acceptodds
Under review as a conference paper at ICLR 2027

BiomECG: Learning Transferable ECG Representations from Small Cohorts

Abstract

Self-supervised learning (SSL) has become a standard approach to ECG representation learning, reducing the need for large clinically annotated datasets. Foundation models have been proposed based on very large cohorts. However, they may not be suited to every target population, making lightweight SSL pipelines that work with smaller datasets beneficial. We propose biomECG, an ECG biometric authentication framework for SSL pretraining and downstream evaluation, training compact Siamese encoders to classify pairs of recordings as belonging to the same or different patients. We obtain results comparable to those of a state-of-the-art ECG foundation model developed from over 10 million clinically annotated ECGs, using approximately three orders of magnitude fewer patients for pretraining and over 100-fold fewer parameters. Crucially, frozen encoders combined through late fusion of ECG leads often outperform full fine-tuning at limited and intermediate downstream training sizes. Late fusion also improves mortality discrimination over early fusion, particularly with smaller training cohorts. We evaluate in detail the impact of both the SSL pretraining dataset size and the downstream training dataset size using MIMIC and CODE-15. Pretraining on separate cohorts of similar size yields comparable transfer performance on MIMIC and CODE-15, indicating that representations learned through biometric pretraining capture information relevant across datasets. Code for biomECG, including pretraining and evaluation on MIMIC and CODE-15, is available at https://github.com/anonymoussubmission1000/biomECG-iclr-anon.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
ecg representation learning - self-supervised learning - biometric authentication - siamese networks - transfer learning - late fusion - frozen-feature evaluation - mortality prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_NMkVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates whether ECG data from a small number of patients can be used to learn representations that transfer to other clinical tasks. The authors propose biomECG, which is pretrained to determine whether two ECG recordings belong to the same patient. They also compare different lead fusion methods and fine-tuning strategies. Experiments are conducted on MIMIC, CODE-15, and CPET, with a primary focus on mortality prediction, along with age and sex prediction. The results show that pretraining on only 1359 patients achieves performance close to ECGFounder on mortality prediction, while the benefits on other tasks remain limited.

Critical strengths

The paper directly examines whether large patient cohorts are necessary for learning effective ECG representations. Pretraining on a small cohort using patient identity verification achieves performance comparable to a large foundation model on two mortality prediction datasets. This finding makes the research direction worth exploring, especially for medical applications with limited data.

The authors use patient data from different sources for pretraining and evaluate transfer performance on MIMIC and CODE-15. Different pretraining sources produce similar results, and even CPET data enables effective transfer. This suggests that identity verification pretraining can capture ECG information that generalizes across datasets to some extent.

Critical weaknesses

Prior work has already explored ECG representation learning based on patient identity, and the Siamese architecture used in this paper largely follows existing methods. Although the paper shows that effective representations can be learned from a small cohort, it lacks direct comparisons with patient level contrastive learning or other common self-supervised methods under matched data sizes and model architectures. It is therefore difficult to determine whether the performance comes from the identity verification objective itself or from the combined effects of the ECG data, architecture, and training strategy. The comparison with ECGFounder does not resolve this issue, since the two methods differ substantially in pretraining data, supervision, and model capacity.

The main benefits are concentrated on mortality prediction, with no comparable gains on age or sex prediction. Mortality prediction also uses age and sex as additional inputs, but the paper lacks a baseline using only these clinical variables. This makes it difficult to isolate the contribution of the learned ECG representations. Notably, on the full MIMIC training set, the late fusion model trained from scratch achieves an AUROC of 0.838, slightly higher than the 0.835 achieved with frozen biomECG. In addition, a proportion of patients in MIMIC are censored before one year but are treated as negative cases. This further raises concerns about the reliability of the mortality prediction results.

The paper is also poorly written and difficult to follow. Adding one or two diagrams in the Intro or Method section would help readers understand the biomECG architecture and training procedure.

Overall, the paper includes a substantial amount of experimentation, but the experiments are poorly organized. Few directly address the central research questions, and the evidence is insufficient to support the main claims.

