Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Electrocardiogram (ECG) signals are critical physiological indicators for assessing human health status and diagnosing cardiovascular diseases. However, substantial inter-subject feature variability poses a major challenge to cross-subject generalization in ECG analysis. Existing methods primarily rely on domain adaptation and subject-invariant feature learning, which are often task-specific and may introduce potential data leakage. To address these issues, we propose Vision Enhanced Generative Augmentation (VEGA), a novel paradigm that enables models to quickly adapt to different subjects through data generation without introducing any data leakage. Specifically, VEGA leverages rich semantic priors from a generative model pre-trained on natural images to guide the synthesis of realistic and diverse ECG signals. To adapt to the visual generator, one-dimensional (1D) ECG signals are first converted into two-dimensional (2D) visual representation. Then, to align the numerical distributions between the visualized ECG signals and natural images, VEGA adopts a lightweight domain adaptation strategy by fine-tuning only the normalization layers of the generator with a small amount of external ECG data. Through augmented ECG data generation, VEGA effectively bridges the feature distribution gap between training and testing subjects. We evaluate VEGA on multiple ECG tasks, including fatigue detection, ventricular premature beat (VPB) detection, and generative heartbeat prediction. Across all tasks and datasets, VEGA consistently improves cross-subject generalization performance, demonstrating its effectiveness as a unified generative augmentation paradigm for ECG analysis. Our code is available at https://github.com/Yuyun1011/VEGA-release.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1818_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Yuyun1011/VEGA-release

Link to the Dataset(s)

MIT-BIH dataset: https://physionet.org/content/mitdb/1.0.0/ BUP-QDB dataset: https://physionet.org/content/butqdb/1.0.0/

BibTex

@InProceedings{ShiTia_VEGA_MICCAI2026,
        author = { Shi, Tianyi AND Zheng, Siyang AND Meng, Zhu AND Cui, Zhe AND Huang, Jin AND Ren, Changrui AND Zhao, Zhicheng},
        title = { { VEGA: Vision Enhanced Generative Augmentation for Cross-Subject Electrocardiogram Generalization } },
        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 core contribution of this paper is the proposal of VEGA (Vision Enhanced Generative Augmentation), a novel cross-modal data augmentation framework specifically designed to tackle the challenge of cross-subject generalization in Electrocardiogram (ECG) analysis. Due to the significant morphological variability of ECG signals across different subjects, deep learning models often overfit to subject-specific identity cues rather than task-relevant features. To address this, VEGA innovatively transforms 1D ECG signals into 2D visual representations and leverages the rich structural priors from a natural-image pre-trained Masked Autoencoder (MAE) to synthesize diverse and realistic ECG samples. To align the natural image modality with the ECG modality, the authors introduce a parameter-efficient fine-tuning strategy coupled with a novel Frequency Adapter. By augmenting the training set with these generated samples, VEGA acts as a module that explicitly bridges the salient feature distribution gap between training and testing subjects, thereby significantly boosting the robustness of downstream tasks such as fatigue and VPB detection.

  • 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 paper directly tackles a well-known challenge in ECG analysis: severe inter-subject variability that leads to poor generalization on testing patients. Instead of relying on traditional fine-tuning techniques, which may risk data leakage, VEGA solves the distribution gap from a data generation perspective. By enriching the training manifold with diverse, generated cross-subject samples, it forces the downstream networks to focus on task-relevant discriminative features rather than memorizing patient-specific backgrounds. The latent space visualization (Fig. 3) provides convincing evidence that VEGA effectively aligns the training and testing distributions.

    2.The methodological highlight of this work is the highly creative cross-modal adaptation. Training high-fidelity generative models (such as Diffusion) from scratch typically requires massive amounts of ECG data, which is clinically impractical. VEGA elegantly bypasses this by exploiting the spatial/structural similarities between natural images and the periodic waveforms of ECGs. The strategy of dividing 1D ECGs into segments and stacking them column-wise into a 2D matrix successfully preserves both intra-cycle morphology and inter-cycle dynamics.

  • 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.While the authors position VEGA primarily as a generative augmentation framework, the empirical validation lacks comprehensive comparisons with existing time-series data augmentation techniques. In the generative heartbeat prediction task (Table 2), VEGA is only compared against NF and DDPM For the core clinical downstream tasks, the experiments only compare the baseline models against themselves with VEGA augmentation (i. e. , Model vs. Model + VEGA). Without more baselines, it is difficult to determine whether the performance gains are uniquely attributed to VEGA or merely the general benefit of having more training data.

