Pith. sign in

REVIEW 3 major objections 6 minor 47 references

ECG foundation models do not transfer clinically meaningful knowledge for Brugada syndrome; pre-training mainly stabilizes large models that fail from scratch.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · grok-4.5

2026-07-12 05:28 UTC pith:6JKIPP4X

load-bearing objection Clean negative result: for Brugada, public ECG FMs mostly buy optimization stability, not transferable clinical knowledge—architecture and site alignment dominate. the 3 major comments →

arxiv 2607.03009 v1 pith:6JKIPP4X submitted 2026-07-03 cs.LG

Do ECG Foundation Models Transfer to Rare Cardiac Diseases? Evidence from Brugada Syndrome Detection

classification cs.LG
keywords ElectrocardiogramFoundation ModelsSelf-Supervised LearningRare Cardiac DiseasesBrugada SyndromeTransfer LearningCross-site Generalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tests whether large ECG foundation models, pre-trained on vast unlabeled heart recordings, actually learn transferable clinical knowledge for rare diseases—or whether their benefits are mostly easier training. Using Brugada syndrome as a hard rare-disease case, the authors evaluate nine public ECG foundation models against identical architectures trained from scratch on two independent clinical cohorts, under linear probing, full fine-tuning, severe data reduction, and zero-shot hospital-to-hospital transfer. Pre-training rescues high-capacity models that otherwise collapse, but compact models that already work from scratch gain little; the best fine-tuned foundation model only marginally exceeds the strongest supervised baseline. Data-efficiency wins on one reduced cohort do not hold on the second site, and every strategy collapses toward chance under cross-site transfer. The finding challenges the assumption that scale alone encodes rare-phenotype knowledge and elevates architecture and domain alignment over pre-training status.

Core claim

For Brugada syndrome detection, foundation-model pre-training is mechanical rather than semantic: it stabilizes optimization for high-capacity architectures that cannot converge from scratch, but does not deliver statistically reliable gains over the best architecture-matched supervised baselines, does not produce a data-efficiency advantage that replicates across independent cohorts, and does not improve zero-shot cross-site robustness. Architectural suitability, not pre-training, is the most consistent predictor of performance.

What carries the argument

Architecture-matched from-scratch baselines for each of nine public ECG foundation models, compared under linear probing and full fine-tuning, plus a 3% window-per-patient data ablation and zero-shot cross-site transfer between the BrSwiss and HUCA cohorts, which isolates the contribution of pre-trained weights from architecture, data volume, and site shift.

Load-bearing premise

The claim depends on Brugada syndrome in these two adult case-control cohorts being a fair rare-disease stress test, and on those cases not having been present in any of the models’ pre-training data.

What would settle it

On a third independent Brugada (or comparable rare ECG) cohort of similar size, show that the best fine-tuned foundation model significantly and reproducibly beats the best architecture-matched from-scratch baseline both under matched training-set size and under zero-shot cross-site transfer.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Practitioners cannot assume off-the-shelf ECG foundation models will help rare cardiac phenotypes; architecture choice may matter more than pre-training.
  • Label-efficiency claims for ECG foundation models need multi-site replication before they are trusted in rare-disease settings.
  • Reliable cross-site rare-disease ECG performance will require explicit domain adaptation, not reliance on large-scale pre-training alone.
  • Compact models that train well from scratch can nearly match fine-tuned foundation models when limited labeled rare-disease data exist.
  • Pre-training strategies aimed at rare disease should deliberately cover clinical variability and multi-site domain shifts rather than scale alone.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar mechanical-versus-semantic patterns may appear for other low-prevalence ECG phenotypes absent from public pre-training corpora.
  • Acquisition-protocol and sampling-rate mismatch may erase any representation advantage faster than pre-training can create one.
  • Rare-disease ECG benchmarks may need mandatory architecture-matched from-scratch controls before foundation-model benefit is claimed.
  • If Brugada-like patterns were deliberately added to pre-training corpora, a reopened between-model gap would test whether semantic coverage was the missing ingredient.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript systematically benchmarks nine publicly available ECG foundation models for Brugada syndrome detection on an internal Swiss cohort (BrSwiss; 294 patients, 87 cases) and an independent external Spanish cohort (HUCA; 363 patients, 76 cases). Each model is evaluated under from-scratch training, linear probing, and full fine-tuning, with matched-architecture supervised baselines, a 3% training-window ablation, and zero-shot cross-site transfer. The central claim is that, for this rare phenotype, FM pre-training is largely mechanical (stabilizing optimization of high-capacity models that fail to converge from scratch) rather than semantic (encoding transferable clinical knowledge): the best fine-tuned FM only marginally exceeds the strongest from-scratch baseline on full BrSwiss (ECG-CPC 0.962 vs 0.932, p=0.091), the BrSwiss-3% data-efficiency gain does not replicate on HUCA, and all strategies collapse toward chance under zero-shot cross-site transfer. Architectural suitability, not pre-training per se, is identified as the most consistent performance driver.

