Pith. sign in

REVIEW 4 major objections 3 minor 1 cited by

From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining

T0 review · 4 major / 3 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 Aligning ECGs with text at three time scales - waveform fragments, heartbeats, and full rhythms - yields representations that transfer to new classification tasks better than any prior ECG-language or self-supervised method they compared…

desk verdict A credible multi-scale ECG-language pretraining paper whose reported gains are consistent but whose beat-level contribution is not cleanly isolated; send it to reviewers, not the reject pile. read the letter →

arxiv 2506.21803 v1 pith:I3UYK2KJ submitted 2025-06-11 eess.SP cs.AIcs.LG

Fuying Wang , Jiacheng Xu , Lequan Yu This is my paper
classification eess.SPcs.AIcs.LG
keywords ECGfoundationmodelmulti-scalepretrainingECG-languagealignmentcontrastivelearningzero-shotclassificationself-supervisedheartbeatembeddingclinicalreportcaptioning
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

This paper claims that ECG signals and their paired clinical text reports align better when the alignment happens at three nested time scales instead of one global scale: waveform fragments (tokens), individual heartbeats (beats), and the full 10-second rhythm. The authors build MELP, a pretraining model that supervises all three levels at once, and they report that the resulting ECG representations beat earlier self-supervised and ECG-language baselines: +3.7 average AUC over the strongest prior model in zero-shot classification, first or second in 16 of 18 linear-probing settings, and the best average score in cross-dataset transfer. The point of caring: if true, this is a recipe for getting clinically usable ECG encoders from paired reports without hand-labeled data, especially useful when labeled training data is scarce.

What carries the argument

The load-bearing machinery is the three-loss supervision stack. A rhythm-level InfoNCE contrastive loss ($L_g$) aligns global ECG and text embeddings; a beat-level local contrastive loss ($L_{Local}$) matches attention-pooled latent beat embeddings, built from 10 learnable query tokens, against sentence embeddings of the report via temperature-scaled attention weights over a cosine similarity matrix; and a token-level captioning loss ($L_{LM}$) trains an autoregressive text decoder to regenerate the paired report from attention-pooled ECG tokens. The total objective is $L = L_g + 2\,L_{LM} + 0.2\,L_{Local}$, and the entire pipeline rests on a pre-stage that fine-tunes a biomedical text encoder on a cardiology corpus with masked language modeling.

What would settle it

One decisive check is to retrain MELP with the beat-level contrastive loss applied to deliberately shuffled beat-sentence pairings, breaking local alignment while keeping the statistics identical, and test whether the zero-shot gain over MERL collapses; if beat-level supervision is really doing the work, the misalignment should erase most of the reported +3.7 AUC improvement. A second check is whether the token-level captioning actually reproduces waveform-level phrases such as 'no P waves' or 'PR interval' on held-out reports; if the decoder only emits high-level summaries like 'sinus rhythm', then the token-level supervision is not learning the fine-grained structure the paper claims.

Watch

Extended reading notes

Core claim

The central discovery claim is that hierarchical, multi-scale cross-modal supervision is what makes ECG-language pretraining effective, not just global contrastive alignment. MELP first pretrains a cardiology-specific text encoder, then jointly optimizes three objectives on MIMIC-IV-ECG report pairs: an ECG captioning loss that forces token-level waveform features to reconstruct the clinical report, a beat-level contrastive loss that matches latent heartbeat embeddings against sentence embeddings, and a rhythm-level InfoNCE loss aligning the whole recording with the whole report. Empirically, each level contributes, the full model is the best, and the representation transfers to zero-shot classification, low-data linear probing with 1% of labels, and cross-institution transfer, where it outperforms the strongest prior ECG-language baseline, MERL, by an average of 3.7 AUC across six zero-shot settings.

Load-bearing premise

The method assumes that the paired clinical reports contain enough waveform- and beat-level detail to supervise local alignment, and that 10 latent tokens can represent heartbeat structure without explicit beat segmentation; the paper itself reports that many MIMIC-IV-ECG reports are high-level summaries and that 16 latent tokens outperform 10 by 0.66 AUC.