Questions for authors
  1. Does the identity verification objective offer a clear advantage over existing alternatives? Under matched pretraining cohort sizes, model architectures, and downstream training settings, does biomECG outperform existing patient level contrastive learning or other self-supervised methods?
  2. How much of the improvement in mortality prediction actually comes from the pretrained representations? How much additional predictive value does biomECG provide over a baseline using only age and sex? Do the main conclusions still hold when censoring is handled appropriately?
  3. Is the advantage of late fusion mainly due to model ensembling? After controlling for model capacity and ensemble size, does modeling each lead independently still outperform joint lead modeling?

If the authors can adequately address these questions and provide clear, focused experiments that support the main claims, I would be willing to raise my score.

5 comments
AuthorsOpenReview ↗

Response to Reviewer NMkV

Thank you for your careful reading and for suggesting concrete ways to strengthen the paper. The three reviews converged on the need for stronger baselines, clearer evidence about what drives mortality prediction, and a more focused presentation. Your feedback led us to substantially revise the study.

Overall changes compared to the previous version

The new experiments support three contributions:

  1. Competitive mortality representations from small cohorts. Several SSL methods pretrained on 1,359 patients approach several much larger ECG foundation models on MIMIC and CODE-15. This had been seen for biometry pretraining, and turned out to be a more general result. Moreover, comparisons now include several foundation models.
  2. Lead-wise learning beyond ensembling. The previous version did not compare lead-wise learning with an ensemble approach for multilead. Across three SSL objectives on a shared ResNet-18 backbone, late fusion generally improves on an eight-model multilead ensemble with comparable encoder capacity.
  3. Simple and transferable biometric pretraining. These were the already present results regarding biometry: biomECG learns competitive representations through direct patient verification without an explicit contrastive loss. Scaling pretraining on MIMIC improves frozen transfer to CODE-15, where biomECG achieves a higher mean AUROC than all three foundation models evaluated.

We added the new baselines to the anonymous GitHub repository.

Q1. SSL and foundation baselines

We now compare with PCLR, JEPA, and CPC using the same pretraining patients and downstream protocol as biomECG. PCLR is the closest identity baseline. JEPA and CPC offer distinct predictive objectives. PCLR and JEPA share biomECG’s ResNet-18. CPC keeps its native multilead CNN–S4 given its temporal objective and higher training cost (Sec. 4.2, Appendix F).

We added ECG-JEPA and ECG-CPC alongside ECGFounder, the top three foundation models in Al-Masud et al.’s frozen mortality benchmark (2026, Tab. 4). With late fusion, biomECG matches JEPA and generally improves on adapted PCLR (Fig. 3). Local CPC scores highest among local methods on MIMIC, with a larger encoder.

