Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Signal, Image, or Symbolic: Exploring the Best Input Representation for Electrocardiogram-Language Models Through a Unified Framework

T0 review · 4 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 Symbolic, tokenized ECG representations are the most effective input modality for ECG-language models, beating raw signals and images in generative benchmarks.

desk verdict Broad, useful benchmark of complete pipelines, but the headline claim that symbolic tokens are the best input representation is confounded by a one-to-one mapping between representation and training paradigm. read the letter →

arxiv 2505.18847 v1 pith:XGLKOTVM submitted 2025-05-24 cs.AI cs.CL

William Han , Chaojing Duan , Zhepeng Cen , Yihang Yao , Xiaoyu Song , Atharva Mhaskar , Dylan Leong , Michael A. Rosenberg
show 2 more authors
Emerson Liu Ding Zhao
This is my paper · ORCID
classification cs.AIcs.CL
keywords electrocardiogram-languagemodelsECGtokenizationbyte-pairencodinginputrepresentationbenchmarkautoregressivegenerationmultimodallargelanguagesymbolictimeseriesrobustness
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 asks which way of feeding an electrocardiogram (ECG) into a generative language model works best: raw time-series values, a rendered image, or a discretized symbolic token sequence. It builds a unified training and evaluation framework covering four input variants, six public datasets, and five text-generation metrics, and it reports that the symbolic representation, produced by ECG-Byte's quantization plus byte-pair encoding, wins the most statistically significant comparisons and also generalizes best in zero-shot evaluation. The practical conclusion, if the finding holds, is that researchers building ECG-language models should tokenize ECGs into discrete symbols rather than invest in ECG-specific encoders or vision adapters for generative tasks. The paper also cautions that even the best representation remains far below expert cardiologist performance.

What carries the argument

The load-bearing mechanism is ECG-Byte, a tokenizer that maps each normalized, amplitude-quantized ECG trace into a sequence over a 26-symbol alphabet and then applies byte-pair encoding, iteratively merging frequent symbol pairs into an extended token vocabulary. This converts the ECG into the same discrete token stream as the text, so the whole model is one autoregressive language model trained end-to-end on a masked next-token loss without a separate ECG encoder or learned projection layer. The evaluation machinery is a paired t-test across five random seeds that converts per-metric differences into counts of statistically significant wins, with robustness checks under random perturbation.

What would settle it

Train raw-signal, image, and symbolic inputs under a single identical training procedure (same LLM, same optimizer, same sequence budget, and an end-to-end objective for all three) and rerun the paired comparisons. If the symbolic advantage shrinks to statistical noise or reverses, the central claim is an artifact of the per-representation training pipelines; if it survives, the conclusion is materially stronger.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the tokenized symbolic input $X_{ID}$ is the most effective modality for autoregressive ECG-language models. In the head-to-head benchmark, $X_{ID}$ trained end-to-end accumulates the most statistically significant wins (9) compared with the best raw-signal encoder MERL (4) and the best image encoders (1 each), and it attains the highest mean scores across all five metrics in the zero-shot PULSE ECG-Bench evaluation. The authors further find that this advantage widens when longer ECG segments are used ($L=2500$) and that the model remains stable under added Gaussian noise and baseline wander, supporting their conclusion that symbolic tokenization is the preferred input representation for generative ELMs.

Load-bearing premise

Each representation is tested only with its own training pipeline—raw signals through a two-stage encoder, images through a frozen pretrained vision encoder with a learned projector, and symbols through direct end-to-end token training—so the paper must assume that the performance differences come from the representation itself rather than from the pipeline.

Editorial extensions

If this is right

  • Tokenized symbolic ECG input should become the default representation for generative ECG-language modeling, since it wins on more dataset-metric pairs than any raw-signal or image-based configuration.
  • End-to-end symbolic training eliminates the need for a dedicated ECG encoder and its pretraining stage, so scaling an ELM reduces to scaling an ordinary language model on token streams.
  • The symbolic advantage grows with longer ECG segments, so longer context windows favor tokenized inputs over single-token compressed representations.
  • At short sequence budgets ($T=512$) symbolic models degrade because signal tokens must be truncated, whereas image and raw-signal pipelines that compress the ECG into one token stay stable; practitioners with tight memory may prefer those pipelines.
  • Even the best representation is not clinically expert-level, so the benchmark should be read as a guide for representation choice, not as a claim of ready-for-bedside accuracy.

