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

Abstract

Wearable and portable devices enable convenient electrocardiography (ECG) acquisition, yet they typically measure only a single limb lead, limiting diagnostic utility compared to the standard 12-lead ECG used in hospitals. We study the problem of generating the full standard 12-lead ECG from a single measured lead by leveraging inherent ECG principles. Specifically, we emphasize two claims: (i) limb lead reconstruction benefits from a hybrid strategy that generates only one additional limb lead via learning and computes the remaining limb leads through vector operations; and (ii) precordial lead reconstruction is more reliable when conditioned on the full set of limb leads. Based on these claims, we propose \textbf{AURORA}, a 12-lead ECG reconstruction framework that combines learning models and deterministic vector operations to generate limb leads, followed by the reconstruction of precordial leads. To instantiate the learning components, we propose \textbf{IGUANA}, a lead-pattern learning model comprising a lead representation learner and a lead generator for precise lead-to-lead transformation. Experiments on two benchmark datasets demonstrate that AURORA consistently outperforms state-of-the-art methods and validate the effectiveness of incorporating vector operations in lead reconstruction. Our code is available at {https://anonymous.4open.science/r/ECG-553F.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4622_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/DHSeo11/AURORA.git

Link to the Dataset(s)

PTB-XL dataset: https://physionet.org/content/ptb-xl/1.0.3/ cusph dataset: https://physionet.org/content/ecg-arrhythmia/1.0.0/

BibTex

@InProceedings{SeoDon_Accurate_MICCAI2026,
        author = { Seo, Dong-hyuk AND Kim, Ui Jong AND Jo, Yong-Yeon AND Kwon, Junmyung AND Shin, Won-Yong AND Kim, Sang-Wook},
        title = { { Accurate Reconstruction of the Standard 12-Lead ECG from a Single Lead Based on Inherent Principles of ECG } },
        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 paper addresses the problem of reconstructing a standard 12-lead ECG from a single measured lead (Lead I), motivated by the widespread use of single-lead wearable ECG devices and the diagnostic limitations of single-lead recordings. The authors propose AURORA, a three-stage reconstruction framework that combines learning-based generation with deterministic vector operations derived from established ECG principles (Einthoven’s triangle and Goldberger’s equations).

    The authors articulate two claims as the foundation of their approach: (C1) limb lead reconstruction benefits from a hybrid strategy that generates only one additional limb lead via learning and computes the remaining limb leads through vector operations, and (C2) precordial lead reconstruction is more reliable when conditioned on the full set of limb leads rather than a single lead. They provide empirical evidence for both claims and evaluate the framework on two standard benchmark datasets (PTB-XL and CUSPH), comparing against six existing methods across three reconstruction metrics (RMSE, DTW, PCC) and assessing clinical utility through downstream disease prediction (myocardial infarction on PTB-XL, arrhythmia on CUSPH).

  • 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 central design choice of incorporating Einthoven’s and Goldberger’s relationships as explicit deterministic operations rather than leaving them to be implicitly learned is a reasonable and well-motivated departure from the prevailing end-to-end learning paradigm. Constraining the solution space with known physical relationships reduces the learning burden and enforces internal consistency among the reconstructed limb leads, and the ablation comparing AURORA with its pure learning-based variant (AURORA_{1to11}) supports this design choice.

    • The three-stage decomposition (1-to-1 generation, vector operations, 6-to-6 generation) provides a clear separation of concerns. Each stage has a well-defined role with explicit interfaces between them, which is a notable improvement over monolithic end-to-end approaches whose failure modes are often difficult to diagnose.

    • The reported results show consistent gains over competitive baselines on two independent datasets with different target diseases. On PTB-XL, AURORA reduces RMSE by 13.0% and DTW by 26.3% compared to the strongest baseline (ANN128); on CUSPH, the corresponding improvements are 14.0% and 12.6%. The consistency across different metrics, datasets, and disease contexts suggests that the observed improvements are not artifacts of a particular evaluation setting.

  • 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 paper frames its goal as “accurate reconstruction of the standard 12-lead ECG from a single measured lead,” but the information-theoretic content of what is actually produced warrants closer examination. In AURORA, Lead II is generated from Lead I through M_{1-to-1}, and the remaining four limb leads (III, aVR, aVL, aVF) are computed as linear combinations of Lead I and the generated Lead II. Consequently, all six limb leads provided to M_{6-to-6} are ultimately functions of Lead I alone. The four vector-derived leads, being linear combinations of two signals, contribute no additional independent information by construction; they are formally redundant with {L_I, L_II^gen}. The six precordial leads are then generated from this set, meaning that all eleven reconstructed leads trace back to a single measured signal.