\text{Method} & \text{MIMIC AUROC} & \text{CODE-15 AUROC} \\\\ \hline \text{biomECG} & 0.84 & 0.84 \\\\ \text{PCLR} & 0.83 & 0.81 \\\\ \text{JEPA} & 0.84 & 0.84 \\\\ \text{ECG-CPC, best foundation reference} & 0.86 & 0.85 \end{array}$$ **Q2a. Contribution of ECG features beyond age and sex** Fig. 3 compares age and sex with and without frozen ECG features using matched MLPs, training subsets, and checkpoint selection. Age and sex alone yield **0.77 on MIMIC** and **0.80–0.81 on CODE-15**. With full training, adding late-fusion biomECG yields **0.84 on both**. At small **downstream** budgets, adding ECG features lowers AUROC, likely because the MLP has too few examples for the higher-dimensional inputs. Full-training biomECG AUROCs (Appendix H.1): $$\begin{array}{lcc} \text{Inputs} & \text{MIMIC} & \text{CODE-15} \\\\ \hline \text{Age and sex only} & 0.77 & 0.81 \\\\ \text{biomECG features only} & 0.80 & 0.77 \\\\ \text{biomECG + age and sex} & 0.84 & 0.84 \end{array}$$ **Q2b. Sensitivity to early censoring** We agree that patients with less than one year of follow-up and no recorded death may have unknown one-year outcomes. We followed the convention of Lukyanenko et al. (2025). Their supplement discusses how excluding patients with short or zero follow-up changes the evaluated population and can increase the observed event proportion. Appendix C tests how this choice affects AUROC and model rankings. Both analyses retain observed deaths within one year. - **Test only.** Excluding early-censored test patients with predictions fixed barely changes AUROC on CODE-15 and modestly lowers it on MIMIC, with nearly unchanged rankings. - **With retraining.** We filter all partitions and retrain the heads with frozen encoders. Results are similar, with a few swaps between close models. Filtering without replacements reduces training sizes and may partly explain the MIMIC decrease. **Q3. Ensemble control** The early ensemble averages predictions from eight independently pretrained models, each using all eight leads jointly and the same patient cohorts. Late fusion uses one model per lead. Encoder capacity is similar (1.70M early vs. 1.69M late). Across biomECG, PCLR, and JEPA, late fusion beats single early fusion in 18/18 settings and the early ensemble in 15/18 (Fig. 3). **Frozen biomECG vs. training from scratch on full MIMIC** Training from scratch slightly outperforms frozen biomECG on full MIMIC with 1,359 pretraining patients. Increasing the pretraining cohort to 8,000 patients is enough for frozen biomECG to surpass the scratch baseline (Figures 4). **Presentation** We worked to make the study as clear and focused as possible. We simplified the prose, added two diagrams, and put the key experiments first.
Reviewer_NMkVOpenReview ↗

To Authors

Thank you for the rebuttal. I have reread the updated PDF. The new method figure and additional experiments help clarify the paper.

However, my second concern remains. Age and sex alone already achieve strong mortality prediction performance, while the additional benefit of biomECG features is limited, especially on CODE-15. The link between patient identification during pretraining and mortality prediction also remains unclear.

My main concern is whether the first two steps, biometric pretraining and representation learning, are really necessary for the final mortality prediction task.

For now, I maintain my rating.

AuthorsOpenReview ↗

Response to Reviewer NMkV - concern 1

Thank you for taking the time to read our rebuttal version and raising these concerns. We believe however that these concerns are already addressed by our results. We give more details below.

Concern 1: Age and sex alone already achieve strong mortality prediction performance, while the additional benefit of biomECG features is limited, especially on CODE-15.

Lukyanenko et al. (2025) [1] benchmarked ECG-based survival models and showed that combining learned ECG representations with age and sex improves mortality prediction compared with ECG-only or age-and-sex-only models. We adopt this evaluation framework: all ECG representations are systematically concatenated with age and sex for downstream mortality prediction, allowing comparison under matched downstream conditions.

The smaller incremental benefit of ECG features on CODE-15 than on MIMIC is also documented and discussed in Lukyanenko et al. (2025). Our results in Figure 3 remain consistent: adding biometric ECG representations to age and sex leads to a +0.04 AUROC increase in mortality prediction on CODE-15 with full downstream training. Despite the increase being modest, this observation should be distinguished from the question of how well biomECG compares with other ECG representations under matched conditions: this gain is among the largest obtained with the compared ECG representations.

Overall these observations does not diminish our claims:

  • the representations that we obtained through pretraining an encoder on a very small subset of 1359 patients with the biometric technic and freezing it
  • contains information useful to mortality prediction

We have updated the Figure 3 caption to emphasize that all ECG representations are concatenated with age and sex, and revised Section 5.1 text to begin with the demographic baseline and explain the additional contribution of ECG representations more explicitly, as these points may not have been sufficiently clear in the previous version. Thank you for the opportunity to emphasize these points.

For further details : For these models that take [ECG representation, age, sex] as input, performance depends on both:

  • the quality of the ECG representation
  • and the size of the downstream mortality training cohort. Figure 3 varies both: it compares different representations, including biomECG, across downstream training sizes. All ecg representations (not only biometric representation) are concatenated with age and sex under the same downstream protocol.