Editorial extensions

If this is right

  • Zero-shot ECG classification on unseen datasets and label sets improves by an average of 3.7 AUC over the prior ECG-language baseline, with the largest gains on rhythm-level labels (PTBXL-Rhythm, +6.9 AUC).
  • In low-label regimes the multi-scale representation is substantially stronger: at 1% training data MELP beats the second-best method by margins of 2.5-7.4 AUC across six datasets, making it useful where annotated ECGs are scarce.
  • The learned features transfer across institutions and label sets: MELP achieves the best average performance in six cross-domain transfer settings, including +3.76 AUC when linear probing on CPSC2018 and testing on CSN.
  • The fine-grained supervision also enables a secondary capability, ECG report generation, where a 284M-parameter model scores higher than a 7B multimodal baseline on BLEU, ROUGE-L, and other natural-language-generation metrics in a preliminary ECGBench evaluation.
  • Each supervision level contributes independently: the full model beats the best partial-supervision variant by an average of 2.86 AUC in linear probing, and the rhythm-level loss is the largest single contributor since every configuration containing it stays above 82 AUC.

Reading between the lines

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

  • The paper's own beat-count ablation, where 16 latent beats beat 10 by +0.66 AUC while the median real beat count is 12-13, suggests the latent-token design under-parametrizes local structure; an adaptive or R-peak-guided beat tokenization could push the local alignment gain beyond the reported numbers.
  • If the captioning loss is truly driving token-level understanding, then pairing MELP with report augmentation that inserts explicit waveform-level phrases, such as expanding 'sinus rhythm' into its P-wave and PR-interval criteria, should measurably improve downstream fine-grained tasks; this is a testable prediction of the paper's own stated motivation.
  • The architecture bundles several improvements at once, including cardiology text pretraining, a Wav2Vec2 backbone, and captioned contrastive training, and the text-pretraining ablation shows only +0.35 average AUC, so a clean decomposition of which ingredient carries the zero-shot gains remains open; readers should attribute the headline number to the package rather than to any single component.
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

4 major / 3 minor

Summary. The paper introduces MELP, a multi-scale ECG-language pretraining framework. MELP first pretrains a cardiology-specific text encoder, then aligns 12-lead ECG recordings with clinical reports at three levels: token-level captioning, beat-sentence contrastive alignment, and global rhythm-level contrastive alignment. The method is trained on MIMIC-IV-ECG and evaluated on PTB-XL, CPSC2018, and CSN for zero-shot classification, linear probing, cross-domain transfer, and ECG report generation. The paper reports consistent improvements over MERL and several SSL baselines, and includes ablations of the three loss components, the ECG encoder backbone, text pretraining, and hyperparameters. Code is made publicly available.

Significance. If the empirical results hold, MELP is a meaningful advance in ECG-language pretraining: it is among the first to organize cross-modal supervision around clinically motivated scales, and the reported gains are consistent across three public benchmarks. The paper's strengths include the release of code, comparison against a broad set of baselines, and ablations of the three supervision levels. The main limitations are the absence of uncertainty estimates, an ablation table whose per-dataset gains do not uniformly support the claim that every supervision level is necessary, and inconsistencies in the reported dataset statistics. These issues weaken but do not invalidate the central claim; they need to be addressed before the paper is accepted.