Reading between the lines

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

  • Beyond the paper: the decisive test that the benchmark does not run is an equal-pipeline comparison—training raw-signal and image inputs with the exact same end-to-end objective used for symbols—so the representation-level conclusion still depends on the training-paradigm confound.
  • Beyond the paper: if symbolic tokenization transfers, the same quantization-plus-BPE recipe could be applied to other physiological traces, such as intracardiac electrograms or photoplethysmography, wherever free-form generative reports are wanted.
  • Beyond the paper: because perturbation robustness differences were small across all representations, deployment choices may ultimately hinge on data availability, hardware budgets, and interoperability with text-only toolchains rather than on generative quality alone.
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 / 5 minor

Summary. The paper presents a benchmark of four ECG input representations (raw signal Xsig, stacked signal X*sig, rendered image Ximg, and tokenized symbolic sequence XID) for autoregressive ECG-language models (ELMs). Using three training paradigms (2-Stage Scratch, End-to-End LLaVA, and End-to-End), the authors evaluate across six public datasets and five text-generation metrics, plus robustness and ablations of LLM backbone, sequence length, and ECG length. The headline claim is that symbolic tokenized representations (ECG-Byte) constitute the most effective input modality, with the greatest number of statistically significant wins, strong robustness, and efficient scaling.

Significance. If the central claim were supported, the paper would offer clear guidance for ELM architecture choices and would be a useful contribution to a rapidly growing area. The open-sourced unified framework, the breadth of datasets and encoders, and the systematic ablations are genuine strengths. However, the study's design does not isolate input representation from the training pipeline: every representation is evaluated with exactly one associated training paradigm, and the paper's own ablations (Table 5) show that token budget, not symbolic content per se, likely drives much of the observed advantage. The significance of the empirical result is therefore currently limited to comparing complete pipelines, not input representations.

major comments (4)
  1. [§5.1, Table 1] The central comparison confounds input representation with training paradigm. Xsig always uses 2-Stage Scratch with a from-scratch ECG encoder, Ximg and X*sig always use End-to-End LLaVA with frozen pretrained encoders, and XID always uses End-to-End training with ECG-Byte tokens. The paper itself notes this in the Figure 2 caption and in §5.1, but the conclusion 'symbolic, tokenized representations of ECGs constitute the most effective input modality' (§6) is a claim about the representation, not about a pipeline. At least four factors vary simultaneously: representation, encoder initialization, trainable parameters, and the number of signal tokens seen by the LLM (one token vs. a long sequence). The strongest counter-explanation is that ECG-Byte wins because End-to-End training lets the LLM see thousands of ECG tokens and optimize directly on the generation loss, whereas the other pipelines compress the 12-lead signal into a single token before the LLM sees it. The T=512 ablation (Table 5) supports this: when the ECG-Byte token budget is truncated, its performance collapses, while the single-token pipelines remain stable. To support the representation-level claim, the authors need either a design that holds the training paradigm fixed (e.g., comparing Xsig and XID both in an End-to-End setup, or Ximg and XID both in the same paradigm), or the conclusion must be reformulated as a claim about the End-to-End ECG-Byte pipeline.
  2. [§6 vs. §5.1] The significance analysis does not support the statement that symbolic representations achieve 'statistically significant wins over both signal and image inputs.' For each (dataset, metric) pair, only the top-2 models are compared with a paired t-test over five seeds. This means ECG-Byte is never directly compared against most signal/image variants. In fact, Table 8 shows numeric losses for ECG-Byte on some metrics (e.g., ECG-QA MIMIC-IV-ECG BLEU-4: MERL 19.77 and CLIP-Ximg 19.92 vs. ECG-Byte 16.47; ECG-QA PTB-XL METEOR: MTAE 27.09 vs. ECG-Byte 28.95? Actually ECG-Byte wins there; but the point stands that not all pairwise comparisons are made). The aggregated 'win count' also omits losses and effect sizes. Reporting only wins over the runner-up, without correction for multiple testing and without pairwise comparisons to all baselines, overstates the evidence for the headline claim.
  3. [§5.3, Table 3] The Discussion (Section 6) states the results are 'significant wins in two-sample z-tests,' while Section 5.1 describes paired t-tests across five seeds. This is an internal inconsistency. Moreover, a paired t-test with n=5 has very low power and is sensitive to the assumption of normally distributed differences; the paper does not report effect sizes or confidence intervals. Given that the 'significant wins' table is the backbone of the conclusion, the statistical methodology needs to be made consistent and more robust, and the number of comparisons should be acknowledged.
  4. The robustness experiment is interpreted as evidence that symbolic models 'effectively leverage compressed symbolic information even when signal fidelity degrades' (Section 6). However, the perturbation deltas are tiny and often positive even for non-symbolic baselines (e.g., SigLIP X*sig accuracy +0.13, CLIP X*sig accuracy +0.11), and the paper itself notes in Section 5.3 that the language model may 'derive limited utility from the raw signal.' This pattern is more consistent with the model mostly ignoring the ECG input than with robust representation use. A negative control that ablates or randomizes the ECG input (e.g., replacing it with noise or a constant signal) would be needed to determine whether any representation is actually being used. As written, the robustness section does not provide discriminating evidence for the symbolic representation.
minor comments (5)
  1. [§3.3] The ECG-Byte tokenizer is trained with 5000 BPE merges and an instance-normalization scheme that deviates from the original ECG-Byte methodology (global percentiles). These are free parameters that could affect the comparison; a brief justification or ablation of the normalization choice would strengthen the paper.
  2. [§5.1, Figure 2] The spider charts are difficult to read because many overlapping series are plotted on the same axes; the reader must rely on Table 8 for exact values. Consider showing per-dataset bar charts with error bars or a faceted layout.
  3. [§2.3] The sentence 'some studies speculate that design choices in signal plotting can significantly affect a model's ability to interpret the visualization' lacks a citation; please add a reference or remove the attribution.
  4. [§6] The phrase 'efficient scaling' in the conclusion is not directly demonstrated: Table 6 shows performance increasing with ECG length for ECG-Byte, but no scaling-law analysis or compute comparison is provided. Please either support this term or replace it with a more specific statement.
  5. [§4.1] The data split mention of '70:10' (rather than 70/10/20 or 70/20/10) is ambiguous; please clarify the actual train/test split proportions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the representation-versus-pipeline confound is a validity threat, not a circularity.