The results with very small downstream size (1000 or 8000) is shared across representation methods, rather than being specific to biomECG representations. It highlights that learning to extract prognostic information from ECG beyond age and sex requires more downstream training patients on CODE-15, rather than a limitation specific to biomECG :

  • for MIMIC, benefit appears starting at 1000 downstream patients for MIMIC
  • for CODE15, benefit appears somewhere in between 8000 patients (no benefit of any kind of ECG representation compared to age+sexe alone) and full training (63k patients, clear benefit when adding ecg representations to age+sexe). This information might be of importance for future applications.

[1] Platon Lukyanenko, Joshua Mayourian, Mingxuan Liu, John K. Triedman, Sunil J. Ghelani, and William G. LaCava. Deep survival analysis from adult and pediatric electrocardiograms: a multi-center benchmark study. BioDataMining, 19(1), December2025. ISSN1756-0381. doi: https://doi.org/10.1186/s13040-025-00510-4.

Reviewer_ViHHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper adapts the Siamese ResNet architecture of Thiebaud et al. (2026) to learn ECG representations through patient authentication and transfer them to downstream prediction tasks. The authors compare three adaptation regimes: frozen transfer, full fine-tuning, and training from scratch, as well as early fusion and prediction-level late fusion. Experiments use MIMIC, CODE-15, and CPET as biometric pretraining cohorts. The primary downstream task is one-year mortality prediction on MIMIC and CODE-15, with age prediction and sex classification as additional evaluation tasks. The results suggest that biometric pretraining on relatively small patient cohorts can achieve mortality prediction performance comparable to ECGFounder in several settings, while using substantially fewer encoder parameters.

Critical strengths
  1. The research question is practically relevant, and the experimental results demonstrate promising data efficiency. The results suggest that biometric pretraining on a relatively small patient cohort can achieve competitive mortality prediction performance. In addition, biomECG uses substantially fewer encoder parameters than ECGFounder, highlighting its potential for resource-constrained settings.
  2. The paper provides a systematic empirical evaluation across multiple experimental settings. It investigates the effects of different biometric pretraining cohorts, pretraining cohort sizes, downstream training-set sizes, adaptation regimes, and lead fusion strategies. The evaluation on both MIMIC and CODE-15 further supports the empirical analysis.
Critical weaknesses
  1. The baseline comparisons are insufficient. ECGFounder (Li et al., 2025) is the only external pretrained model used for comparison. Although ECGFounder is a relevant baseline, comparing against a single pretrained model is insufficient to establish the competitiveness of biomECG. Given that this submission is for ICLR 2027, a broader comparison with representative ECG foundation models and self-supervised learning methods is necessary. The authors could refer to the already cited Benchmarking ECG FMs: A Reality Check Across Clinical Tasks (Al-Masud et al., 2026) to select additional baselines. At the very least, the authors should provide a more convincing justification for choosing ECGFounder as the sole external pretrained baseline.
  2. The contribution of the learned ECG representations to mortality prediction is not sufficiently examined. As described in Section 3.2, the mortality prediction head uses both ECG representations and demographic covariates, including chronological age and sex. However, the experiments do not clearly distinguish the contributions of these different sources of information. Since age and sex may already provide substantial predictive information for one-year mortality, it is unclear how much of the reported performance can be attributed to the ECG representations learned through small-scale biometric pretraining. An ablation study comparing age and sex alone, ECG representations alone, and their combination would help clarify the contribution of the learned representations and better support the paper's central claim.
Questions for authors

As described in Section 3.2, the mortality prediction head uses both ECG representations and demographic covariates, including chronological age and sex. Could the authors provide an ablation study comparing age and sex alone, ECG representations alone, and their combination? This would help clarify the additional predictive value of the ECG representations learned through biometric pretraining.

Appendix C states that patients censored before 365 days without a recorded death are assigned to the negative class. Could the authors clarify how this labeling choice may affect the mortality prediction results? A sensitivity analysis excluding these patients would help determine whether the main findings and relative model rankings remain consistent.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ViHH

Thank you for your careful reading and for suggesting concrete ways to strengthen the paper. The three reviews converged on the need for stronger baselines, clearer evidence about what drives mortality prediction, and a more focused presentation. Your feedback led us to substantially revise the study.