major comments (4)
  1. [§4.3.1, Tables 5 and 6] The claim that hierarchical supervision drives the reported gains is not cleanly supported by the ablations. In Table 5, the Lg-only model already reaches 94.98 AUROC at 100% data on PTBXL-Rhythm, and the full-model gain is only 1.93 points; on PTBXL-Form at 100%, the LLocal-only row reaches 81.41 versus 83.30 for the full model. In the zero-shot ablation (Table 6), the LLM+LLocal row (76.4 average) is below the Lg-only row (76.9), and the full model's gain over Lg-only (79.0 versus 76.9) is not uniform across datasets. Because no repeated-seed standard deviations or significance tests are reported, the claimed 2.86% and 2.1% average improvements could be within run-to-run noise. Please add error bars and a per-variant significance analysis, or temper the conclusion that each supervision level is individually necessary.
  2. [§3.3, Eqs. (2)-(3)] Equations (2) and (3) contain index errors. Equation (2) writes \hat{B}^{(l)} = \sum_{l=1}^{N_B} \alpha_l S^{(l)}, but the right-hand side should sum over beat indices j and use beat embeddings B^{(j)}; as written, the attention-weighted beat representation is not well defined. Equation (3) has numerator \exp(\langle S^{(l)}, B^{(l)}\rangle/\tau_1) while the denominator sums over B^{(j)}, making the softmax weights inconsistent. Please rewrite these equations with distinct indices for the sentence (l), the beat being attended (j), and the query sentence.
  3. [Table 1 and Appendix D] The dataset statistics are contradictory. Table 1 lists PTBXL-Rhythm as 16,832/2,100/2,098 train/validation/test and CPSC2018 as 4,950/551/1,376, while Appendix D (Table 16) lists PTB-XL as 12,978/1,642/1,652 total and CPSC2018 as 8,958/1,303/2,598. The CSN numbers also differ (16,546/1,860/4,620 in Table 1 versus 7,651/851/2,126 in Appendix D). Since downstream results are reported on these splits, please reconcile these numbers or explicitly state which split protocol is used for each table.
  4. [§3.3, Appendix C.2] The default number of learnable beat tokens is acknowledged to be suboptimal. Appendix C.2 shows that 16 tokens outperform the default 10 by +0.66 average AUROC on linear probing. Because the beat-level loss is one of the three claimed contributions, the paper should either adopt the better-tuned configuration or justify reporting all main results with the suboptimal setting. This is especially relevant to the conclusions drawn from the ablation in Table 5.
minor comments (3)
  1. [Abstract and Section 1] The abstract typesets MELP as 'Multi-scaleECG-LanguagePretraining' without spaces; please fix the formatting.
  2. [Appendix A.1] Appendix A.1 contains a duplicated sentence ('To ensure content relevance, we only employ abstract section for PudMed dataset.') and the typo 'PudMed' instead of 'PubMed'.
  3. [Table 4] Table 4's header 'Zero-shot Training Ratio' and the placeholder-looking entries (e.g., '%100%') appear corrupted in the preprint and should be reformatted for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; MELP's gains are empirical evaluations on held-out public benchmarks with standard SSL objectives.

full rationale

MELP's central claim is empirical: pretraining with three cross-modal losses on MIMIC-IV-ECG pairs and then transferring to held-out PTB-XL, CPSC2018, and CSN benchmarks. The losses (captioning LLM, local contrastive LLocal, and global InfoNCE Lg) are standard SSL objectives whose weighted sum is optimized, not quantities derived from the evaluation metric. No parameter is fitted to the downstream test sets: the loss weights lambda1 and lambda2 and the beat-token count are chosen on validation/zero-shot validation (Tables 10 and 15), and the paper explicitly reports that its default 10 beat tokens is suboptimal relative to 16 (+0.66 AUC), which cuts against any claim that the configuration was reverse-engineered to the reported results. The text encoder is domain-adapted on MIMIC-IV-ECG reports, but the downstream zero-shot prompts come from GPT-4 following MERL and the linear-probing labels come from independent benchmarks, so this is standard domain pretraining rather than circularity. There is no uniqueness theorem, no load-bearing self-citation chain, and the token/beat/rhythm decomposition is a design choice rather than a renaming of the evaluation result. The only oddity is a likely typographical slip in Eq. (2), where the attention-weighted beat embedding is printed with S(l) on the right-hand side; taken literally this would make the local loss text-self-similar, but the surrounding text and Eq. (3) make the intended B(j) weighting clear, and this is an editing error, not a circular step. Overall, the derivation is self-contained against external benchmarks, so no circularity is present.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central empirical claim relies on standard SSL objectives plus a small set of domain assumptions about ECG-report alignment. The main assumptions are that MIMIC-IV-ECG text provides local supervision and that latent beat pooling is a faithful substitute for explicit heartbeat segmentation. The free parameters are loss weights, temperatures, and the beat-token count; the latter is demonstrably suboptimal. No new physical entities are introduced.