    • The IGUANA training procedure includes a source reconstruction phase (Eq. 6) with its corresponding loss ℓ^s_rec (Eq. 10). The ablation in Table 2(b) shows that removing this loss degrades RMSE from 0.127 to 0.144, but the mechanism underlying this improvement is not established. The authors’ stated rationale - that it “encourages faithful transformation” - leaves several alternative explanations unexamined: the loss may simply act as a regularizer achievable through simpler means; it may alleviate under-training of G by doubling gradient signal per step; or, most concerning, it may incentivize an unintended failure mode in which L_I information is encoded into perceptually inconspicuous components of L_II^gen, a form of steganography documented in cycle-consistent translation models (Chu et al. , 2017). Under this last interpretation, the cycle loss would inflate apparent reconstruction fidelity without improving physiological validity.
    • There is also a tension with the representation consistency loss ℓ^s_con (Eq. 9), which already enforces alignment between latent codes of real and generated signals. It is unclear what additional constraint source reconstruction provides beyond invertibility of G on these aligned representations. Distinguishing among these explanations would require targeted analyses - such as perturbation sensitivity tests on L_II^gen or matched-complexity regularization baselines-that the paper does not provide.
    • Key training details including the loss weights λ_gen, λ_rec, λ_con, learning rate, batch size, and number of epochs are not reported in the paper, which limits reproducibility and makes it difficult to assess whether baselines were tuned under comparable protocols.
  • 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

    Could the authors clarify what the name IGUANA stands for, or whether it is intended as a thematic pairing with AURORA rather than an acronym?

  • 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 paper addresses a practically relevant problem with a reasonable solution. Combining learned lead-to-lead generation with deterministic vector operations grounded in Einthoven’s and Goldberger’s relationships is well-motivated, the three-stage pipeline is clearly structured, and the empirical results show consistent improvements over multiple baselines across two datasets. The downstream diagnostic evaluation provides useful evidence of practical utility beyond signal-level metrics. Several concerns temper my enthusiasm. The framing of “physiology-grounded claims” overstates contributions that are, in substance, applications of well-established linear algebraic relationships and general machine learning principles. The term “reconstruction” obscures the fact that all eleven generated leads are ultimately derived from a single input signal, making the output closer to a structured estimate under a learned prior. The mechanism underlying the source reconstruction objective is not adequately analyzed, and the downstream evaluation lacks a single-lead baseline needed to attribute the observed performance to the reconstruction rather than to the diagnostic sufficiency of Lead I itself. Reproducibility is also a concern: key training details (loss weights, learning rate, batch size, epochs) are not reported, and the description of data splitting and preprocessing is minimal. The anonymized code repository partially mitigates this, but the paper alone does not provide sufficient information to reproduce the results or verify that baselines were tuned under comparable protocols. On balance, the work makes a useful contribution, and the concerns above are addressable through revision rather than indicative of fundamental flaws. I recommend weak accept.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [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

    The manuscript proposes AURORA, a hybrid framework for reconstructing the standard 12-lead ECG from a single measured lead by explicitly incorporating inherent electrophysiological principles into the reconstruction process. Instead of relying purely on end-to-end learning-based translation, the method combines minimal learning with deterministic vector operations grounded in classical ECG theory. Its main contributions are: 1) Physiology-guided hybrid reconstruction strategy: The paper introduces a two-stage formulation in which only one additional limb lead is generated via a learning model, while the remaining limb leads are deterministically computed using well-established vector relationships (Einthoven’s and Goldberger’s equations). 2) Structured reconstruction pipeline: The method decomposes the problem into (i) 1-to-1 limb lead reconstruction, (ii) deterministic completion of limb leads, and (iii) conditional reconstruction of precordial leads from the full limb lead set, reflecting clinically meaningful dependencies between lead groups. 3) IGUANA model for lead-to-lead transformation: A unified learning architecture consisting of a lead representation learner and a generator is proposed, with a two-phase training objective that enforces both target reconstruction and source consistency. 4) Empirical validation across datasets and tasks: The framework is evaluated on PTB-XL and CUSPH datasets, demonstrating improved reconstruction performance over prior methods, and is further validated through downstream disease prediction tasks to assess clinical utility.