    2.The assumption of ‘label preservation’ during the generative augmentation process. The authors utilize a MAE that masks a certain proportion of the input and reconstructs it, claiming that ‘the original task labels are preserved for the generated outputs. However, this claim lacks validation. In cardiovascular diagnostics, diseases like VPB are often characterized by highly localized, subtle morphological anomalies. If the random masking mechanism happens to obscure these critical pathological regions, the MAE might utilize its learned priors to reconstruct a morphologically ‘normal’ waveform. Assigning the original pathological label to this smoothed or normalized reconstruction would inadvertently introduce noisy labels into the training set, potentially confusing the downstream classifier.

  • 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 does not provide sufficient information for reproducibility.

  • 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?

    The major factors driving my recommendation are the highly innovative cross-modal methodology and its strong practical implications for ECG analysis. While I have noted a few areas for improvement in my detailed comments, the need for 1) broader comparisons with traditional data augmentation baselines, and 2) a verification regarding how localized pathological features are needed.

  • Reviewer confidence

    Very confident (4)

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

    Accept

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

    i’m good with the rebuttal, partially address my concerns.



Review #2

  • Please describe the contribution of the paper

    VEGA is a generative augmentation method for cross-subject ECG generalization. It proposes parameter-efficient fine-tuning, updating only layer normalization parameters to adapt the 2D model to ECG morphology with minimal data. The generated ECG samples augment the training distribution, reducing the feature gap between training and test subjects.

  • 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.
    • Strong few-shot efficiency results. Table 2 shows that VEGA with only 10% training data outperforms DDPM and normalizing flows.
    • Task-agnostic design. VEGA is applied across multiple downstream architectures and distinct tasks without task-specific modification.
  • 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.
    • Missing baselines for the generative component and the adaptation component. The prior works (Chen et al. and Yoon et al.) are cited in related work but absent from comparison.
    • The inference-time sampling procedure is not clearly described. It is ambiguous whether generation proceeds by masking a contiguous future segment of an existing training sample and reconstructing it, or by generating from scratch?
    • The paper motivates 2D conversion, but providees no comparison against 1D-based generative baselines. Without this comparison.
    • The paper repeatedly claims that VEGA produces “diverse” augmented samples, but provides no quantitative evidence of diversity.
  • 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 does not provide sufficient information for reproducibility.

  • 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?

    All datasets used are established public benchmarks with no privacy issues raised.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    While the paper presents a clever parameter-efficient adaptation module (Frequency Adapter) and demonstrates solid empirical improvements in few-shot scenarios, the core methodological contribution—converting 1D ECGs to 2D to exploit natural image priors—is insufficiently validated. The complete absence of 1D generative baselines and missing comparisons against directly related works (Chen et al., Yoon et al.) leaves a major gap in proving the necessity of the proposed pipeline. Furthermore, the lack of clarity regarding the sampling procedure and unsubstantiated claims of generation diversity push the paper below the acceptance threshold in its current state. A strong rebuttal addressing the baseline comparisons is required.

  • Reviewer confidence

    Very confident (4)

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

    Reject

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

    The rebuttal clarifies the sampling process, but my main concern remains. The key claim is that VEGA improves downstream cross-subject generalization, but the paper does not compare against baselines on the downstream tasks. I also remain unconvinced by the diversity evidence, since figure 3 appears to use more reference/training samples in the right panel, making the visual comparison difficult to interpret. Therefore, I keep my score unchanged.



Review #3

  • Please describe the contribution of the paper

    This paper proposes AURORA to address the limitation of existing learning-based approaches for reconstructing standard 12-lead ECG from a single-lead measurement, which often fail to adequately reflect the intrinsic physiological principles of ECG signals. The proposed method first reconstructs one additional limb lead using a learning-based model, and then computes the remaining limb leads based on the Einthoven and Goldberger equations. Subsequently, the full limb leads are used as conditions to reconstruct the precordial leads. Experimental results on the PTB-XL and CUSPH datasets demonstrate improvements in RMSE, DTW, and PCC metrics. The paper also reports both reconstruction performance and downstream disease prediction results. However, the proposed approach appears to be a hybrid extension of existing lead reconstruction frameworks with the incorporation of ECG-related equations, resulting in limited methodological novelty. Furthermore, the potential issue of error propagation in the staged reconstruction pipeline is not sufficiently analyzed. Overall, while the study direction is meaningful, several aspects require further refinement at the current stage.

  • 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.

    The paper formulates the ECG reconstruction problem not as a purely end-to-end generation task, but as a structured problem leveraging ECG vector relationships. It introduces a hybrid strategy in which only one limb lead is reconstructed via learning, while the remaining limb leads are derived analytically using physiological equations. The proposed IGUANA framework combines a lead representation learner with an AdaIN-based 1D U-Net generator and incorporates both source reconstruction and representation consistency in the training process. Comprehensive experiments are conducted on both PTB-XL and CUSPH datasets, including comparisons with competing methods, claim validation, and loss ablation studies. The reconstructed 12-lead ECG achieves AUROC and AUPRC performance close to that of the original ECG in downstream disease prediction tasks.