Significance. If the result holds under the paper’s stated scope, it is a timely and useful reality check for the ECG foundation-model literature, which has largely evaluated common, well-labeled conditions. The design strengths are real and should be credited: one-to-one matched architecture from-scratch baselines, patient-level AUC with bootstrap CIs, DeLong tests with FDR control across large comparison blocks, early stopping keyed to validation AUC to equalize effective training exposure under heterogeneous native input windows, and an independent external cohort released after all pre-training cutoffs. These choices make the within-model versus between-model decomposition and the non-replication of data-efficiency gains more persuasive than typical single-model transfer reports. The work does not claim a universal negative about FMs; it scopes a concrete failure mode (rare phenotype outside pre-training coverage, site shift) that the field needs to confront when arguing for clinical deployment of large ECG FMs.

major comments (3)
  1. Title, abstract, and conclusion frame the result as evidence about transfer to “rare cardiac diseases,” but all empirical support is for a single adult case–control phenotype (Brugada) on two specialized cohorts. Methods 3.1 and the Limitations section already note this; the load-bearing issue is that the abstract’s closing sentence and the paper title still invite a broader reading than the data justify. Either narrow the title/abstract claims to Brugada (or “a rare ECG phenotype”) or add a short, explicit boundary statement that generalization to other rare cardiovascular conditions is untested and may differ when the phenotype is better represented in pre-training corpora.
  2. The cross-site claim (“pre-training does not confer cross-site robustness”) is central (Results §4.2 Cross-site generalization; Discussion theme 3; Figure 4; Tables 8–10). Limitations correctly notes the native-rate mismatch (BrSwiss 1000 Hz vs HUCA 100 Hz) and that up-sampling cannot restore content above the original Nyquist frequency. This is not a minor preprocessing detail: several FMs were pre-trained at 250–500 Hz, so HUCA inputs are spectrally truncated relative to their pre-training distribution. Without a controlled same-rate ablation (e.g., all pipelines forced to 100 Hz on both sites, or BrSwiss downsampled before transfer), it remains ambiguous how much of the near-chance collapse is population/protocol shift versus spectral domain shift. A same-rate control, or at minimum a quantitative sensitivity analysis, is needed before the “no robustness from pre-training” conclusion
  3. The interpretive slogan “mechanical rather than semantic” (Abstract Conclusion; Discussion first theme; Conclusions) is stronger than the direct evidence. Within-model FT–FS gains for large transformers that fail from scratch, non-significant between-model gaps on full BrSwiss and HUCA, non-replication of BrSwiss-3% efficiency on HUCA, and chance-level zero-shot transfer jointly support that pre-training does not deliver a practical, transferable clinical advantage over the best architecture trained from scratch. They do not, by themselves, prove that pre-trained representations contain no Brugada-relevant structure—only that any such structure is insufficient, non-transferable across sites, or dominated by optimization effects under the tested adaptation protocols. Soften or operationalize the language (e.g., “pre-training primarily stabilizes optimization and does not yield transferabl
minor comments (6)
  1. Secondary metrics (Appendix Tables 11–16) are reported at a fixed threshold of 0.5. The Clinical utility paragraph correctly notes that Brugada screening needs sensitivity-oriented operating points; consider adding one sensitivity-at-fixed-specificity (or Youden/validation-tuned) column for the top models so readers can see clinical operating characteristics without re-thresholding themselves.
  2. Patient-level aggregation uses mean probability pooling over segments (Appendix C). Brugada type-1 pattern can be intermittent (spontaneous vs drug/fever-induced). A brief sensitivity check with max pooling (or fraction of positive segments) would clarify whether relative FM vs FS rankings are robust to aggregation choice.
  3. Table 1 and Methods 3.3 assert that BrSwiss was never in any pre-training corpus and that HUCA’s public release postdates all pre-training. For completeness, a short note on whether any of the public pre-training sources (MIMIC-IV-ECG, CODE, PTB-XL, etc.) are known to contain labeled or unlabeled Brugada-pattern ECGs would strengthen the “outside pre-training coverage” premise used in the Discussion.
  4. Figure 2–4 error bars and asterisks are dense; ensure the caption explicitly states that asterisks are within-model (LP/FT vs same-architecture FS) and that between-model claims are only in the text. A small legend or footnote would reduce misreading.
  5. Saliency appendix (Figures 5–7) usefully cautions that IG magnitude scales with logit scale and that maps are per-panel normalized. Consider moving one sentence of that caveat into the main-text Results or Discussion so readers who skip the appendix do not over-interpret visual saturation differences between FS and FT.
  6. Minor typography: abstract and body occasionally drop spaces after periods or in compound terms (e.g., “Foundationmodels”, “rarediseases”); a pass for spacing and consistent hyphenation of “fine-tuning” / “from-scratch” would help.

