Abstract

Electrocardiography (ECG) is a low-cost, widely used modality for diagnosing electrical abnormalities like atrial fibrillation by capturing the heart’s electrical activity. However, it cannot directly measure cardiac morphological phenotypes, such as left ventricular ejection fraction (LVEF), which typically require echocardiography (Echo). Predicting these phenotypes from ECG would enable early, accessible health screening. Existing self-supervised methods suffer from a representational mismatch by aligning ECGs to single-view Echos, which only capture local, spatially restricted anatomical snapshots. To address this, we propose Echo2ECG, a multimodal self-supervised learning framework that enriches ECG representations with the heart’s morphological structure captured in multi-view Echos. We evaluate Echo2ECG as an ECG feature extractor on two clinically relevant tasks that fundamentally require morphological information: (1) classification of structural cardiac phenotypes across three datasets, and (2) retrieval of Echo studies with similar morphological characteristics using ECG queries. Our extracted ECG representations consistently outperform those of state-of-the-art unimodal and multimodal baselines across both tasks, despite being 18x smaller than the largest baseline. These results demonstrate that Echo2ECG is a robust, powerful ECG feature extractor. Our code is accessible at: https://github.com/michelleespranita/Echo2ECG.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1932_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/michelleespranita/Echo2ECG

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LimMic_Echo2ECG_MICCAI2026,
        author = { Liman, Michelle Espranita AND Turgut, Özgün AND Müller, Alexander AND Martens, Eimo AND Rueckert, Daniel AND Müller, Philip},
        title = { { Echo2ECG: Enhancing ECG Representations with Cardiac Morphology from Multi-view Echos } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16896},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a multimodal self-supervised learning framework to improve ECG representations by incorporating multi-view echocardiography (Echo). The paper aims to address the challenge of measuring cardiac morphological phenotypes, such as left ventricular ejection fraction (LVEF), which traditionally require Echo. On the other hand, ECG is a more accessible and cost-effective modality. Therefore, training a model that aligns ECG with Echo could help estimate LVEF using ECG alone. Compared to previous work such as EchoingECG, the authors utilize multi-view echocardiography without relying on the text modality.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • Usability: Incorporating structural information from Echo into ECG representations addresses a meaningful real-world clinical problem. -Well-written paper: The manuscript is clearly structured, easy to follow, and effectively motivates the problem.
    • Comprehensive experiments: The inclusion of a retrieval task for Echo studies.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • Limited novelty: Based on the stated contributions, the primary novelty is in using multi-view echocardiography instead of single-view inputs. While the use of multi-view data and the architectural design are valuable, the overall novelty may be limited for a venue like MICCAI. Limited evaluation metrics: The evaluation mainly reports classification performance using AUROC.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
    • Adding references to support the statement: “. . . it cannot directly quantify cardiac morphological phenotypes like left ventricular ejection fraction (LVEF). Traditionally, these structural measurements require echocardiography (Echo), a modality often limited by cost, medical expertise, and accessibility. ” There are several studies that estimate LVEF from ECG. The authors should clarify the limitations of those approaches. Why is OTiS selected as the ECG encoder? While it is a time-series foundation model, there are strong ECG-specific foundation models such as ECG-FM and ECG-Founder. If there is evidence supporting OTiS as a better choice, please provide references.
    • Adding a reference for the statement: “In clinical practice, cardiologists do not base diagnoses on a single view; instead, they integrate information across multiple views from a complete Echo study to obtain a comprehensive cardiac assessment. ”
    • It would be helpful to clarify the rationale for selecting KNN as the classifier.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Limited novelty is the main concern about the paper, but its strengths are enough not to be rejected.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper presents Echo2ECG, a multimodal self-supervised framework that aligns ECG signals with multi-view echocardiographic representations, addressing the representational mismatch where single-view Echo captures only local anatomy while ECG reflects global electrical activity. Using CLIP-style contrastive learning with frozen pretrained encoders, the resulting ECG encoder is evaluated on LVEF classification across three datasets, structural heart disease classification under varying annotation scales, and morphological ECG-to-Echo retrieval. Results demonstrate superior performance over unimodal baselines and existing multimodal methods, particularly in low-data scenarios.