  • 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 method largely resembles an integration of existing lead reconstruction architectures with ECG-based analytical relationships, and thus does not strongly highlight methodological novelty. The experimental setting fixes the input lead to Lead I and the auxiliary lead to Lead II, but lacks sufficient validation of generalization across different lead configurations. The rationale for adopting the current 6-to-6 conditioning scheme in precordial lead reconstruction is not sufficiently compared with alternative design choices. Since the proposed approach follows a staged reconstruction pipeline, further analysis is required to examine the potential accumulation of reconstruction errors across stages. Some competing methods rely on the authors’ own implementations, leaving potential bias due to differences in reproduction settings.

  • 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 does not provide sufficient information for reproducibility.

  • 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?

    I selected Weak Accept because the paper presents a technically coherent framework and reports consistent improvements across multiple evaluation settings, which supports the practical relevance of the proposed approach. In particular, the combination of learning-based lead reconstruction with ECG-related analytical relationships is well motivated, and the empirical results on PTB-XL and CUSPH suggest that the method is effective in both reconstruction quality and downstream utility. At the same time, the methodological novelty is somewhat limited, as the overall design remains close to a hybrid extension of existing reconstruction pipelines, and several design choices would benefit from stronger validation. In addition, the staged reconstruction process raises questions about possible error propagation, and the current experiments do not fully resolve these concerns. Overall, I consider the paper to be slightly above the acceptance threshold because it offers a meaningful contribution with reasonable experimental support, although it still leaves room for further clarification and refinement.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    Accept

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

    After reading the authors’ rebuttal, I maintain my Accept recommendation.

    The rebuttal provides clarifications on the role of visual priors, the necessity of the Frequency Adapter, the inference-time sampling procedure, label preservation, and generation diversity. The authors also explain that the generation process is based on single-step parallel decoding rather than long-horizon autoregressive generation. This explanation partially alleviates my concern about error accumulation.

    The experimental results show performance improvements and reduced variance after applying VEGA in fatigue detection, VPB detection, and generative heartbeat prediction. These results support the view that the proposed method provides a practical contribution to cross-subject ECG generalization.

    However, some concerns remain. The connection between natural image structures and ECG waveforms is still not sufficiently justified. The comparison with simple 1D augmentation baselines is also limited, as it does not include methods such as Gaussian noise injection, temporal shifting, or scaling. The use of a Vision Transformer backbone requires further discussion in terms of computational cost. More analysis of the external ECG adaptation process and its effect would also strengthen the paper.

    Overall, I believe that the rebuttal addresses some of my concerns, but it does not fully resolve all remaining issues. Nevertheless, given the observed performance improvements across multiple tasks, I consider the paper to be slightly above the acceptance threshold and therefore recommend Accept.