full rationale

This paper is an empirical benchmark, not a derivation chain, and no prediction reduces to a fitted input by construction. The central comparison does vary representation and training pipeline together (Section 5.1: 'We denote input modality variants as Xsig, X*sig, Ximg, and Xid, corresponding to 2-Stage Scratch, End-to-End LLaVA, End-to-End LLaVA, and End-to-End training, respectively'), and the paper's own Figure 2 caption states that each representation 'utilize[s]' a specific paradigm. That is a genuine experimental confound that weakens the causal claim about representation, but it is not a circular reduction: the reported metrics are not defined in terms of the conclusion, and no parameter is fitted to the test labels and then relabeled as a prediction. The ECG-Byte tokenizer comes from the authors' prior work, but its hyperparameters (5000 merges, 300,000 sampled signals) were fixed before evaluation, and the paper even modifies the normalization scheme; this is reuse of a published tool, not a self-citation that smuggles in the target result. The benchmark is externally falsifiable through public datasets such as PULSE ECG-Bench. The confound should be fixed with a matched design, but under the circularity rubric it does not raise the score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

This is an empirical benchmark rather than a derivation, so the axiom ledger is limited. The main load-bearing choices are the pairing of each representation with a fixed training paradigm and the ECG-Byte tokenizer hyperparameters. No new theoretical entities are introduced.

free parameters (2)
  • ECG-Byte BPE merge count = 5000 merges
    Chosen when training the ECG-Byte tokenizer on 300,000 ECG signals. The symbolic representation, and therefore the central comparison, depends on this hyperparameter.
  • ECG-Byte normalization scheme = Instance min-max normalization to [0,1]
    The paper deviates from the original ECG-Byte global-percentile normalization by using per-record min-max scaling. This choice affects the tokenizer and is specific to the symbolic pipeline.
assumptions (4)
  • domain assumption Autoregressive next-token prediction with cross-entropy loss is a valid objective for ECG-language models.
    The entire benchmark is built on this training objective; other approaches such as classification or non-autoregressive generation are out of scope.
  • standard math Paired t-tests over 5 random seeds are a valid way to compare model performance.
    The test assumes independent and normally distributed per-seed differences. With n=5 this is fragile, and the paper aggregates many tests without multiple-comparison correction.
  • domain assumption Each representation is tested only with its canonical training paradigm, and observed differences are attributed to the representation.
    Xsig uses 2-Stage Scratch, Ximg and X*sig use End-to-End LLaVA, and XID uses End-to-End training. This confound is load-bearing for the central claim.
  • domain assumption The ECG preprocessing pipeline (filtering, resampling to 250 Hz, segmentation) preserves clinically relevant information equally across all input representations.
    If preprocessing disproportionately harms one representation, the comparison is biased. The paper applies the same preprocessing to all signals before transformation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Signal, Image, or Symbolic: Exploring the Best Input Representation for Electrocardiogram-Language Models Through a Unified Framework." pith.science (2026). https://pith.science/paper/XGLKOTVM

