masked ECG-Text Auto-Encoders as Discriminative Learners. International Conference on Machine Learning (ICML)

  4. [12]

    H-Tuning: Toward Low-Cost and Efficient ECG-Based Cardiovascular Disease Detection with Pre-Trained Models

    Zhou, Rushuang and Zhang, Yuan-Ting and Dong, Yining. H-Tuning: Toward Low-Cost and Efficient ECG-Based Cardiovascular Disease Detection with Pre-Trained Models. International Conference on Machine Learning (ICML)

  5. [13]

    Knowledge-Enhanced Multimodal ECG Representation Learning with Arbitrary-Lead Inputs

    Liu, Che and Ouyang, Cheng and Wan, Zhongwei and Wang, Haozhe and Bai, Wenjia and Arcucci, Rossella. Knowledge-Enhanced Multimodal ECG Representation Learning with Arbitrary-Lead Inputs. arXiv:2502.17900

    arXiv

  6. [14]

    Reading Your Heart: Learning ECG Words and Sentences via Pre-Training ECG Language Model

    Jin, Jiarui and Wang, Haoyu and Li, Hongyan and Li, Jun and Pan, Jiahui and Hong, Shenda. Reading Your Heart: Learning ECG Words and Sentences via Pre-Training ECG Language Model. International Conference on Learning Representations (ICLR)

  7. [15]

    MEIT: Multimodal Electrocardiogram Instruction Tuning on Large Language Models for Report Generation

    Wan, Zhongwei and Liu, Che and Wang, Xin and Tao, Chaofan and Shen, Hui and Xiong, Jing and Arcucci, Rossella and Yao, Huaxiu and Zhang, Mi. MEIT: Multimodal Electrocardiogram Instruction Tuning on Large Language Models for Report Generation. Findings of the Association for Co...

    2025

  8. [16]

    BiECG-LLM: An Approach to ECG Classification and Report Generation Using a Fine-Tuned LLM with Bi-Modal ECG

    Qiang, Yao and Dong, Xin and Liu, Xin and Hu, Fang and Wang, Rui. BiECG-LLM: An Approach to ECG Classification and Report Generation Using a Fine-Tuned LLM with Bi-Modal ECG. Proceedings of the IEEE International Conference on Bioinformatics and Biomedicine (BIBM)

  9. [17]

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

    Zhao, Yubao and Kang, Jiaju and Zhang, Tian and Han, Puyu and Chen, Tony. ECG-Chat: A Large ECG-Language Model for Cardiac Disease Diagnosis. Proceedings of the IEEE International Conference on Multimedia and Expo (ICME)

  10. [18]

    ECG-aBcDe: Overcoming Model Dependence, Encoding ECG into a Universal Language for Any LLM

    Xia, Yong and Li, Jian and Sun, Yang and Bu, Jiajun. ECG-aBcDe: Overcoming Model Dependence, Encoding ECG into a Universal Language for Any LLM. arXiv:2509.12625

  11. [19]

    HeartLLM: Discretized ECG Tokenization for LLM Based Diagnostic Reasoning

    Yang, Wei and Sun, Wen and Shi, Wei. HeartLLM: Discretized ECG Tokenization for LLM Based Diagnostic Reasoning. arXiv:2508.15338

  12. [20]

    RadTextAid: A CNN-Guided Framework Utilizing Lightweight Vision-Language Models for Assistive Radiology Reporting

    Nafee, Mahmud Wasif and Aanika, Tasmia Rahman and Hasan, Taufiq. RadTextAid: A CNN-Guided Framework Utilizing Lightweight Vision-Language Models for Assistive Radiology Reporting

  13. [21]

    A Multiresolution Mutual Learning Network for Multi-Label ECG Classification

    Huang, Wei and Wang, Ning and Feng, Panpan and Wang, Haiyan and Wang, Zongmin and Zhou, Bing. A Multiresolution Mutual Learning Network for Multi-Label ECG Classification. arXiv:2406.16928

    arXiv

  14. [22]

    and Johnson, Alistair and Moody, Benjamin and Fernandes, Chrystinne and Greenbaum, Nathaniel and Waks, Jonathan W

    Gow, Brian and Pollard, Tom and Nathanson, Larry A. and Johnson, Alistair and Moody, Benjamin and Fernandes, Chrystinne and Greenbaum, Nathaniel and Waks, Jonathan W. and Eslami, Parastou and Carbonati, Tanner and Chaudhari, Ashish and Herbst, Elizabeth and Moukheiber, Dana an...