  • 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) Well-motivated integration of domain knowledge: The central idea of combining learning-based reconstruction with deterministic ECG vector operations is both intuitive and well grounded in electrophysiological principles. This reduces the learning burden while enforcing physiologically consistent relationships between limb leads, which is a meaningful design choice. 2) Clear and structured problem decomposition: The separation into limb lead reconstruction and precordial lead reconstruction is well justified and aligns with known geometric and physiological differences between lead types. This structured pipeline improves interpretability compared to fully end-to-end approaches. 3) Consistent empirical improvements across metrics and datasets: The method shows clear gains over multiple baselines on RMSE, DTW, and PCC across two benchmark datasets. The improvements are consistent rather than isolated to a single metric, strengthening the empirical claims. 4) Clinical relevance through downstream evaluation: The inclusion of disease prediction experiments using reconstructed ECG signals is a strong aspect of the paper. Demonstrating that diagnostic performance remains close to that of original 12-lead ECG signals provides evidence that the reconstruction preserves clinically meaningful information. 5) Thoughtful ablation studies: The paper includes ablations that directly validate the two core claims (hybrid strategy and staged reconstruction) as well as the contribution of different loss components, which helps support the design choices. 6) Good clarity and organization: The paper is well structured, with clearly stated claims (C1, C2), a coherent methodological pipeline, and logical progression from motivation to validation.

  • 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) Limited methodological novelty at the modeling level: While the hybrid formulation is well motivated, the underlying learning components (e.g., U-Net-based generator, representation learning, two-phase reconstruction) are largely based on existing architectures and training strategies. The main novelty lies more in the integration of known principles than in fundamentally new modeling techniques. 2) Incremental nature relative to prior reconstruction approaches: The task of reconstructing 12-lead ECG from reduced leads has been extensively studied, and many prior works already explore deep learning-based lead translation. The proposed method improves performance, but the conceptual leap beyond prior work is somewhat limited. 3) Fairness of baseline comparisons is not fully clear: Some baselines are re-implemented due to lack of public code. While this is understandable, the paper does not provide sufficient detail to ensure that these implementations are fully comparable to the original methods, which introduces some uncertainty in the reported gains. 4) Limited analysis of robustness and generalization: The experiments focus on standard benchmark datasets, but do not explore robustness to noise, device variability, or domain shifts, which are important considerations for real-world wearable ECG scenarios. 5) Design choices could be better justified: Certain decisions, such as the selection of specific limb leads (e.g., reconstructing Lead II first), are empirically motivated but not deeply analyzed. A more thorough justification or comparison with alternative configurations would strengthen the methodological argument. 6) Reproducibility details are partially incomplete: Although the authors provide a code link and indicate that code will be released, some implementation details (e.g., training hyperparameters, exact preprocessing steps, and architecture specifics) are not fully described in the manuscript, which may hinder exact replication.

  • 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

    The paper addresses an important and practically relevant problem in wearable ECG analysis, and the idea of incorporating physiological constraints into reconstruction is a promising direction. A few suggestions that could further strengthen the work: 1) Clarify the novelty more explicitly relative to prior ECG reconstruction methods, particularly distinguishing what is conceptually new versus what is an effective integration of existing components. 2) Provide more detailed discussion on the fairness of baseline comparisons, especially for methods that were re-implemented without official code. 3) Expand the analysis of robustness, for example by evaluating performance under noise, different sampling conditions, or cross-dataset generalization scenarios. 4) Include additional discussion or experiments to justify key design choices, such as the selection of specific limb leads and the ordering of reconstruction steps. 5) Improve reproducibility by providing more detailed implementation descriptions (e.g., architectural configurations, training schedules, and preprocessing steps).

  • 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 manuscript presents a well-motivated and practically relevant approach to reconstructing 12-lead ECG signals from a single lead by integrating physiological principles with learning-based methods. Its main strength lies in the hybrid formulation, which leverages deterministic vector operations to enforce consistency while reducing the reliance on purely data-driven learning. The empirical results are consistently positive across multiple datasets and evaluation metrics, and the inclusion of downstream disease prediction experiments provides convincing evidence of clinical utility. At the same time, the work has several limitations. The methodological novelty is more evident at the formulation level than at the modeling level, as the core learning components are largely based on existing architectures. The improvements over prior work, while consistent, can be viewed as incremental. In addition, the fairness of baseline comparisons and the robustness of the approach under real-world conditions are not fully established, and some implementation details are insufficiently described. Overall, I place the paper marginally above the acceptance threshold. The integration of domain knowledge is meaningful, the experimental validation is solid, and the clinical relevance is clear. However, the paper would benefit from clarification in rebuttal regarding novelty positioning, evaluation fairness, and reproducibility details.

  • 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 rebuttal, I remain marginally on the accept side.

    The authors addressed several of the major concerns raised in the initial reviews with clear and technically grounded responses. In particular, the rebuttal clarified that the novelty of the work lies primarily at the reconstruction-formulation level rather than in introducing a fundamentally new backbone architecture. The proposed staged reconstruction strategy—reconstructing only one additional limb lead through learning, deterministically deriving the remaining limb leads using established ECG vector relationships, and conditioning precordial lead reconstruction on the completed limb-lead set—is now more clearly positioned relative to prior ECG reconstruction approaches.