Overall changes compared to the previous version

The new experiments support three contributions:

  1. Competitive mortality representations from small cohorts. Several SSL methods pretrained on 1,359 patients approach several much larger ECG foundation models on MIMIC and CODE-15. This had been seen for biometry pretraining, and turned out to be a more general result. Moreover, comparisons now include several foundation models.
  2. Lead-wise learning beyond ensembling. The previous version did not compare lead-wise learning with an ensemble approach for multilead. Across three SSL objectives on a shared ResNet-18 backbone, late fusion generally improves on an eight-model multilead ensemble with comparable encoder capacity.
  3. Simple and transferable biometric pretraining. These were the already present results regarding biometry: biomECG learns competitive representations through direct patient verification without an explicit contrastive loss. Scaling pretraining on MIMIC improves frozen transfer to CODE-15, where biomECG achieves a higher mean AUROC than all three foundation models evaluated.

We added the new baselines to the anonymous GitHub repository.

Q1. Contribution of ECG features beyond age and sex

Fig. 3 compares age and sex with and without frozen ECG features using matched MLPs, training subsets, and checkpoint selection. Age and sex alone yield 0.77 on MIMIC and 0.80–0.81 on CODE-15. With full training, adding late-fusion biomECG yields 0.84 on both. At small downstream budgets, adding ECG features lowers AUROC, likely because the MLP has too few examples for the higher-dimensional inputs.

Appendix H.1 adds ECG-only results for biomECG, PCLR, JEPA, and ECG-CPC. The table shows biomECG with full downstream training.

\hline \text{Inputs} & \text{MIMIC AUROC} & \text{CODE-15 AUROC} \\\\ \hline \text{Age and sex only} & 0.77 & 0.81 \\\\ \text{biomECG features only} & 0.80 & 0.77 \\\\ \text{biomECG features + age and sex} & 0.84 & 0.84 \\\\ \hline \end{array}$$ **Q2. Sensitivity to early censoring** We agree that short follow-up does not establish one-year survival. We followed Lukyanenko et al. (2025), whose supplement explains how excluding patients with short or zero follow-up can change the evaluated population and increase the observed event proportion. Appendix C reports two analyses, both retaining observed deaths within one year. - **Test only.** Excluding early-censored test patients while keeping the original predictions changes AUROC little on CODE-15 and lowers it modestly on MIMIC, with nearly unchanged rankings. - **With retraining.** We exclude these patients from training, validation, and test and retrain the mortality heads with frozen encoders. The same pattern holds, with a few swaps between close models. Filtering the nested cohorts without replacements reduces training sizes, which may contribute to the MIMIC decrease. Appendix C also reports retained patient and event counts. **Broader baseline comparisons** Following your suggestion, we added ECG-JEPA and ECG-CPC alongside ECGFounder, the three highest-scoring foundation models in the frozen MIMIC mortality benchmark of Al-Masud et al. (2026, Table 4). We pretrained PCLR, JEPA, and CPC on the same patients as biomECG. PCLR is the closest patient-identity baseline. JEPA and CPC offer distinct predictive objectives. PCLR and JEPA share biomECG’s ResNet-18. CPC’s temporal prediction architecture and higher training cost motivated retaining its native multilead CNN–S4 (Section 4.2, Appendix F). Figure 3 reports mean frozen AUROC with full downstream training. The local methods shown use 1,359 pretraining patients. $$\begin{array}{lcc} \hline \text{Method} & \text{MIMIC AUROC} & \text{CODE-15 AUROC} \\\\ \hline \text{biomECG, late fusion} & 0.84 & 0.84 \\\\ \text{PCLR, late fusion} & 0.83 & 0.81 \\\\ \text{JEPA, late fusion} & 0.84 & 0.84 \\\\ \text{ECG-CPC, best foundation reference} & 0.86 & 0.85 \\\\ \hline \end{array}$$ With 50,000 MIMIC pretraining patients, frozen biomECG reaches **0.85 ± 0.01 AUROC on CODE-15**, above all three foundation references in the scaling analysis (0.83–0.84 mean AUROC, Figure 4). **Presentation and additional experiments** We revised the prose, added two diagrams in place of the dense mathematical description, and put the main comparisons first. The new eight-model early-fusion ensemble tests whether lead-wise learning helps beyond prediction averaging. See our responses to Reviewers 9Dcj and NMkV for details.