  15. [23]

    and Samek, Wojciech and Schaeffter, Tobias

    Wagner, Patrick and Strodthoff, Nils and Bousseljot, Ralf-Dieter and Kreiseler, Dieter and Lunze, Fatima I. and Samek, Wojciech and Schaeffter, Tobias. PTB-XL, a Large Publicly Available Electrocardiography Dataset. Scientific Data

  16. [24]

    and Amaral, Luis A

    Goldberger, Ary L. and Amaral, Luis A. N. and Glass, Leon and Hausdorff, Jeffrey M. and Ivanov, Plamen C. and Mark, Roger G. and Mietus, Joseph E. and Moody, George B. and Peng, Chung-Kang and Stanley, H. Eugene. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Re...

  17. [25]

    and Rajpurkar, Pranav and Haghpanahi, Masoumeh and Tison, Geoffrey H

    Hannun, Awni Y. and Rajpurkar, Pranav and Haghpanahi, Masoumeh and Tison, Geoffrey H. and Bourn, Codie and Turakhia, Mintu P. and Ng, Andrew Y. Cardiologist-Level Arrhythmia Detection and Classification in Ambulatory Electrocardiograms Using a Deep Neural Network. Nature Medicine

  18. [26]

    and Ribeiro, Manoel Horta and Paix \ a o, Gabriela M

    Ribeiro, Antonio H. and Ribeiro, Manoel Horta and Paix \ a o, Gabriela M. M. and Oliveira, Derick M. and Gomes, Paulo R. and Canazart, Jessica A. and Ferreira, Milton P. S. and Andersson, Carl R. and Macfarlane, Peter W. and Meira Jr., Wagner and Sch \"o n, Thomas B. and Ribei...

  19. [27]

    Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL

    Strodthoff, Nils and Wagner, Patrick and Schaeffter, Tobias and Samek, Wojciech. Deep Learning for ECG Analysis: Benchmarks and Insights from PTB-XL. IEEE Journal of Biomedical and Health Informatics

  20. [28]

    and Kaiser, Lukasz and Polosukhin, Illia

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser, Lukasz and Polosukhin, Illia. Attention Is All You Need. Advances in Neural Information Processing Systems (NeurIPS)

  21. [29]

    BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)

    2019

  22. [30]

    Learning Transferable Visual Models from Natural Language Supervision

    Radford, Alec and Kim, Jong Wook and Hallacy, Chris and Ramesh, Aditya and Goh, Gabriel and Agarwal, Sandhini and Sastry, Girish and Askell, Amanda and Mishkin, Pamela and Clark, Jack and Krueger, Gretchen and Sutskever, Ilya. Learning Transferable Visual Models from Natural L...

  23. [31]

    Language Models Are Unsupervised Multitask Learners

    Radford, Alec and Wu, Jeffrey and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya. Language Models Are Unsupervised Multitask Learners. OpenAI Technical Report

  24. [32]

    ROUGE: A Package for Automatic Evaluation of Summaries

    Lin, Chin-Yew. ROUGE: A Package for Automatic Evaluation of Summaries. Text Summarization Branches Out

  25. [33]

    BLEU: A Method for Automatic Evaluation of Machine Translation

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing. BLEU: A Method for Automatic Evaluation of Machine Translation. Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL)

  26. [34]

    METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments

    Banerjee, Satanjeev and Lavie, Alon. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization

  27. [35]

    and Artzi, Yoav

    Zhang, Tianyi and Kishore, Varsha and Wu, Felix and Weinberger, Kilian Q. and Artzi, Yoav. BERTScore: Evaluating Text Generation with BERT. International Conference on Learning Representations (ICLR)

  28. [36]

    Generating Radiology Reports via Memory-Driven Transformer

    Chen, Zhihong and Song, Yan and Chang, Tsung-Hui and Wan, Xiang. Generating Radiology Reports via Memory-Driven Transformer. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2020

  29. [37]

    High-Performance Medicine: The Convergence of Human and Artificial Intelligence

    Topol, Eric J. High-Performance Medicine: The Convergence of Human and Artificial Intelligence. Nature Medicine

  30. [38]

    Circulation , volume=

    Fourth Universal Definition of Myocardial Infarction (2018) , author=. Circulation , volume=

    2018

  31. [39]

    Circulation , volume=

    AHA/ACCF/HRS Recommendations for the Standardization and Interpretation of the Electrocardiogram: Part VI: Acute Ischemia/Infarction , author=. Circulation , volume=

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.