@misc{pith2026250518847,
  author       = {Pith},
  title        = {Pith review of: Signal, Image, or Symbolic: Exploring the Best Input Representation for Electrocardiogram-Language Models Through a Unified Framework},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XGLKOTVM}},
  note         = {Machine review of arXiv:2505.18847}
}
read the original abstract

Recent advances have increasingly applied large language models (LLMs) to electrocardiogram (ECG) interpretation, giving rise to Electrocardiogram-Language Models (ELMs). Conditioned on an ECG and a textual query, an ELM autoregressively generates a free-form textual response. Unlike traditional classification-based systems, ELMs emulate expert cardiac electrophysiologists by issuing diagnoses, analyzing waveform morphology, identifying contributing factors, and proposing patient-specific action plans. To realize this potential, researchers are curating instruction-tuning datasets that pair ECGs with textual dialogues and are training ELMs on these resources. Yet before scaling ELMs further, there is a fundamental question yet to be explored: What is the most effective ECG input representation? In recent works, three candidate representations have emerged-raw time-series signals, rendered images, and discretized symbolic sequences. We present the first comprehensive benchmark of these modalities across 6 public datasets and 5 evaluation metrics. We find symbolic representations achieve the greatest number of statistically significant wins over both signal and image inputs. We further ablate the LLM backbone, ECG duration, and token budget, and we evaluate robustness to signal perturbations. We hope that our findings offer clear guidance for selecting input representations when developing the next generation of ELMs.

Figures

Figures reproduced from arXiv: 2505.18847 by the authors.

Figure 1
Figure 1. A high-level overview of our training and evaluation pipeline. The input data is represented as X = {Xsig, X∗ sig, Ximg, XID} as seen in Step 1. Step 2 comprises two modes: Encoder training and LLM training. During Encoder training, an ECG-specific encoder is trained from scratch. In LLM training, any method used to compress the ECG into an appropriate representation for the ELM is denoted by F(∗) for simplicity; th… view at source ↗
Figure 2
Figure 2. Spider charts for the performance of each model and training paradigm. We want to note that Xsig, X∗ sig, Ximg, and XID utilize 2-Stage Scratch, End￾to-End LLaVA, End-to-End LLaVA, and End-to-End training paradigms respectively. We include a table representing the same results in [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

Discussion (0). Continue with ORCID 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. ELF: A Family of Encoder-Free ECG-Language Models

    cs.MM 2026-01 conditional novelty 6.0 of 10

    A single linear projection from raw ECG to LLM embeddings matches complex encoder-based ECG-language models, while perturbation tests show such models largely ignore the ECG signal.

Reference graph

Works this paper leans on

64 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    The complexities of physician supply and demand: Projections from 2021 to 2036, 03 2024

    AAMC. The complexities of physician supply and demand: Projections from 2021 to 2036, 03 2024. URL https://www.aamc.org/media/75236/download

    work page 2021

  2. [2]

    A brief review: history to understand fundamentals of electrocardiography

    Majd AlGhatrif and Joseph Lindsay. A brief review: history to understand fundamentals of electrocardiography. Journal of community hospital internal medicine perspectives, 2, 04 2012. doi:10.3402/jchimp.v2i1.14383

    work page doi:10.3402/jchimp.v2i1.14383 2012

  3. [3]

    Classification of 12-lead ecgs: The physionet/computing in cardiology challenge 2020, 07 2022

    Erick Andres, Annie Gu, Amit Shah, Chengyu Liu, Ashish Sharma, Salman Seyedi, Ali Bahrami Rad, Matthew Reyna, and Gari Clifford. Classification of 12-lead ecgs: The physionet/computing in cardiology challenge 2020, 07 2022. URL https://physionet.org/content/challenge-2020/1.0.2/

    work page 2020

  4. [4]