Circularity Check

0 steps flagged

No circularity: pure empirical benchmarking of public ECG FMs against matched from-scratch baselines; conclusions are direct statistical comparisons of observed AUCs, not forced by definition or self-citation.

full rationale

The paper is a self-contained empirical evaluation. It takes nine publicly released ECG foundation models (with their native architectures, weights, and input specs), trains identically architected models from random initialization under the same protocol, and reports patient-level AUCs (with DeLong tests and FDR control) under FS/LP/FT, a 3% window ablation, and zero-shot cross-site transfer on two independent cohorts. The central claim that pre-training is “mechanical rather than semantic” is an interpretation of those measured differences (large gains only for high-capacity models that fail to converge from scratch; non-significant between-model advantage of the best FT over the best FS; non-replication of the BrSwiss-3% data-efficiency gap on HUCA; near-chance zero-shot transfer for all strategies). No parameter is fitted to a subset and then re-labeled a prediction; no uniqueness theorem or ansatz is imported via self-citation; no quantity is defined in terms of the result it is claimed to predict. The 3% ratio and LLRD schedule are free experimental design choices that do not force the ranking or the mechanical-vs-semantic conclusion. Minor citations to prior FM benchmarks (including the source of ECG-CPC) supply context and models but are not load-bearing for the statistical comparisons performed here. Hence the derivation chain contains no circular steps.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard supervised-learning assumptions plus a handful of design choices (3 % ablation, LLRD rates, native input windows) that are free parameters of the experimental protocol rather than fitted scientific constants. No new physical or biological entities are postulated.