free parameters (4)
  • loss weights lambda1 and lambda2 = lambda1 = 2, lambda2 = 0.2
    Selected by a preliminary search on zero-shot validation performance (Sec 4.3.5); only four configurations are reported, so the optimum is weakly identified.
  • number of learnable beat tokens = 10, with 16 better (+0.66% AUC) in Appendix C.2
    Chosen because the authors assumed most 10-second ECGs contain 10 beats; Table 14 shows median beat count is 12-13 and 16 tokens performs best.
  • temperatures tau1 and tau2 = tau1 = 0.25, tau2 = 0.1
    Set by hand in Sec 3.3 and not ablated; they directly affect the contrastive alignment sharpness.
  • attention pooler query tokens for ECG = 128
    Chosen in Sec 3.3 without ablation; this architectural choice determines the capacity of the token-level summary.
assumptions (5)
  • standard math InfoNCE and masked language modeling objectives provide valid gradients for representation alignment.
    Used in Eqs. (1), (5), and (6); this is the standard assumption from Oord et al. 2018 and Devlin 2018.
  • domain assumption MIMIC-IV-ECG text reports contain enough waveform-level descriptions to supervise token- and beat-level alignment.
    Core motivation in Sec 3.3 and Appendix B; the paper itself notes many reports lack explicit local descriptions and calls this a limitation.
  • domain assumption Attention pooling over 10 learnable tokens can summarize heartbeats without explicit R-peak segmentation.
    Introduced in Sec 3.3 Beat view; Appendix C.2 shows the chosen count is suboptimal and median beat counts are 12-13, weakening this assumption.
  • domain assumption Sentence-level text embeddings obtained by averaging word tokens are a good match for beat-level ECG embeddings.
    Used in Sec 3.3 Beat view, Eqs. (3) to (4); no validation that sentence boundaries in reports align with cardiac beats.
  • standard math Evaluation on held-out public datasets measures generalization of the pretrained representations.
    Standard empirical evaluation, but the consistency of splits is undermined by conflicting statistics in Table 1 versus Appendix D.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining." pith.science (2026). https://pith.science/paper/I3UYK2KJ

@misc{pith2026250621803,
  author       = {Pith},
  title        = {Pith review of: From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I3UYK2KJ}},
  note         = {Machine review of arXiv:2506.21803}
}
read the original abstract

Electrocardiograms (ECGs) play a vital role in monitoring cardiac health and diagnosing heart diseases. However, traditional deep learning approaches for ECG analysis rely heavily on large-scale manual annotations, which are both time-consuming and resource-intensive to obtain. To overcome this limitation, self-supervised learning (SSL) has emerged as a promising alternative, enabling the extraction of robust ECG representations that can be efficiently transferred to various downstream tasks. While previous studies have explored SSL for ECG pretraining and multi-modal ECG-language alignment, they often fail to capture the multi-scale nature of ECG signals. As a result, these methods struggle to learn generalized representations due to their inability to model the hierarchical structure of ECG data. To address this gap, we introduce MELP, a novel Multi-scale ECG-Language Pretraining (MELP) model that fully leverages hierarchical supervision from ECG-text pairs. MELP first pretrains a cardiology-specific language model to enhance its understanding of clinical text. It then applies three levels of cross-modal supervision-at the token, beat, and rhythm levels-to align ECG signals with textual reports, capturing structured information across different time scales. We evaluate MELP on three public ECG datasets across multiple tasks, including zero-shot ECG classification, linear probing, and transfer learning. Experimental results demonstrate that MELP outperforms existing SSL methods, underscoring its effectiveness and adaptability across diverse clinical applications. Our code is available at https://github.com/HKU-MedAI/MELP.

Figures

Figures reproduced from arXiv: 2506.21803 by the authors.