    Meteor: An automatic metric for mt evaluation with improved correlation with human judgments

    Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In IEEvaluation@ACL, 2005

    work page 2005

  5. [5]

    Llm-abba: Understand time series via symbolic approximation, 2024 a

    Erin Carson, Xinye Chen, and Cheng Kang. Llm-abba: Understand time series via symbolic approximation, 2024 a . URL https://arxiv.org/abs/2411.18506

    arXiv 2024

  6. [6]

    Quantized symbolic time series approximation

    Erin Carson, Xinye Chen, and Cheng Kang. Quantized symbolic time series approximation, 2024 b . URL https://arxiv.org/abs/2411.15209

    work page Pith review arXiv 2024

  7. [7]

    Yhdego, Fatemeh Khadem, and Fatemeh Afghah

    Seokmin Choi, Sajad Mousavi, Phillip Si, Haben G. Yhdego, Fatemeh Khadem, and Fatemeh Afghah. Ecgbert: Understanding hidden language of ecgs with self-supervised representation learning, 2023

    work page 2023

  8. [8]

    Flashattention-2: Faster attention with better parallelism and work partitioning, 2023

    Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning, 2023. URL https://arxiv.org/abs/2307.08691

    arXiv 2023

Show all 64 references
  1. [9]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  2. [10]

    Abba: Adaptive brownian bridge-based symbolic aggregation of time series, 2020

    Steven Elsworth and Stefan Güttel. Abba: Adaptive brownian bridge-based symbolic aggregation of time series, 2020. URL https://arxiv.org/abs/2003.12469

    2020 arXiv

  3. [11]

    A new algorithm for data compression

    Philip Gage. A new algorithm for data compression. The C Users Journal archive, 12: 0 23--38, 1994. URL https://api.semanticscholar.org/CorpusID:59804030

    1994

  4. [12]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  5. [13]

    Rosenberg, Emerson Liu, and Ding Zhao

    William Han, Chaojing Duan, Michael A. Rosenberg, Emerson Liu, and Ding Zhao. Ecg-byte: A tokenizer for end-to-end generative electrocardiogram language modeling, 2024 a . URL https://arxiv.org/abs/2412.14373

    2024 arXiv

  6. [14]

    Rosenberg, Douglas Weber, Emerson Liu, and Ding Zhao

    William Jongwon Han, Diana Gomez, Avi Alok, Chaojing Duan, Michael A. Rosenberg, Douglas Weber, Emerson Liu, and Ding Zhao. Interpretation of intracardiac electrograms through textual representations, 2024 b . URL https://arxiv.org/abs/2402.01115

    2024 arXiv

  7. [15]

    Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H

    Awni Y. Hannun, Pranav Rajpurkar, Masoumeh Haghpanahi, Geoffrey H. Tison, Codie Bourn, Mintu P. Turakhia, and Andrew Y. Ng. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature Medicine, 25: 0 65–69, 01...

    2019 doi

  8. [16]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. URL https://arxiv.org/abs/2106.09685

    2021 arXiv

  9. [17]

    Reading your heart: Learning ecg words and sentences via pre-training ecg language model, 2025

    Jiarui Jin, Haoyu Wang, Hongyan Li, Jun Li, Jiahui Pan, and Shenda Hong. Reading your heart: Learning ecg words and sentences via pre-training ecg language model, 2025. URL https://arxiv.org/abs/2502.10707

    2025 arXiv

  10. [18]

    Alistair E. W. Johnson, Lucas Bulgarelli, Lu Shen, Alvin Gayles, Ayad Shammout, Steven Horng, Tom J. Pollard, Benjamin Moody, Brian Gow, Li-wei H. Lehman, Leo A. Celi, and Roger G. Mark. Mimic-iv, a freely accessible electronic health record dataset. Scientific Data, 10, 01 20...

    2023 doi

  11. [19]

    Counties most in need of cardiologists are the most likely to have none, 07 2024

    Mark Johnson. Counties most in need of cardiologists are the most likely to have none, 07 2024. URL https://www.washingtonpost.com/science/2024/07/29/cardiologists-rural-counties-shortage/

    2024

  12. [20]

    Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. URL https://arxiv.org/abs/2001.08361

    2020 arXiv

  13. [21]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017

    2017

  14. [22]

    Dani Kiyasseh, Tingting Zhu, and David A. Clifton. Clocs: Contrastive learning of cardiac signals across space, time, and patients, 2021. URL https://arxiv.org/abs/2005.13249

    2021 arXiv

  15. [23]

    A simple weight decay can improve generalization

    Anders Krogh and John Hertz. A simple weight decay can improve generalization. In J. Moody, S. Hanson, and R.P. Lippmann, editors, Advances in Neural Information Processing Systems, volume 4. Morgan-Kaufmann, 1991. URL https://proceedings.neurips.cc/paper_files/paper/1991/file...