free parameters (4)
  • BrSwiss-3% window retention ratio
    Chosen by hand to match HUCA training-set size and to stress data scarcity; the exact 3 % value is not derived from first principles.
  • Layer-wise learning-rate decay schedule (1e-3 / 1e-4 / 1e-5)
    Hand-chosen three-tier LLRD used for all fine-tuning runs; different schedules could alter the FT vs FS gap.
  • Early-stopping patience (15 epochs, delta 0.001) and max 200 epochs
    Fixed hyper-parameters that equalize training exposure across heterogeneous input lengths but remain free design choices.
  • Decision threshold 0.5 for secondary metrics
    Fixed operating point used for accuracy/sensitivity etc.; clinical utility would require prevalence-aware threshold tuning.
axioms (4)
  • domain assumption Brugada cases are absent from all nine public pre-training corpora
    Asserted in Methods 3.1 on the basis of release dates and public dataset lists; if any model saw Brugada examples the 'semantic transfer' interpretation would be compromised.
  • domain assumption Patient-level mean-probability pooling is a valid aggregation of segment predictions
    Standard in ECG literature; used throughout for all AUC calculations.
  • standard math DeLong test with Benjamini-Hochberg FDR control correctly ranks paired AUCs
    Standard non-parametric comparison of ROC curves applied to pooled out-of-fold predictions.
  • ad hoc to paper Native input specifications of each FM package (window length, sampling rate) should be respected rather than forced to a common format
    Explicit design choice that treats each FM as a complete deployable package; alternative would be to re-sample everything to a common rate.

pith-pipeline@v1.1.0-grok45 · 41059 in / 2657 out tokens · 23715 ms · 2026-07-12T05:28:27.699141+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Do ECG Foundation Models Transfer to Rare Cardiac Diseases? Evidence from Brugada Syndrome Detection." pith.science (2026). https://pith.science/paper/6JKIPP4X

@misc{pith2026260703009,
  author       = {Pith},
  title        = {Pith review of: Do ECG Foundation Models Transfer to Rare Cardiac Diseases? Evidence from Brugada Syndrome Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JKIPP4X}},
  note         = {Machine review of arXiv:2607.03009}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Background: Foundation models (FMs) trained on large-scale unlabeled physiological data have emerged as a promising paradigm for medical artificial intelligence. Their ability to capture clinically meaningful, transferable representations for rare diseases remains largely unproven. This study investigates whether FM pre-training provides genuine clinical generalization benefits beyond improved optimization for rare electrocardiographic (ECG) phenotypes. Methods: We systematically evaluated nine publicly available ECG FMs for Brugada syndrome detection on the BrSwiss cohort (294 patients, 87 cases) and the independent external HUCA cohort (363 patients, 76 cases), under three strategies (from-scratch training, linear probing, full fine-tuning) across several configurations, including a 3% data ablation and zero-shot cross-site transfers. Results: Pre-training was necessary for high-capacity architectures unable to converge from scratch (AUC gain up to 0.411, p < 0.05), but gave no significant gain for compact architectures already converged on labeled data alone. On full BrSwiss, the best fine-tuned FM (ECG-CPC, AUC = 0.962) only marginally exceeded the strongest supervised baseline (ECG-CPC from scratch, AUC = 0.932; p = 0.091). At matched training-set size, the data-efficiency advantage on BrSwiss-3% (AUC gain = 0.055, p < 0.01) did not replicate on HUCA. Under zero-shot cross-site transfer, FM-based pipelines did not generalize better than supervised baselines, all approaching chance-level performance. Conclusion: For Brugada syndrome detection, FM pre-training is mechanical rather than semantic, providing optimization stability rather than transferable clinical knowledge. These findings challenge the assumption that large-scale pre-training inherently encodes clinically meaningful representations, highlighting the central role of model architecture and data-domain alignment.

Figures

Figures reproduced from arXiv: 2607.03009 by Alvise Dei Rossi, Beatrice Zanchi, Francesca Dalia Faraci, Georgia Sarquella-Brugada, Giuliana Monachino, Giulio Conte, Luigi Fiorillo.