Author Feedback

We thank AC and all reviewers for their efforts and constructive feedback. All your constructive comments are considered. 1.The novelty, integration concern and error accumulation (MR[3]) (1) Novelty of the Paradigm: VEGA is the first systematic framework that bridges visual priors and 1D ECG generation to enable cross-subject augmentation without data leakage. Its tailored frequency adapter freezes self-attention to preserve spatial priors learned from images, while explicitly injecting global spectral information to model ECG’s periodic nature. The design effectively disentangles spatial and spectral features, and its superiority is quantitatively validated in Table 2. (2) Integration Innovation: All model components are elaborately designed rather than arbitrarily assembled, forming a minimal yet indispensable architecture. Specifically, the MAE captures reconstructive visual priors, instance normalization enables amplitude bridging, frozen self-attention preserves pre-trained prior knowledge under limited data, and the Frequency Adapter achieves spectral alignment. Our contribution lies in the synergistic combination of these modules, instead of their isolated individual functions. (3) Error Accumulation Issue: Unlike conventional autoregression, VEGA adopts single-step parallel decoding for future ECG segments, thereby inherently avoiding long-horizon error accumulation—a deliberate architectural design to enhance model robustness. 2.Label preservation and inference-time sampling procedure (MR[2]/R1[2]/R2[2]) During inference, VEGA converts the ECG into a 2D representation and maps it to the left image half via bilinear interpolation without masking, so no information is lost. The right half is fully masked and reconstructed by the MAE. For augmentation, the alignment loss targets the original sequence; for prediction, it targets the future sequence. The substantial Recall and F1-score improvements after VEGA augmentation (Fig. 4) confirm that sensitivity to positive samples is preserved without distorting pathological waveforms. Predictive generation visualizations (Fig. 5) further verify that VEGA faithfully reconstructs key waveform morphologies under varying conditions. 3.Diversity issue (R2[4]) Regarding generation diversity, we provide both mechanistic explanations and visual evidence. (1) Source of diversity: Diversity is inherent in our design. First, bilinear interpolation during 2D resizing introduces sub-patch shifts (hyper-parameter η). Second, the MAE’s stochastic masking naturally produces varying outputs from the same input, since different visible patch mosaics yield distinct completions. (2) Latent space evidence (Fig. 3): Latent feature visualization shows that VEGA substantially broadens the training distribution toward the test set, directly evidencing the generation of diverse, subject-variant features that bridge the generalization gap. 4.Lack of comprehensive baselines (MR[1]/R1[1]/R2[3]/R2[1]) As clarified in Q1, VEGA is the first framework bridging visual priors and 1D ECG generation for cross-subject augmentation. It adopts a cross-modal paradigm of visual pre-training and fine-tuning on a small external ECG dataset. In contrast, diffusion models and normalizing flows typically require large amounts of data; fine-tuned on the same limited data as VEGA, they would suffer mode collapse, produce nearly pure noise, and cause a sharp performance drop. To demonstrate VEGA’s few-shot generative capability, we employ conditional generation (i.e., predictive generation), which effectively alleviates the mode collapse problem prevalent in diffusion and other mainstream methods. As Table 2 shows, VEGA achieves better generative performance with far fewer training samples than other 1D generative methods. In the revised version, we will rigorously add more baseline comparisons under fair experimental settings to validate the advantages of the proposed method.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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.

    While reviewers agreed that the paper addresses important challenges in ECG analysis and features innovative design elements, they also raised serious concerns about the paper’s evaluation rigor, methodological clarity, and novelty: First, the empirical validation lacks comprehensive baselines, and relevant prior works are absent from comparisons, making it hard to attribute performance gains to the proposed methods. Second, key claims and procedures lack validation or clarity. Third, methodological novelty is not strongly highlighted, with one method resembling an integration of existing architectures and insufficient justification for design choices. Since the reviews are mixed, we would like to invite the authors to provide a rebuttal. In addition to clarify other concerns, please try to address the following points during rebuttal: include comprehensive comparisons with existing time-series augmentation techniques, 1D-based generative baselines, and relevant prior works; validate the “label preservation” claim and clarify the inference-time sampling procedure; highlight methodological novelty, justify key design choices, and address potential reconstruction error accumulation.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    This is a borderline paper with mixed reviews. While reviewers agreed that VEGA effectively addresses a critical challenge in ECG analysis—poor cross-subject generalization—by creatively adapting 2D generative models to enrich the training manifold and align distributions, they also raised serious concerns about the paper’s soundness and significance: first, the empirical validation lacks comprehensive baselines, particularly comparisons with existing time-series augmentation techniques and alternative generative approaches, making it difficult to isolate VEGA’s unique contribution versus generic data expansion benefits; second, the assumption of label preservation during masked reconstruction is insufficiently validated, raising the risk of introducing noisy labels when critical pathological features are obscured; third, the evidence for generated sample diversity is purely qualitative and unsubstantiated, while the experimental design omits key comparisons, such as 1D generative baselines and alternative lead configurations. The rebuttal clarified certain procedural details but did not fully resolve concerns regarding novelty, baseline comparisons, diversity quantification, or label preservation validation. After careful deliberation, I lean a bit towards rejecting this manuscript.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal clarifies several key aspects of the method, particularly the sampling procedure, the role of the Frequency Adapter, and the rationale behind the cross-modal design.

    However, some limitations remain only partially resolved. In particular, the lack of direct comparisons with standard time-series augmentation techniques and 1D generative baselines in downstream tasks is still an important gap, and the evidence supporting generation diversity and label preservation remains somewhat indirect.

    Given the novelty of the cross-modal augmentation strategy and the consistent empirical improvements across tasks, the rebuttal sufficiently strengthens the paper.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    The rebuttal clarifies several methodological details and partially addresses concerns regarding model design, label preservation, and baseline comparisons. I appreciate the interesting idea of leveraging visual generative priors for ECG augmentation, and the reported empirical improvements suggest potential value. However, I still have some reservations about the core motivation: ECG is inherently a 1D periodic physiological signal with strong cycle- and morphology-specific constraints, and it remains insufficiently justified why a vision generative model should be more suitable than signal-specific generative models for this task. Overall, while this concern remains, I am willing to support acceptance given the novelty of the cross-modal formulation and the promising experimental results.



back to top