    1991

  16. [24]

    Gem: Empowering mllm for grounded ecg understanding with time series and images, 2025

    Xiang Lan, Feng Wu, Kai He, Qinghao Zhao, Shenda Hong, and Mengling Feng. Gem: Empowering mllm for grounded ecg understanding with time series and images, 2025. URL https://arxiv.org/abs/2503.06073

    2025

  17. [25]

    Time series as images: Vision transformer for irregularly sampled time series, 2023

    Zekun Li, Shiyang Li, and Xifeng Yan. Time series as images: Vision transformer for irregularly sampled time series, 2023. URL https://arxiv.org/abs/2303.12799

    2023 arXiv

  18. [26]

    Rouge: A package for automatic evaluation of summaries

    Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In ACL 2004, 2004

    2004

  19. [27]

    Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement, 2024 a

    Che Liu, Zhongwei Wan, Cheng Ouyang, Anand Shah, Wenjia Bai, and Rossella Arcucci. Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement, 2024 a . URL https://arxiv.org/abs/2403.06659

    2024 arXiv

  20. [28]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. URL https://arxiv.org/abs/2304.08485

    2023 arXiv

  21. [29]

    Teach multimodal llms to comprehend electrocardiographic images, 2024 b

    Ruoqi Liu, Yuelin Bai, Xiang Yue, and Ping Zhang. Teach multimodal llms to comprehend electrocardiographic images, 2024 b . URL https://arxiv.org/abs/2410.19008

    2024 arXiv

  22. [30]

    Real-time frequency-independent single-lead and single-beat myocardial infarction detection

    Harold Martin, Ulyana Morar, Walter Izquierdo, Mercedes Cabrerizo, Anastasio Cabrera, and Malek Adjouadi. Real-time frequency-independent single-lead and single-beat myocardial infarction detection. Artificial intelligence in medicine, 121: 0 102179, 2021

    2021

  23. [31]

    Guiding masked representation learning to capture spatio-temporal relationship of electrocardiogram, 2024

    Yeongyeon Na, Minje Park, Yunwon Tae, and Sunghoon Joo. Guiding masked representation learning to capture spatio-temporal relationship of electrocardiogram, 2024. URL https://arxiv.org/abs/2402.09450

    2024 arXiv

  24. [32]

    Harnessing vision models for time series analysis: A survey, 2025

    Jingchao Ni, Ziming Zhao, ChengAo Shen, Hanghang Tong, Dongjin Song, Wei Cheng, Dongsheng Luo, and Haifeng Chen. Harnessing vision models for time series analysis: A survey, 2025. URL https://arxiv.org/abs/2502.08869

    2025 arXiv

  25. [33]

    Electrocardiogram classification by modified efficientnet with data augmentation

    Naoki Nonaka and Jun Seita. Electrocardiogram classification by modified efficientnet with data augmentation. In 2020 Computing in Cardiology, pages 1--4. IEEE, 2020

    2020

  26. [34]

    Lead-agnostic self-supervised learning for local and global representations of electrocardiogram, 2022

    Jungwoo Oh, Hyunseung Chung, Joon myoung Kwon, Dong gyun Hong, and Edward Choi. Lead-agnostic self-supervised learning for local and global representations of electrocardiogram, 2022. URL https://arxiv.org/abs/2203.06889

    2022 arXiv

  27. [35]

    Ecg-qa: A comprehensive question answering dataset combined with electrocardiogram, 2023

    Jungwoo Oh, Gyubok Lee, Seongsu Bae, Joon myoung Kwon, and Edward Choi. Ecg-qa: A comprehensive question answering dataset combined with electrocardiogram, 2023. URL https://arxiv.org/abs/2306.15681

    2023 arXiv

  28. [36]

    Cardiovascular diseases, 2024

    World Health Organization. Cardiovascular diseases, 2024. URL https://www.who.int/health-topics/cardiovascular-diseases#tab=tab_1