Figure 1
Figure 1. Figure 1: Schematic overview of the experimental settings in our study. In the Train from Scratch approach (a), the model is initialized with random weights. In the Linear Evaluation approach (b), the pre-trained encoder weights are kept frozen to act as a static feature extractor, while only the classification head is trained. The Full Fine-Tuning approach (c) involves updating the entire pre-trained network. All s… view at source ↗
Figure 2
Figure 2. Figure 2: Patient-level AUC for the nine ECG foundation models tested on the same BrSwiss test set. Each model group contains six bars representing the three evaluation strategies (blue = From scratch training; orange = Linear probing; and red = Full fine-tuning) trained on both the full BrSwiss-100% training set (full opacity) and a reduced 3% subset (reduced opacity). Models are ordered by parameter count. Error b… view at source ↗
Figure 3
Figure 3. Figure 3: Patient-level AUC for the nine ECG foundation models under the three adaptation strategies (blue = From scratch training; orange = Linear probing; and red = Full fine-tuning). The left panel shows models trained and tested on the reduced BrSwiss training set (BrSwiss-3%); the right panel shows models trained and tested on the HUCA dataset. Models are ordered by parameter count. Error bars denote 95% bootst… view at source ↗
Figure 4
Figure 4. Figure 4: Patient-level AUC for the nine ECG foundation models across all six experimental configurations, under the three adaptation strategies (blue = From scratch training; orange = Linear probing; and red = Full fine-tuning). Rows correspond to the training source: the full BrSwiss training set (BrSwiss-100%, top), the reduced BrSwiss training set (BrSwiss-3 %), and the HUCA training set (bottom). The left colum… view at source ↗
Figure 5
Figure 5. Figure 5: Integrated Gradients saliency maps on the BrSwiss-100% dataset for a representative Brugada syndrome patient, under the three adaptation strategies. Each panel shows all 12 ECG leads with the attribution heatmap overlaid. Each map is normalised by its own maximum absolute value. B. Zanchi et al.: Preprint submitted to Elsevier Page 27 of 19 [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Integrated Gradients saliency maps on the BrSwiss-3% dataset for a representative Brugada syndrome patient, under the three adaptation strategies. Each panel shows all 12 ECG leads with the attribution heatmap overlaid. Each map is normalised by its own maximum absolute value. B. Zanchi et al.: Preprint submitted to Elsevier Page 28 of 19 [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Integrated Gradients saliency maps on the HUCA dataset for a representative Brugada syndrome patient, under the three adaptation strategies. Each panel shows all 12 ECG leads with the attribution heatmap overlaid. Each map is normalised by its own maximum absolute value. B. Zanchi et al.: Preprint submitted to Elsevier Page 29 of 19 [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 13 linked inside Pith

  1. [1]

    Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network.Nature medicine, 25(1):65–69, 2019

    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(1):65–69, 2019

  2. [2]

    Artificial intelligence in electrocardiography: From automated arrhythmia detection to predicting hidden cardiovascular disease.Cureus, 17(10), 2025

    Ramy Elantary and Samar Othman. Artificial intelligence in electrocardiography: From automated arrhythmia detection to predicting hidden cardiovascular disease.Cureus, 17(10), 2025

  3. [3]

    Performance of a convolutional neural network and explainability technique for 12-lead electrocardiogram interpretation.JAMA cardiology, 6(11):1285–1295, 2021

    J Weston Hughes, Jeffrey E Olgin, Robert Avram, Sean A Abreau, Taylor Sittler, Kaahan Radia, Henry Hsia, Tomos Walters, Byron Lee, Joseph E Gonzalez, et al. Performance of a convolutional neural network and explainability technique for 12-lead electrocardiogram interpretation.JAMA cardiology, 6(11):1285–1295, 2021

  4. [4]

    Performance of a convolutional neural network derived from an ecg database in recognizing myocardial infarction.Scientific reports, 10(1):8445, 2020

    Hisaki Makimoto, Moritz Höckmann, Tina Lin, David Glöckner, Shqipe Gerguri, Lukas Clasen, Jan Schmidt, Athena Assadi-Schmidt, Alexandru Bejinariu, Patrick Müller, et al. Performance of a convolutional neural network derived from an ecg database in recognizing myocardial infarction.Scientific reports, 10(1):8445, 2020

  5. [5]

    Deep learning for cardiac computer-aided diagnosis: benefits, issues & solutions.Mhealth, 3:45, 2017

    Brian CS Loh and Patrick HH Then. Deep learning for cardiac computer-aided diagnosis: benefits, issues & solutions.Mhealth, 3:45, 2017

  6. [6]

    Overcoming data scarcity in life-threatening arrhythmia detection through transfer learning.Communications Medicine, 5(1):248, 2025