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.This work addresses a clinically significant limitation of ECG by leveraging echocardiography as a morphological supervision signal, establishing clear relevance for cardiac imaging. 2.The key insight that ECG should align with multi-view Echo representations rather than single views represents a sound and intuitively compelling design. 3.The experiment evaluation spans multiple downstream tasks and datasets, including classification and cross-modal retrieval, providing a comprehensive assessment of representation quality. 4.The compact model size achieves superior results with fewer parameters than key multimodal baselines, enhancing practical utility.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    The framework stacks standard components (pretrained encoders, view aggregation, contrastive learning), with multi-view substitution for single-view alignment constituting incremental rather than innovative contribution. Certain claims appear overstated: assertions regarding the distillation of a complete picture of the heart’s morphological structure or generalized robustness lack sufficient evidentiary support. The text report-independent claim is ambiguous because the Echo encoder initialization relies on report-supervised pretraining, undermining strict differentiation from text-assisted methods. Reproducibility details are insufficient regarding training configurations, aggregation module behavior, and data processing. The retrieval evaluation appears heuristic: the ±0.5 standard deviation matching rule lacks theoretical or empirical justification, rendering results less convincing than classification outcomes. The source of performance gains requires clarification, whether they come from multi-view informational completeness, increased paired supervision, or attention aggregation itself. The Echo aggregation requires elaboration on typical view counts, missing view handling, and permutation invariance. Pretrained components transparency should be discussed. EchoPrime initialization should be described clearly to avoid misleading impressions of complete text independence.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper proposes a multimodal self-supervised learning framework to align ECG signals with multi-view echocardiography, enriching ECG representations with cardiac morphological information. It fits the topic of the venue and has some merits for acceptance.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposes a multimodal self-supervised learning framework, Echo2ECG, which aims to enhance ECG representations using structural and functional information from multi-view echocardiography (Echo). The method aligns ECG embeddings with study-level embeddings derived from multiple Echo views, enabling cross-modal self-supervised learning without relying on text labels or manual annotations. The framework is evaluated on three datasets for structure-related cardiac phenotype tasks.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The problem is clinically meaningful. The paper addresses the important question of whether richer morphological information from Echo can be used to improve ECG representations. This direction is well motivated from a medical perspective and has potential value for low-cost screening of structural heart disease. 2.The method has a certain degree of novelty. Compared with prior single-view alignment strategies, the paper emphasizes aligning ECG with a multi-view Echo study, thereby aggregating more complete cardiac structural information. This setting is more consistent with real echocardiographic examinations and constitutes the main methodological contribution of the paper. 3.The proposed approach is fully self-supervised and does not require text labels or additional manual annotations, which improves its scalability and general applicability. 4.The evaluation covers multiple datasets. The experiments are conducted on three datasets, which provides a broader empirical basis for assessing the method.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.Some experimental details remain insufficiently described. For example, the design of the comparison setting Echo2ECG (single-view Echo) is not clearly explained. The current manuscript does not sufficiently specify how the single-view variant is constructed, including the view selection strategy, input form, and its exact difference from the multi-view setting. This weakens the interpretability of one of the paper’s core comparisons. 2.Some results do not fully support a clear generalization advantage. Based on the reported results, the proposed model does not consistently outperform competing methods across all settings. The paper would benefit from a more careful discussion of the method’s generalization ability and its limitations. 3.Although the paper argues that multi-view Echo provides advantages over single-view alignment, the current version lacks more direct visualization or representation analysis to explain what is learned through the multi-view design and why it is beneficial. 4.The reported evaluation metrics are somewhat limited. For example, for the LVEF three-class classification task, additional metrics such as accuracy, macro-F1, or confusion matrices would provide a more complete picture of performance across classes. 5.The study uses private institutional data, but the current manuscript provides limited details regarding ethical approval, data access, and compliance procedures. 6.Releasing pretrained weights would substantially increase the reproducibility and impact of this work.