    2024

  29. [37]

    Bleu: a method for automatic evaluation of machine translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002

    2002

  30. [38]

    C-melt: Contrastive enhanced masked auto-encoders for ecg-language pre-training, 2024

    Manh Pham, Aaqib Saeed, and Dong Ma. C-melt: Contrastive enhanced masked auto-encoders for ecg-language pre-training, 2024. URL https://arxiv.org/abs/2410.02131

    2024 arXiv

  31. [39]

    Jielin Qiu, William Han, Jiacheng Zhu, Mengdi Xu, Michael Rosenberg, Emerson Liu, Douglas Weber, and Ding Zhao. Transfer knowledge from natural language to electrocardiography: Can we detect cardiovascular disease through language models? In Andreas Vlachos and Isabelle Augens...

    2023 doi

  32. [40]

    Rosenberg, Emerson Liu, Douglas Weber, and Ding Zhao

    Jielin Qiu, Jiacheng Zhu, Shiqi Liu, William Han, Jingqi Zhang, Chaojing Duan, Michael A. Rosenberg, Emerson Liu, Douglas Weber, and Ding Zhao. Automated cardiovascular record retrieval by multimodal learning between electrocardiogram and clinical report. In Stefan Hegselmann,...

    2023

  33. [41]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

    2025 arXiv

  34. [42]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021

    2021

  35. [43]

    Hannun, Masoumeh Haghpanahi, Codie Bourn, and Andrew Y

    Pranav Rajpurkar, Awni Y. Hannun, Masoumeh Haghpanahi, Codie Bourn, and Andrew Y. Ng. Cardiologist-level arrhythmia detection with convolutional neural networks, 2017. URL https://arxiv.org/abs/1707.01836

    2017 arXiv

  36. [44]

    Ribeiro, Gabriela M.M

    Antônio H. Ribeiro, Gabriela M.M. Paixao, Emilly M. Lima, Manoel Horta Ribeiro, Marcelo M. Pinto Filho, Paulo R. Gomes, Derick M. Oliveira, Wagner Meira Jr, Thömas B Schon, and Antonio Luiz P. Ribeiro. Code-15 URL https://zenodo.org/records/4916206

  37. [45]

    Mortazavi, Adrian D

    Veer Sangha, Bobak J. Mortazavi, Adrian D. Haimovich, Ant \^o nio H. Ribeiro, Cynthia A. Brandt, Daniel L. Jacoby, Wade L. Schulz, Harlan M. Krumholz, Antonio Luiz P. Ribeiro, and Rohan Khera. Automated multilabel diagnosis on electrocardiographic images and signals. medRxiv, ...

    2021 doi

  38. [46]

    Image-based time series forecasting: A deep convolutional neural network approach

    Artemios-Anargyros Semenoglou, Evangelos Spiliotis, and Vassilios Assimakopoulos. Image-based time series forecasting: A deep convolutional neural network approach. Neural Networks, 157: 0 39--53, 2023. ISSN 0893-6080. doi:https://doi.org/10.1016/j.neunet.2022.10.006. URL http...

    2023 doi

  39. [47]

    Clifford, Matthew A

    Kshama Kodthalu Shivashankara, Deepanshi, Afagh Mehri Shervedani, Gari D. Clifford, Matthew A. Reyna, and Reza Sameni. Ecg-image-kit: A synthetic image generation toolbox to facilitate deep learning-based electrocardiogram digitization, 2024. URL https://arxiv.org/abs/2307.01946

    2024 arXiv

  40. [48]

    Deep learning for ecg analysis: Benchmarks and insights from ptb-xl

    Nils Strodthoff, Patrick Wagner, Tobias Schaeffter, and Wojciech Samek. Deep learning for ecg analysis: Benchmarks and insights from ptb-xl. IEEE Journal of Biomedical and Health Informatics, 25: 0 1519--1528, 2021

    2021

  41. [49]

    Heartbert: A self-supervised ecg embedding model for efficient and effective medical signal analysis, 2024

    Saedeh Tahery, Fatemeh Hamid Akhlaghi, Termeh Amirsoleimani, and Saeed Farzi. Heartbert: A self-supervised ecg embedding model for efficient and effective medical signal analysis, 2024. URL https://arxiv.org/abs/2411.11896

    2024 arXiv

  42. [50]

    Electrocardiogram-language model for few-shot question answering with meta learning, 2024 a