    Giuliana Monachino, Beatrice Zanchi, Michael Wand, Giulio Conte, Athina Tzovara, and Francesca Dalia Faraci. Overcoming data scarcity in life-threatening arrhythmia detection through transfer learning.Communications Medicine, 5(1):248, 2025

  7. [7]

    Advancesinmachineand deep learning for ecg beat classification: a systematic review.Frontiers in Digital Health, 7:1649923, 2025

    AllamJayaPrakash,AbdelkaderNasreddineBelkacem,IbrahimMElfadel,HerbertFJelinek,andMohamedAtef. Advancesinmachineand deep learning for ecg beat classification: a systematic review.Frontiers in Digital Health, 7:1649923, 2025. B. Zanchi et al.:Preprint submitted to ElsevierPage 11 of 19 Do ECG FMs Transfer to Rare Cardiac Diseases?

  8. [8]

    Foundationmodelsinelectrocardiogram:Areview.arXivpreprintarXiv:2410.19877, 2024

    YuHan,XiaofengLiu,XiangZhang,andChengDing. Foundationmodelsinelectrocardiogram:Areview.arXivpreprintarXiv:2410.19877, 2024

  9. [9]

    An electrocardiogram foundation model built on over 10 million recordings with external evaluation across multiple domains.arXiv preprint arXiv:2410.04133, 2024

    Jun Li, Aaron Aguirre, Junior Moura, Che Liu, Lanhai Zhong, Chenxi Sun, Gari Clifford, Brandon Westover, and Shenda Hong. An electrocardiogram foundation model built on over 10 million recordings with external evaluation across multiple domains.arXiv preprint arXiv:2410.04133, 2024

  10. [10]

    Learning general representation of 12-lead electrocardiogram with a joint-embedding predictive architecture.arXiv preprint arXiv:2410.08559, 2024

    Sehun Kim. Learning general representation of 12-lead electrocardiogram with a joint-embedding predictive architecture.arXiv preprint arXiv:2410.08559, 2024

  11. [11]

    Guidingmaskedrepresentationlearningtocapturespatio-temporalrelationship of electrocardiogram.arXiv preprint arXiv:2402.09450, 2024

    YeongyeonNa,MinjePark,YunwonTae,andSunghoonJoo. Guidingmaskedrepresentationlearningtocapturespatio-temporalrelationship of electrocardiogram.arXiv preprint arXiv:2402.09450, 2024

  12. [12]

    Zero-shot ecg classification with multimodal learning and test-time clinical knowledge enhancement.arXiv preprint arXiv:2403.06659, 2024

    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.arXiv preprint arXiv:2403.06659, 2024

  13. [13]

    YuanyuanTian,ZhiyuanLi,YanruiJin,MengxiaoWang,XiaoyangWei,LiqunZhao,YunqingLiu,JinleiLiu,andChengliangLiu.Foundation model of ecg diagnosis: Diagnostics and explanations of any form and rhythm on ecg.Cell Reports Medicine, 5(12), 2024

  14. [14]

    Hubert-ecg as a self-supervised foundation model for broad and scalable cardiac applications.medRxiv, pages 2024–11, 2024

    Edoardo Coppola, Mattia Savardi, Mauro Massussi, Marianna Adamo, Marco Metra, and Alberto Signoroni. Hubert-ecg as a self-supervised foundation model for broad and scalable cardiac applications.medRxiv, pages 2024–11, 2024

  15. [15]

    Ecg-fm:Anopenelectrocardiogramfoundationmodel.JAMIA open, 8(5):ooaf122, 2025

    KadenMcKeen,SameerMasood,AugustinToma,BarryRubin,andBoWang. Ecg-fm:Anopenelectrocardiogramfoundationmodel.JAMIA open, 8(5):ooaf122, 2025

  16. [16]

    On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021

    RishiBommasani,DrewAHudson,EhsanAdeli,RussAltman,SimranArora,SydneyvonArx,MichaelSBernstein,JeannetteBohg,Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021

  17. [17]