    The rebuttal also strengthened several methodological aspects through additional analyses. The authors provided empirical and physiological justification for selecting Lead II as the auxiliary reconstructed lead, including quantitative comparisons against alternative lead choices. Concerns regarding potential error accumulation in the staged reconstruction pipeline were addressed through separate V1–V6 evaluations, which still demonstrated improved performance relative to direct reconstruction variants. In addition, the response to the cycle-consistency and steganography-related concerns was thoughtful and technically relevant, including sanity checks using low-pass filtering and matched gradient-doubling baselines.

    The reproducibility and baseline fairness concerns were also partially alleviated by updating the anonymous repository with missing implementation details, hyperparameters, preprocessing information, and reproduced baseline configurations.

    Some limitations nevertheless remain. The modeling components themselves are still largely based on existing architectures, and the overall contribution is more incremental than fundamentally novel at the modeling level. Robustness under realistic deployment scenarios (e.g., noise, device variability, or domain shifts) also remains insufficiently explored. However, I believe the integration of physiological priors into the reconstruction formulation is meaningful, the empirical evaluation is consistently solid across datasets and metrics, and the downstream diagnostic validation supports the clinical relevance of the approach.

    Overall, I continue to consider the paper a borderline but acceptable contribution for MICCAI.



Review #3

  • Please describe the contribution of the paper

    This manuscript 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 manuscript 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 that leverages vector relationships inherent in ECG signals. It introduces a hybrid strategy in which only one limb lead is reconstructed through 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 into the training process. Comprehensive experiments are conducted on both PTB-XL and CUSPH, including comparisons with competing methods, claim validation, and loss ablation studies. The reconstructed 12-lead ECG also achieves AUROC and AUPRC values close to those 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 proposed method remains close to an integration of existing lead reconstruction frameworks with ECG-based analytical relationships, and therefore does not strongly highlight methodological novelty. The experimental setting fixes the input lead to Lead I and the auxiliary lead to Lead II, but provides limited 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. In addition, 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 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.

    (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?

    I selected Weak Reject because the paper addresses a meaningful problem and presents a technically coherent framework for reconstructing a standard 12-lead ECG from a single-lead measurement. The combination of learning-based lead reconstruction with ECG-related analytical relationships is well structured, and the reported results on PTB-XL and CUSPH suggest that the approach is effective in both reconstruction quality and downstream utility. However, the overall method remains close to a hybrid extension of existing reconstruction frameworks, which limits the degree of methodological novelty. In addition, several design choices, including the fixed lead configuration and the staged reconstruction process, would benefit from stronger validation and more detailed analysis, particularly with respect to generalization and possible error propagation. Overall, while the study direction is meaningful and the empirical results are encouraging, I find the paper to be slightly below the acceptance threshold in its current form.

  • 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 considering the authors’ rebuttal, I change my final recommendation from Reject to Accept.

    The rebuttal provides relatively concrete responses to the main concerns I raised in my review. The authors clarify that the novelty of AURORA lies not in replacing the backbone with a more complex architecture, but in the reconstruction formulation that incorporates ECG domain knowledge. They also explain that the Lead-I-based setting is motivated by wearable and portable ECG scenarios, and they provide additional evidence supporting the choice of Lead II as the auxiliary limb lead. Regarding possible error accumulation in the staged reconstruction pipeline, the authors report an evaluation focused on V1–V6 and explain that conditioning on the full limb-lead set contributes to precordial lead reconstruction. In addition, the concerns about hidden signals or gradient-doubling effects in the source reconstruction process are partially addressed through a low-pass filtering sanity check and a matched baseline.

    Some limitations still remain. The novelty of AURORA is closer to a formulation-level contribution that combines ECG vector operations with learning-based reconstruction, rather than a fundamentally new deep architecture. Generalization to different input-lead configurations is still outside the scope of the current submission, and the fairness of reproduced baselines cannot be fully verified. In addition, some of the analyses provided in the rebuttal should be clearly reflected in the final manuscript.

    Nevertheless, I find that the rebuttal alleviates a substantial part of my previous concerns. Considering the performance improvements reported in the manuscript and the additional clarifications provided in the rebuttal, I now consider the paper to be slightly above the acceptance threshold. Therefore, I recommend Accept.