Figure 1
Figure 1. Illustration of the multi-scale view of ECG signals. Rhythm Level: Captures the full ECG recording, reflecting the heart’s global electrical activity over time. Beats Level: Segments each rhythm into discrete heartbeat tokens, isolating individual cardiac cycles for localized analysis. Token Level: Further de￾composes each heartbeat into finer-grained temporal components, enabling granular feature extraction. 1. Int… view at source ↗
Figure 2
Figure 2. Overview of MELP: MELP incorporates three levels of supervision—token, beat, and rhythm—to guide ECG-language pretraining. At the token level, fine-grained ECG representations serve as queries for a transformer decoder, which reconstructs paired ECG reports using a captioning loss (LLM). At the beat level, token-level ECG features are aggregated into beat-level representations via an attention pooling layer, while t… view at source ↗
Figure 3
Figure 3. Comparison of T-SNE visualizations of the embedding space for MERL and MELP on the CSN test set. 13 out of 18 settings, achieving an average improvement of 0.35%. These results suggest that text pretraining en￾hances downstream generalization by improving feature separation in the joint embedding space. They also high￾light the potential benefits of developing more specialized cardiology-specific language models to … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: illustrates the overall training workflow of our cardiology-specific language model pretraining. We initialize our model using the query encoder from MedCPT (Jin et al., 2023) 3 , which was originally trained on PubMed search logs. To better adapt it to the cardiology …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Across 14 sensor generation settings, flow-matching models are the strongest overall baseline, while demographic conditioning, time-frequency modeling, and moderate synthetic augmentation improve hard regimes and down...

Reference graph

Works this paper leans on

52 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    wav2vec 2.0: A framework for self-supervised learning of speech representations

    Baevski, A., Zhou, Y., Mohamed, A., and Auli, M. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33: 0 12449--12460, 2020

    work page 2020

  2. [2]

    C., Schwaighofer, A., Hyland, S., Wetscherek, M., Naumann, T., Nori, A., Alvarez-Valle, J., et al

    Boecking, B., Usuyama, N., Bannur, S., Castro, D. C., Schwaighofer, A., Hyland, S., Wetscherek, M., Naumann, T., Nori, A., Alvarez-Valle, J., et al. Making the most of text semantics to improve biomedical vision--language processing. In European conference on computer vision, pp.\ 1--21. Springer, 2022

    work page 2022

  3. [3]

    Emerging properties in self-supervised vision transformers

    Caron, M., Touvron, H., Misra, I., J \'e gou, H., Mairal, J., Bojanowski, P., and Joulin, A. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 9650--9660, 2021

    2021

  4. [4]

    A simple framework for contrastive learning of visual representations

    Chen, T., Kornblith, S., Norouzi, M., and Hinton, G. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PMLR, 2020

    2020

  5. [5]

    and He, K

    Chen, X. and He, K. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 15750--15758, 2021

    work page 2021

  6. [6]

    An empirical study of training self-supervised vision transformers

    Chen, X., Xie, S., and He, K. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 9640--9649, 2021

    2021

  7. [7]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Devlin, J. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018

    arXiv 2018

  8. [8]

    Unified language model pre-training for natural language understanding and generation

    Dong, L., Yang, N., Wang, W., Wei, F., Liu, X., Wang, Y., Gao, J., Zhou, M., and Hon, H.-W. Unified language model pre-training for natural language understanding and generation. Advances in neural information processing systems, 32, 2019

    work page 2019

Show all 52 references
  1. [9]

    A review on deep learning methods for ecg arrhythmia classification

    Ebrahimi, Z., Loni, M., Daneshtalab, M., and Gharehbaghi, A. A review on deep learning methods for ecg arrhythmia classification. Expert Systems with Applications: X, 7: 0 100033, 2020

    2020

  2. [10]

    K., Li, X., and Guan, C

    Eldele, E., Ragab, M., Chen, Z., Wu, M., Kwoh, C. K., Li, X., and Guan, C. Time-series representation learning via temporal and contextual contrasting. arXiv preprint arXiv:2106.14112, 2021

    2021 arXiv

  3. [11]