    Jialu Tang, Tong Xia, Yuan Lu, Cecilia Mascolo, and Aaqib Saeed. Electrocardiogram-language model for few-shot question answering with meta learning, 2024 a . URL https://arxiv.org/abs/2410.14464

    2024 arXiv

  43. [51]

    Electrocardiogram report generation and question answering via retrieval-augmented self-supervised modeling, 2024 b

    Jialu Tang, Tong Xia, Yuan Lu, Cecilia Mascolo, and Aaqib Saeed. Electrocardiogram report generation and question answering via retrieval-augmented self-supervised modeling, 2024 b . URL https://arxiv.org/abs/2409.08788

    2024 arXiv

  44. [52]

    Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le Lan...

    2024 arXiv

  45. [53]

    Heartbeit: Vision transformer for electrocardiogram data improves diagnostic performance at low sample sizes, 2022

    Akhil Vaid, Joy Jiang, Ashwin Sawant, Stamatios Lerakis, Edgar Argulian, Yuri Ahuja, Joshua Lampert, Alexander Charney, Hayit Greenspan, Benjamin Glicksberg, Jagat Narula, and Girish Nadkarni. Heartbeit: Vision transformer for electrocardiogram data improves diagnostic perform...

    2022

  46. [54]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023

    2023

  47. [55]

    Lunze, Wojciech Samek, and Tobias Schaeffter

    Patrick Wagner, Nils Strodthoff, Ralf-Dieter Bousseljot, Dieter Kreiseler, Fatima I. Lunze, Wojciech Samek, and Tobias Schaeffter. PTB - XL , a large publicly available electrocardiography dataset. Scientific Data, 7 0 (1): 0 154, May 2020. ISSN 2052-4463. doi:10.1038/s41597-0...

    2020 doi

  48. [56]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  49. [57]

    Routine electrocardiogram screening and cardiovascular disease events in adults

    Ryuichiro Yagi, Yuichiro Mori, Shinichi Goto, Taku Iwami, and Kosuke Inoue. Routine electrocardiogram screening and cardiovascular disease events in adults. JAMA Internal Medicine, 184 0 (9): 0 1035--1044, 09 2024. ISSN 2168-6106. doi:10.1001/jamainternmed.2024.2270. URL https...

    2024

  50. [58]

    Sigmoid loss for language image pre-training, 2023

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training, 2023. URL https://arxiv.org/abs/2303.15343

    2023 arXiv

  51. [59]

    Maefe: Masked autoencoders family of electrocardiogram for self-supervised pretraining and transfer learning

    Huaicheng Zhang, Wenhan Liu, Jiguang Shi, Sheng Chang, Hao Wang, Jin He, and Qijun Huang. Maefe: Masked autoencoders family of electrocardiogram for self-supervised pretraining and transfer learning. IEEE Transactions on Instrumentation and Measurement, 72: 0 1--15, 2023. doi:...

    2023

  52. [60]

    Weinberger, and Yoav Artzi

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. ArXiv, abs/1904.09675, 2020

    1904 arXiv

  53. [61]

    Ecg-chat: A large ecg-language model for cardiac disease diagnosis, 2024

    Yubao Zhao, Tian Zhang, Xu Wang, Puyu Han, Tong Chen, Linlin Huang, Youzhu Jin, and Jiaju Kang. Ecg-chat: A large ecg-language model for cardiac disease diagnosis, 2024. URL https://arxiv.org/abs/2408.08849

    2024 arXiv

  54. [62]

    A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients

    Jianwei Zheng, Jianming Zhang, Sidy Danioko, Hai Yao, Hangyuan Guo, and Cyril Rakovski. A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients. Scientific Data, 7: 0 48, 02 2020. doi:10.1038/s41597-020-0386-x. URL https://www.nature.com...

    2020 doi

  55. [63]

    P Xing, Hao Zhang, Joseph E

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric. P Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging llm-as-a-judge with mt-bench and chatbot arena, 2023

    2023

  56. [64]

    See it, think it, sorted: Large multimodal models are few-shot time series anomaly analyzers, 2024

    Jiaxin Zhuang, Leon Yan, Zhenwei Zhang, Ruiqi Wang, Jiawei Zhang, and Yuantao Gu. See it, think it, sorted: Large multimodal models are few-shot time series anomaly analyzers, 2024. URL https://arxiv.org/abs/2411.02465

    2024 arXiv

Pith tools

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