  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    The study uses private institutional data, but the current manuscript provides limited details regarding ethical approval, data access, and compliance procedures.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Overall, the paper studies a clinically relevant problem and proposes a multimodal self-supervised framework with a reasonably clear methodological motivation. Its main strength lies in the idea of enhancing ECG representations through multi-view Echo alignment, which is intuitively meaningful and potentially useful. With stronger clarification and more comprehensive analysis, the work could become substantially more convincing.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Author Feedback

We thank all reviewers for their thoughtful feedback and for recognizing the key strengths of our work: (1) the clinical relevance of predicting Echo-derived labels from the low-cost and easily accessible ECG, (2) the novelty of aligning ECGs with multi-view Echos, and (3) the comprehensive evaluation on multiple datasets and multiple downstream tasks.

We now address the raised concerns:

R1:

  • OTiS as ECG encoder: We selected OTiS primarily for its computational efficiency: at 7.1M parameters, it is much more lightweight than ECG-specific alternatives such as ECGFounder (30.7M), xECG (57M), and ECG-FM (90.9M). Despite being a general time-series model, OTiS performs on par with these domain-specific alternatives on our downstream tasks (Table 3 & Figure 2).
  • kNN classifier: We use kNN because it evaluates frozen embeddings directly without any parameter updates, providing a parameter-free measure of representation quality. This ensures that performance differences reflect embedding structure rather than classifier capacity.

R2:

  • Implementation details: Due to space constraints, training configurations are provided in our GitHub repository. Data processing is described in Section 4.1, and we refer the reader to the original OTiS and EchoPrime papers for further details.
  • Echo view aggregation: The typical number of views per study can be derived from Table 1.To accommodate the variable number of views across studies, we pad studies to a uniform size (to the configured max. number of views per study=128) within each batch using zero-padding, and apply an attention mask during view aggregation to ensure padded views do not contribute to the resulting study embedding. Furthermore, our aggregation is permutation-invariant: since views within a study have no meaningful sequential order, the aggregator treats them as an unordered set, making the resulting embedding independent of the order in which views are presented.
  • Retrieval threshold (±0.5 standard deviation): This threshold balances two considerations: (1) a single ECG is compatible with a range of structural phenotype values — which is precisely why we frame LVEF prediction as classification rather than regression — making exact matching overly strict and clinically unrealistic; and (2) thresholds beyond ±0.5 standard deviation produce too many chance-level correct matches, rendering Precision@1 uninformative.

R3:

  • Single-view vs. multi-view setting: Both settings use identical train/test data. In the single-view setting, each Echo view is treated as an independent sample, forming separate ECG↔view pairs (instead of ECG↔study pairs) without any aggregation module. No view selection heuristic is applied. At test time, in the single-view setting, the study-level embedding for retrieval is obtained by averaging individual view embeddings (as noted in Section 4.3). The key architectural difference from the multi-view setting is the absence of the view aggregator.
  • Ethics and data compliance: Ethical approval, data access, and compliance details for our private institutional dataset will be included in the final manuscript.
  • Pre-trained weights: We will release pre-trained weights in our GitHub repository.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    The paper addresses a clinically relevant problem, aiming to improve ECG representations using multi-view echocardiography. The motivation is clear and aligned with clinical practice.

    Reviewers consistently find the approach effective, with results across multiple datasets and tasks supporting the method.

    There are some limitations. The novelty is somewhat incremental, and some claims and experimental details could be better clarified. One reviewer also noted that details on ethical approval and data usage should be more clearly stated.

    Overall, the strengths outweigh the limitations and the paper meets the acceptance bar.



back to top