    3kg: Contrastive learning of 12-lead electrocardiograms using physiologically-inspired augmentations

    Gopal, B., Han, R., Raghupathi, G., Ng, A., Tison, G., and Rajpurkar, P. 3kg: Contrastive learning of 12-lead electrocardiograms using physiologically-inspired augmentations. In Machine Learning for Health, pp.\ 156--167. PMLR, 2021

    2021

  4. [12]

    A., Johnson, A., Moody, B., Fernandes, C., Greenbaum, N., Berkowitz, S., Moukheiber, D., Eslami, P., et al

    Gow, B., Pollard, T., Nathanson, L. A., Johnson, A., Moody, B., Fernandes, C., Greenbaum, N., Berkowitz, S., Moukheiber, D., Eslami, P., et al. Mimic-iv-ecg-diagnostic electrocardiogram matched subset. Type: dataset, 2023

    2023

  5. [13]

    Bootstrap your own latent-a new approach to self-supervised learning

    Grill, J.-B., Strub, F., Altch \'e , F., Tallec, C., Richemond, P., Buchatskaya, E., Doersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33: ...

    2020

  6. [14]

    Gwon, H., Seo, J., Park, S., Kim, Y.-H., and Jun, T. J. Medical language model specialized in extracting cardiac knowledge. Scientific Reports, 14 0 (1): 0 29059, 2024

    2024

  7. [15]

    Foundation models in electrocardiogram: A review

    Han, Y., Liu, X., Zhang, X., and Ding, C. Foundation models in electrocardiogram: A review. arXiv preprint arXiv:2410.19877, 2024

    2024

  8. [16]

    Momentum contrast for unsupervised visual representation learning

    He, K., Fan, H., Wu, Y., Xie, S., and Girshick, R. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 9729--9738, 2020 a

    2020

  9. [17]

    Deberta: Decoding-enhanced bert with disentangled attention

    He, P., Liu, X., Gao, J., and Chen, W. Deberta: Decoding-enhanced bert with disentangled attention. arXiv preprint arXiv:2006.03654, 2020 b

    2006 arXiv

  10. [18]

    Spatiotemporal self-supervised representation learning from multi-lead ecg signals

    Hu, R., Chen, J., and Zhou, L. Spatiotemporal self-supervised representation learning from multi-lead ecg signals. Biomedical Signal Processing and Control, 84: 0 104772, 2023

    2023

  11. [19]

    Reading your heart: Learning ECG words and sentences via pre-training ECG language model

    Jin, J., Wang, H., Li, H., Li, J., Pan, J., and Hong, S. Reading your heart: Learning ECG words and sentences via pre-training ECG language model. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=6Hz1Ko087B

    2025

  12. [20]

    C., Yeganova, L., Wilbur, W

    Jin, Q., Kim, W., Chen, Q., Comeau, D. C., Yeganova, L., Wilbur, W. J., and Lu, Z. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search logs for zero-shot biomedical information retrieval. Bioinformatics, 39 0 (11): 0 btad651, 2023

    2023

  13. [21]

    Kiyasseh, D., Zhu, T., and Clifton, D. A. Clocs: Contrastive learning of cardiac signals across space, time, and patients. In International Conference on Machine Learning, pp.\ 5606--5615. PMLR, 2021

    2021

  14. [22]

    LB, M. Ccs atrial fibrillation guidelines committee: Canadian cardiovascular society atrial fibrillation guidelines 2010: Prevention and treatment of atrial fibrillation following cardiac surgery. Can J Cardiol, 27: 0 91--97, 2011

    2010

  15. [23]

    Frozen language model helps ecg zero-shot learning

    Li, J., Liu, C., Cheng, S., Arcucci, R., and Hong, S. Frozen language model helps ecg zero-shot learning. In Medical Imaging with Deep Learning, pp.\ 402--415. PMLR, 2024

    2024

  16. [24]

    Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement

    Liu, C., Wan, Z., Ouyang, C., Shah, A., Bai, W., and Arcucci, R. Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement. arXiv preprint arXiv:2403.06659, 2024 a

    2024 arXiv

  17. [25]

    An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detection

    Liu, F., Liu, C., Zhao, L., Zhang, X., Wu, X., Xu, X., Liu, Y., Ma, C., Wei, S., He, Z., et al. An open access database for evaluating the algorithms of electrocardiogram rhythm and morphology abnormality detection. Journal of Medical Imaging and Health Informatics, 8 0 (7): 0...