Comments on the paper

4 comments
AuthorsOpenReview ↗

Response to Reviewer 9Dcj

Thank you for your careful reading and for suggesting concrete ways to strengthen the paper. The three reviews converged on the need for stronger baselines, clearer evidence about what drives mortality prediction, and a more focused presentation. Your feedback led us to substantially revise the study.

Overall changes compared to the previous version

The new experiments support three contributions:

  1. Competitive mortality representations from small cohorts. Several SSL methods pretrained on 1,359 patients approach several much larger ECG foundation models on MIMIC and CODE-15. This had been seen for biometry pretraining, and turned out to be a more general result. Moreover, comparisons now include several foundation models.
  2. Lead-wise learning beyond ensembling. The previous version did not compare lead-wise learning with an ensemble approach for multilead. Across three SSL objectives on a shared ResNet-18 backbone, late fusion generally improves on an eight-model multilead ensemble with comparable encoder capacity.
  3. Simple and transferable biometric pretraining. These were the already present results regarding biometry: biomECG learns competitive representations through direct patient verification without an explicit contrastive loss. Scaling pretraining on MIMIC improves frozen transfer to CODE-15, where biomECG achieves a higher mean AUROC than all three foundation models evaluated.

We added the new baselines to the anonymous GitHub repository.

Q1. Sensitivity to early censoring

We lack evidence of complete one-year ascertainment for these patients. We followed Lukyanenko et al. (2025), whose supplement discusses how excluding short or zero follow-up changes the cohort and event proportion.

Both analyses in Appendix C retain observed deaths within one year.

  • Test only. Excluding early-censored test patients with predictions fixed barely changes AUROC on CODE-15 and modestly lowers it on MIMIC, with nearly unchanged rankings.
  • With retraining. We filter all partitions and retrain the heads with frozen encoders. Results are similar, with a few swaps between close models. Filtering the nested cohorts without replacements reduces training sizes, which may contribute to the MIMIC decrease.

Q2. Contribution of ECG features beyond age and sex

Fig. 3 compares age and sex with and without frozen ECG features using matched MLPs, training subsets, and checkpoint selection. Age and sex alone yield 0.77 on MIMIC and 0.80–0.81 on CODE-15. With full training, adding late-fusion biomECG yields 0.84 on both. At small budgets, adding ECG features lowers AUROC, likely because the MLP has too few examples for the higher-dimensional inputs.

Appendix H.1 adds ECG-only results. Below are full-training biomECG AUROCs.

\hline \text{Inputs} & \text{MIMIC} & \text{CODE-15} \\\\ \hline \text{Age and sex only} & 0.77 & 0.81 \\\\ \text{biomECG features only} & 0.80 & 0.77 \\\\ \text{biomECG + age and sex} & 0.84 & 0.84 \\\\ \hline \end{array}$$ **Q3a. Validation and training data efficiency** We split patients 70/15/15 and vary only the nested training subsets. The 23,863 MIMIC validation patients serve only for checkpoint selection. We now specify **downstream training data efficiency** at a fixed validation-label budget (Section 4.3, Appendix B, Discussion). **Q3b. Scaling** Across five partitions, scaling MIMIC pretraining from 100 to 50,000 patients raises frozen AUROC from 0.82 to 0.85 on MIMIC and from 0.83 to **0.85 ± 0.01 on CODE-15**, above all three foundation references there (Figure 4). **Q4b. SSL and foundation baselines** We now compare with PCLR, JEPA, and CPC SSL-methods using the same pretraining patients as biomECG. PCLR is the closest identity baseline. JEPA and CPC offer distinct predictive objectives. PCLR and JEPA share biomECG’s ResNet-18. CPC keeps its native multilead CNN–S4 given its temporal objective and higher training cost (Section 4.2, Appendix F). We added ECG-JEPA and ECG-CPC alongside ECGFounder, the top three foundation models in Al-Masud et al.’s frozen mortality benchmark (2026, Table 4). With late fusion, biomECG matches JEPA and generally improves on adapted PCLR (Figure 3). $$\begin{array}{lcc} \hline \text{Method} & \text{MIMIC AUROC} & \text{CODE-15 AUROC} \\\\ \hline \text{biomECG} & 0.84 & 0.84 \\\\ \text{PCLR} & 0.83 & 0.81 \\\\ \text{JEPA} & 0.84 & 0.84 \\\\ \text{ECG-CPC, best foundation reference} & 0.86 & 0.85 \\\\ \hline \end{array}$$ **Q4a. Ensemble control** The early ensemble averages eight independently pretrained models, each using all eight leads and the same patient cohorts. Encoder capacity is similar (1.70M early vs. 1.69M late). Across biomECG, PCLR, and JEPA, late fusion beats single early fusion in 18/18 settings and the early ensemble in 15/18 (Figure 3). **Presentation** We revised the prose, replaced dense math with two diagrams, and put the main comparisons first.