    Benchmarking ECG FMs: A reality check across clinical tasks

    M A Al-Masud, Juan Lopez Alcaraz, and Nils Strodthoff. Benchmarking ECG FMs: A reality check across clinical tasks. InThe Fourteenth International Conference on Learning Representations, 2026

  18. [18]

    Openecg:Benchmarkingecgfoundationmodelswithpublic1.2million records.arXiv preprint arXiv:2503.00711, 2025

    ZhijiangWan,QianhaoYu,JiaMao,WenfengDuan,andChengDing. Openecg:Benchmarkingecgfoundationmodelswithpublic1.2million records.arXiv preprint arXiv:2503.00711, 2025

  19. [19]

    An electrocardiogram multi-task benchmark with comprehensive evaluations and insightful findings.arXiv preprint arXiv:2512.08954, 2025

    Yuhao Xu, Jiaying Lu, Sirui Ding, Defu Cao, Xiao Hu, and Carl Yang. An electrocardiogram multi-task benchmark with comprehensive evaluations and insightful findings.arXiv preprint arXiv:2512.08954, 2025

  20. [20]

    Benchecg and xecg: a benchmark and baseline for ecg foundation models.arXiv preprint arXiv:2509.10151, 2025

    Riccardo Lunelli, Angus Nicolson, Samuel Martin Pröll, Sebastian Johannes Reinstadler, Axel Bauer, and Clemens Dlaska. Benchecg and xecg: a benchmark and baseline for ecg foundation models.arXiv preprint arXiv:2509.10151, 2025

  21. [21]

    Adeepfoundationmodelforelectrocardiograminterpretation:enablingraredisease detection through transfer learning.European Heart Journal-Digital Health, 6(4):619–623, 2025

    StephanieMHu,JoshuaPBarrios,andGeoffreyHTison. Adeepfoundationmodelforelectrocardiograminterpretation:enablingraredisease detection through transfer learning.European Heart Journal-Digital Health, 6(4):619–623, 2025

  22. [22]

    Self-supervised vicreg pre-training for brugada ecg detection

    Robert Ronan, Constantine Tarabanis, Larry Chinitz, and Lior Jankelson. Self-supervised vicreg pre-training for brugada ecg detection. Scientific Reports, 15(1):9396, 2025

  23. [23]

    Satoshi Kawada, Hiroshi Morita, Charles Antzelevitch, Yoshimasa Morimoto, Koji Nakagawa, Atsuyuki Watanabe, Nobuhiro Nishii, Kazufumi Nakamura, and Hiroshi Ito. Shanghai score system for diagnosis of brugada syndrome: validation of the score system and system and reclassification of the patients.Clinical Electrophysiology, 4(6):724–730, 2018

  24. [24]

    Brugada-HUCA:12-LeadECGRecordingsfortheStudyofBrugadaSyndrome.PhysioNet, February 2026

    NahuelCostaCortezandDanielGarciaIglesias. Brugada-HUCA:12-LeadECGRecordingsfortheStudyofBrugadaSyndrome.PhysioNet, February 2026. Version 1.0.0

  25. [25]

    Nils Strodthoff, Juan Miguel Lopez Alcaraz, and Wilhelm Haverkamp. Prospects for artificial intelligence-enhanced electrocardiogram as a unified screening tool for cardiac and non-cardiac conditions: an explorative study in emergency care.European Heart Journal-Digital Health, 5(4):454–460, 2024

  26. [26]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

  27. [27]

    Hubert: Self- supervised speech representation learning by masked prediction of hidden units.IEEE/ACM transactions on audio, speech, and language processing, 29:3451–3460, 2021

    Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, and Abdelrahman Mohamed. Hubert: Self- supervised speech representation learning by masked prediction of hidden units.IEEE/ACM transactions on audio, speech, and language processing, 29:3451–3460, 2021

  28. [28]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

  29. [29]

    wav2vec 2.0: A framework for self-supervised learning of speech representations.Advances in neural information processing systems, 33:12449–12460, 2020

    Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations.Advances in neural information processing systems, 33:12449–12460, 2020