    2018

  18. [26]

    Teach multimodal llms to comprehend electrocardiographic images

    Liu, R., Bai, Y., Yue, X., and Zhang, P. Teach multimodal llms to comprehend electrocardiographic images. arXiv preprint arXiv:2410.19008, 2024 b

    2024 arXiv

  19. [27]

    A., and Brady, W

    Mattu, A., Tabas, J. A., and Brady, W. J. Electrocardiography in emergency, acute, and critical care. American College of Emergency Physicians, 2019

    2019

  20. [28]

    Ecg-fm: An open electrocardiogram foundation model

    McKeen, K., Oliva, L., Masood, S., Toma, A., Rubin, B., and Wang, B. Ecg-fm: An open electrocardiogram foundation model. arXiv preprint arXiv:2408.05178, 2024

    2024 arXiv

  21. [29]

    Guiding masked representation learning to capture spatio-temporal relationship of electrocardiogram

    Na, Y., Park, M., Tae, Y., and Joo, S. Guiding masked representation learning to capture spatio-temporal relationship of electrocardiogram. arXiv preprint arXiv:2402.09450, 2024

    2024 arXiv

  22. [30]

    H., Sinthong, P., and Kalagnanam, J

    Nie, Y., Nguyen, N. H., Sinthong, P., and Kalagnanam, J. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730, 2022

    2022 arXiv

  23. [31]

    Lead-agnostic self-supervised learning for local and global representations of electrocardiogram

    Oh, J., Chung, H., Kwon, J.-m., Hong, D.-g., and Choi, E. Lead-agnostic self-supervised learning for local and global representations of electrocardiogram. In Conference on Health, Inference, and Learning, pp.\ 338--353. PMLR, 2022

    2022

  24. [32]

    Oord, A. v. d., Li, Y., and Vinyals, O. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    2018 arXiv

  25. [33]

    C-melt: Contrastive enhanced masked auto-encoders for ecg-language pre-training

    Pham, M., Saeed, A., and Ma, D. C-melt: Contrastive enhanced masked auto-encoders for ecg-language pre-training. arXiv preprint arXiv:2410.02131, 2024

    2024 arXiv

  26. [34]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PMLR, 2021

    2021

  27. [35]

    J., Wang, Z., Oikonomou, E

    Sangha, V., Khunte, A., Holste, G., Mortazavi, B. J., Wang, Z., Oikonomou, E. K., and Khera, R. Biometric contrastive learning for data-efficient deep learning from electrocardiographic images. Journal of the American Medical Informatics Association, 31 0 (4): 0 855--865, 2024

    2024

  28. [36]

    C., Noseworthy, P

    Siontis, K. C., Noseworthy, P. A., Attia, Z. I., and Friedman, P. A. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nature Reviews Cardiology, 18 0 (7): 0 465--478, 2021

    2021

  29. [37]

    T., Hong, D., Kwon, J.-m., and Jo, Y.-Y

    Song, J., Jang, J.-H., Lee, B. T., Hong, D., Kwon, J.-m., and Jo, Y.-Y. Foundation models for ecg: Leveraging hybrid self-supervised learning for advanced cardiac diagnostics. arXiv preprint arXiv:2407.07110, 2024

    2024 arXiv

  30. [38]

    Foundation model of ecg diagnosis: Diagnostics and explanations of any form and rhythm on ecg

    Tian, Y., Li, Z., Jin, Y., Wang, M., Wei, X., Zhao, L., Liu, Y., Liu, J., and Liu, C. Foundation model of ecg diagnosis: Diagnostics and explanations of any form and rhythm on ecg. Cell Reports Medicine, 5 0 (12), 2024

    2024

  31. [39]