The additional analyses and baselines substantially strengthen the paper, particularly regarding censoring sensitivity, the contribution of ECG features beyond demographic covariates, and lead-wise fusion versus ensembling. Most of my technical concerns have therefore been alleviated. However, methodological novelty remains limited, and the evidence for added value in the smallest-data settings is still mixed. Overall, I view this as a useful empirical study, but the contribution remains somewhat incremental. I therefore maintain my weak reject recommendation.

Thank you for reassessing the manuscript and acknowledging that the additional analyses alleviate most of your technical concerns. We would like to clarify “the smallest-data settings,” as two distinct cohort sizes are involved.

Are you referring to the smallest downstream training cohorts in Figure 3, also discussed in our response to Q2? Our statement that “at small budgets, adding ECG features lowers AUROC” refers to downstream training. Throughout Figure 3, each pretrained encoder and its extracted ECG representations remain unchanged. Only the number of patients used to train the downstream mortality predictor varies. All locally pretrained SSL models in this figure were pretrained on 1,359 patients. We have revised the caption to make this explicit.

With very few downstream patients, adding these fixed ECG embeddings to age and sex can lower AUROC, plausibly because the downstream MLP has insufficient examples for the higher-dimensional inputs. The same phenomenon is observed with representations from the foundation models evaluated. As the downstream cohort grows, the predictor can increasingly exploit the information already contained in these representations, with gains emerging from around 8,000 patients. It reflects the downstream predictor’s data requirements, while the representations themselves were learned using the same small pretraining cohort.

We use this small pretraining budget throughout the main comparisons. Section 5.2 explicitly varies it and provides further evidence of data efficiency in learning useful representations. In Figure 4, pretraining on just 100 MIMIC patients yields downstream AUROCs of approximately 0.82 on MIMIC and 0.83 on CODE-15, already comparable to the foundation model references on CODE-15. With 50,000 pretraining patients, frozen transfer reaches 0.85 on CODE-15, exceeding the mean AUROC of all three foundation models evaluated. We consider this evidence that useful, transferable mortality representations can be learned with remarkably little pretraining data a substantive empirical contribution.

Regarding methodological novelty, biomECG develops a simple patient-verification approach using binary cross-entropy, without an explicit contrastive loss. With matched pretraining patients and a shared ResNet-18 backbone, it generally outperforms adapted PCLR in late fusion. Appendix G.2 illustrates that same-patient embeddings nevertheless cluster naturally. These results support direct patient verification as an effective alternative for learning transferable representations.

Finally, the lead-wise learning contribution is supported across SSL objectives by the new ensemble controls. Late fusion improves over single early fusion in 18/18 settings and over a capacity-matched eight-model multilead ensemble in 15/18 settings, demonstrating benefits beyond prediction averaging. We acknowledge that these methodological choices build on prior work, while their evaluation provides evidence about how little pretraining data is needed, how well the resulting representations transfer, and how lead-wise learning improves performance.