  30. [30]

    Clocs: Contrastive learning of cardiac signals across space, time, and patients

    Dani Kiyasseh, Tingting Zhu, and David A Clifton. Clocs: Contrastive learning of cardiac signals across space, time, and patients. In International Conference on Machine Learning, pages 5606–5615. PMLR, 2021

  31. [31]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

  32. [32]

    A path towards autonomous machine intelligence version 0.9

    Yann LeCun. A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27.Open Review, 62(1):1–62, 2022

  33. [33]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022

  34. [34]

    Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396, 2021

    Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces.arXiv preprint arXiv:2111.00396, 2021

  35. [35]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

  36. [36]

    Holmes: health online model ensemble serving for deep learning models in intensive care units

    Shenda Hong, Yanbo Xu, Alind Khare, Satria Priambada, Kevin Maher, Alaa Aljiffry, Jimeng Sun, and Alexey Tumanov. Holmes: health online model ensemble serving for deep learning models in intensive care units. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 1614–1624, 2020

  37. [37]

    Designing network design spaces

    Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dollár. Designing network design spaces. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428–10436, 2020. B. Zanchi et al.:Preprint submitted to ElsevierPage 12 of 19 Do ECG FMs Transfer to Rare Cardiac Diseases?

  38. [38]

    Dist-pu: Positive-unlabeled learning from a label distribution perspective

    Yunrui Zhao, Qianqian Xu, Yangbangyan Jiang, Peisong Wen, and Qingming Huang. Dist-pu: Positive-unlabeled learning from a label distribution perspective. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14461–14470, 2022

  39. [39]

    Deeplearningforecganalysis:Benchmarksandinsightsfromptb-xl

    NilsStrodthoff,PatrickWagner,TobiasSchaeffter,andWojciechSamek. Deeplearningforecganalysis:Benchmarksandinsightsfromptb-xl. IEEE journal of biomedical and health informatics, 25(5):1519–1528, 2020

  40. [40]

    Beit: Bert pre-training of image transformers.arXiv preprint arXiv:2106.08254, 2021

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers.arXiv preprint arXiv:2106.08254, 2021

  41. [41]

    Amethodforstochasticoptimization

    DiederikKinga,JimmyBaAdam,etal. Amethodforstochasticoptimization. InInternationalconferenceonlearningrepresentations(ICLR), volume 5. California;, 2015

  42. [42]

    Training data-efficient image transformers & distillation through attention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. InInternational conference on machine learning, pages 10347–10357. PMLR, 2021

  43. [43]

    Convit:Improvingvisiontransformers with soft convolutional inductive biases

    Stéphaned’Ascoli,HugoTouvron,MatthewLLeavitt,AriSMorcos,GiulioBiroli,andLeventSagun. Convit:Improvingvisiontransformers with soft convolutional inductive biases. InInternational conference on machine learning, pages 2286–2296. PMLR, 2021

  44. [44]

    Scaling laws for neural language models.arXiv preprint arXiv:2001.08361, 2020

    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.arXiv preprint arXiv:2001.08361, 2020

  45. [45]

    Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 2022

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. Training compute-optimal large language models.arXiv preprint arXiv:2203.15556, 2022

  46. [46]

    Inflationoftestaccuracyduetodataleakageindeeplearning-basedclassification of oct images.Scientific Data, 9(1):580, 2022

    IulianEmilTampu,AndersEklund,andNedaHaj-Hosseini. Inflationoftestaccuracyduetodataleakageindeeplearning-basedclassification of oct images.Scientific Data, 9(1):580, 2022

  47. [47]

    Axiomaticattributionfordeepnetworks

    MukundSundararajan,AnkurTaly,andQiqiYan. Axiomaticattributionfordeepnetworks. InInternationalconferenceonmachinelearning, pages 3319–3328. PMLR, 2017. B. Zanchi et al.:Preprint submitted to ElsevierPage 13 of 19 Do ECG FMs Transfer to Rare Cardiac Diseases? Figure 1:Schematic overview of the experimental settings in our study. In the Train from Scratch ap...