Author Feedback

Novelty (R2-W1,2/R3-W1): The novelty of AURORA is at the reconstruction-formulation level, not in replacing the backbone with a more-complex architecture. Rather than reconstructing all missing leads once by learning-based translation, AURORA reconstructs only one additional limb lead, computes the remaining limb leads through deterministic vector operations, and reconstructs precordial leads conditioned on the full limb-lead set. This approach reduces the learning burden and constrains the output space to physiologically consistent ECGs, which is not achieved by simply applying a generic lead-translation network. We believe it is a meaningful methodological contribution to medical AI, aligned with recent MICCAI studies deriving values from domain knowledge or physiological/physical priors in reconstruction formulation [MRF-DiPh(MICCAI’25)/TA-fMRI-GM(MICCAI’25)/ACIS(MICCAI’24)]. We will clarify this positioning in the final version. 1-to-1 validation (R3-W2): Our target scenario is Lead-I-based reconstruction because many wearable/portable devices acquire a Lead-I-like signal [mECGNet(MICCAI’25)]. We agree that arbitrary input-lead generalization is broader, but it is not the focus of this submission. Within this setting, the auxiliary lead was not arbitrary: comparing candidate auxiliary limb leads, Lead II was most reliable, achieving 0.127/0.722 versus 0.203/0.519 in RMSE/PCC for the best alternative, Lead III. This supports the Lead-II choice as empirically justified and physiologically plausible, as Lead II often aligns with the dominant cardiac electrical axis.

Lead I information (R1-W1): We agree with R1 that AURORA does not create independent information beyond measured Lead I. Rather, we view it as diagnostic-view augmentation: Lead I may contain disease-relevant cues, but these can be difficult for downstream models to exploit from a single directional view. Since ECG leads reflect the same cardiac activity from different directions, AURORA presents such cues in the standard 12-lead form. To test whether the gain comes from Lead I itself, we added a Lead I-only baseline in this rebuttal; it achieved AUROC 0.87/0.78, lower than AURORA and prior reconstruction baselines.

Error accumulation (R3-W3): We agree that a staged pipeline could introduce error accumulation, particularly for precordial leads. To examine this directly, we evaluated V1–V6 only, beyond the 11-lead average in Table 2.On V1–V6, AURORA achieved 0.16/0.76 in RMSE/PCC, outperforming AURORA_{1to11} (0.18/0.62). This suggests that structured limb-lead conditioning benefits precordial reconstruction and is not dominated by possible error propagation.

Cycle concerns (R1-W2,3): We agree that source reconstruction may raise steganography or gradient-doubling concerns. However, unlike unpaired cycle-consistent translation, the generated Lead II is directly supervised by paired ground-truth Lead II. As a sanity check, 40Hz low-pass filtering generated Lead II before source reconstruction changed RMSE/PCC only marginally from 0.316/0.941 to 0.317/0.940, suggesting no strong reliance on fragile high-frequency hidden signals. A matched gradient-doubling baseline without source reconstruction achieved 0.328/0.884, still worse than AURORA. Thus, the gain is not simply explained by hidden channels or doubled gradients.

Reproducibility/baselines (R1-W4/R2-W3,6/R3-W4): We appreciate the concerns regarding AURORA reproducibility and baseline fairness. We updated the missing AURORA training/preprocessing details, including hyperparameters, in the anonymous repository. Since official implementations of several baselines were unavailable, we followed the original papers and updated the anonymous repository with reproduced baseline implementations and training configurations.

Naming (R1): IGUANA is an acronym of multI-lead ECGs Generation throUgh leAd patterN leArning. We apologize for the confusion and will revise it accordingly.



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.

    This is a interesting study which employs the intrinsic physiological principles of ECG signals to reconstruct 12-lead ECG. The method is simple but intuitive and reasonable. Some minor issue raised by the reviewers need to be solved further

  • 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.

    I am happy with the rebuttal, which solved my concerns.



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.

    Major concerns from the reviewers have been addressed. It is not clear whether the newly added baseline and “low-pass filtering and gradient-double checks” involved any possible violation of MICCAI’s rebuttal rules, and these updates were not in the manuscript. Based on the updated reviews and meta-reviews, I recommend a plain-poster presentation binding the following camera-ready conditions:

    “Analyses provided in the rebuttal should be clearly reflected in the final manuscript” – R3, these include: 1.Lead-I-only baseline, the V1 to V6 error-accumulation breakdown, and the low-pass and matched gradient-doubling sanity checks 2.Report the training hyperparameters, preprocessing, and baseline-reproduction settings now held only in the repository. 3.define the IGUANA acronym at first use (critical)



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.

    Rebuttal answered questions, remaining issues with R3 are non-fatal. “Cautiously optimistic” with respect to novelty, but certainly worth acceptance.



back to top