AuthorsOpenReview ↗

Precision in response to Reviewer 9Dcj

Thank you again for your feedbacks. We have provided additional clarifications to Reviewer NMkV that are also relevant to your concerns, and would like to build on our previous response to emphasize points that may not have been sufficiently clear. We have also revised the manuscript to make these distinctions more explicit.

Concern 1: Mixed evidence for added value in the smallest-data settings.

We would like to emphasize the distinction between two different data budgets: the number of patients used for representation pretraining (as few as 1,359), and the number used for downstream mortality prediction (1,000 to the full cohort). Our main data-efficiency claim concerns the pretraining cohort.

Lukyanenko et al. (2025) [1] benchmarked ECG-based survival models and showed that combining learned ECG representations with age and sex improves mortality prediction compared with ECG-only or age-and-sex-only models. We adopt this evaluation framework: all ECG representations are systematically concatenated with age and sex for downstream mortality prediction, allowing comparison under matched downstream conditions.

The smaller incremental benefit of ECG features on CODE-15 than on MIMIC is also documented and discussed in Lukyanenko et al. (2025). Our results in Figure 3 remain consistent: adding biometric ECG representations to age and sex leads to a +0.04 AUROC increase in mortality prediction on CODE-15 with full downstream training. Despite the increase being modest, this observation should be distinguished from the question of how well biomECG compares with other ECG representations under matched conditions: this gain is among the largest obtained with the compared ECG representations.

Regarding the smallest downstream training settings, the trends are shared across representation methods, rather than being specific to biomECG. Figure 3 compares different ECG representations across downstream training sizes, always concatenated with age and sex under the same protocol.

  • On MIMIC, the benefit of ECG representations emerges with as few as 1,000 downstream patients and increases substantially with training cohort size.
  • On CODE-15, no evaluated ECG representation improves over age and sex alone at 8,000 downstream patients, whereas clear gains emerge with the full cohort (approximately 63,000 patients). This highlights that learning to exploit prognostic information from ECG beyond age and sex requires more downstream training patients on CODE-15, rather than reflecting a limitation specific to biomECG. We believe this dataset-dependent behavior is itself an informative result for future applications.

Importantly, representations learned through biometric pretraining on only 1,359 patients remain competitive with those from much larger pretrained models, under matched downstream conditions. This is the central data-efficiency finding, independently of the downstream cohort size required to exploit these representations.

To make these distinctions clearer, we have revised Section 5.1 to explicitly separate the demographic baseline from the additional contribution of ECG representations, emphasize that the observed trends apply across representation methods, and clarify their dependence on downstream training size.

Concern 2: Limited methodological novelty.

As detailed in our previous response, we believe the revised study makes substantive methodological and empirical contributions.

First, small-cohort SSL competes with much larger foundation models in frozen mortality transfer. Second, lead-wise learning provides benefits beyond ensembling, outperforming single early fusion in 18/18 settings and a capacity-matched eight-model ensemble in 15/18.

Finally, biomECG introduces a simple patient-verification approach using binary cross-entropy, without an explicit contrastive loss. With matched pretraining patients and a shared ResNet-18 backbone, it performs comparably to JEPA and generally outperforms adapted PCLR (Figure 3). Increasing biometric pretraining size further improves frozen mortality transfer, eventually outperforming all three much larger foundation-model references on CODE-15 (Figure 4).

We believe these findings constitute a meaningful contribution to ECG representation learning, both through the biometric approach and through new empirical evidence on data efficiency, lead-wise learning, and prognostic transfer.

We thank the reviewer again for the constructive feedback and hope these additional clarifications better convey the contributions and scope of our work.

[1] Platon Lukyanenko, Joshua Mayourian, Mingxuan Liu, John K. Triedman, Sunil J. Ghelani, and William G. La Cava. Deep survival analysis from adult and pediatric electrocardiograms: a multi-center benchmark study. BioData Mining, 19(1), December 2025. https://doi.org/10.1186/s13040-025-00510-4.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.