    I., Samek, W., and Schaeffter, T

    Wagner, P., Strodthoff, N., Bousseljot, R.-D., Kreiseler, D., Lunze, F. I., Samek, W., and Schaeffter, T. Ptb-xl, a large publicly available electrocardiography dataset. Scientific data, 7 0 (1): 0 1--15, 2020

    2020

  32. [40]

    Adversarial spatiotemporal contrastive learning for electrocardiogram signals

    Wang, N., Feng, P., Ge, Z., Zhou, Y., Zhou, B., and Wang, Z. Adversarial spatiotemporal contrastive learning for electrocardiogram signals. IEEE Transactions on Neural Networks and Learning Systems, 35 0 (10): 0 13845--13859, 2024. doi:10.1109/TNNLS.2023.3272153

    2024

  33. [41]

    Williams, R. J. and Zipser, D. A learning algorithm for continually running fully recurrent neural networks. Neural computation, 1 0 (2): 0 270--280, 1989

    1989

  34. [42]

    X., and Lin, D

    Wu, Z., Xiong, Y., Yu, S. X., and Lin, D. Unsupervised feature learning via non-parametric instance discrimination. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 3733--3742, 2018

    2018

  35. [43]

    Fusing transformer model with temporal features for ecg heartbeat classification

    Yan, G., Liang, S., Zhang, Y., and Liu, F. Fusing transformer model with temporal features for ecg heartbeat classification. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp.\ 898--905. IEEE, 2019

    2019

  36. [44]

    Ecg-sl: Electrocardiogram (ecg) segment learning, a deep learning method for ecg signal

    Yu, H., Yang, H., and Sano, A. Ecg-sl: Electrocardiogram (ecg) segment learning, a deep learning method for ecg signal. arXiv preprint arXiv:2310.00818, 2023

    2023 arXiv

  37. [45]

    Ecg semantic integrator (esi): A foundation ecg model pretrained with llm-enhanced cardiological text

    Yu, H., Guo, P., and Sano, A. Ecg semantic integrator (esi): A foundation ecg model pretrained with llm-enhanced cardiological text. arXiv preprint arXiv:2405.19366, 2024

    2024 arXiv

  38. [46]

    Coca: Contrastive captioners are image-text foundation models

    Yu, J., Wang, Z., Vasudevan, V., Yeung, L., Seyedhosseini, M., and Wu, Y. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022

    2022 arXiv

  39. [47]

    Ts2vec: Towards universal representation of time series

    Yue, Z., Wang, Y., Duan, J., Yang, T., Huang, C., Tong, Y., and Xu, B. Ts2vec: Towards universal representation of time series. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 8980--8987, 2022

    2022

  40. [48]

    Barlow twins: Self-supervised learning via redundancy reduction

    Zbontar, J., Jing, L., Misra, I., LeCun, Y., and Deny, S. Barlow twins: Self-supervised learning via redundancy reduction. In International conference on machine learning, pp.\ 12310--12320. PMLR, 2021

    2021

  41. [49]

    Self-supervised time series representation learning via cross reconstruction transformer

    Zhang, W., Yang, L., Geng, S., and Hong, S. Self-supervised time series representation learning via cross reconstruction transformer. IEEE Transactions on Neural Networks and Learning Systems, 2023

    2023

  42. [50]

    Ecg-chat: A large ecg-language model for cardiac disease diagnosis

    Zhao, Y., Zhang, T., Wang, X., Han, P., Chen, T., Huang, L., Jin, Y., and Kang, J. Ecg-chat: A large ecg-language model for cardiac disease diagnosis. arXiv preprint arXiv:2408.08849, 2024

    2024 arXiv

  43. [51]

    A large scale 12-lead electrocardiogram database for arrhythmia study (version 1.0

    Zheng, J., Guo, H., and Chu, H. A large scale 12-lead electrocardiogram database for arrhythmia study (version 1.0. 0). PhysioNet 2022Available online httpphysionet orgcontentecg arrhythmia10 0accessed on, 23, 2022

    2022

  44. [52]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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