Abstract

With the rise of wearable IoT devices such as smartwatches and smart rings, ECG signals have become more accessible and made cardiovascular monitoring a reality. However, analyzing the ECG signals for complex conditions, such as bundle branch blocks and myocardial infarction, requires multi-lead ECG data. Although various deep learning models for ECG reconstruction have been proposed, they are computationally expensive and unsuitable on resource-constrained wearable IoT devices. To address this challenge, we propose mEcgNet, a parameter-efficient model for reconstructing 12-lead ECG signals from a single lead. mEcgNet introduces a modular deep learning architecture for parameter efficiency and separates the single lead-I signal into multiple frequency segments to improve accuracy. Our experiments demonstrate that mEcgNet significantly reduces the number of parameters and inference time by ∼23.1× and ∼5.4×, respectively, compared to existing state-of-the-art models. Furthermore, it reduces the reconstruction error by ∼22.1%, demonstrating its high accuracy and efficiency.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1826_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{LeeJun_ParameterEfficient_MICCAI2025,
        author = { Lee, Junseok and Yoo, Yeonho and Kim, Jinkyu and Lim, Dosun and Yang, Gyeongsik and Yoo, Chuck},
        title = { { Parameter-Efficient 12-Lead ECG Reconstruction from a Single Lead } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {430 -- 440}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposes mEcgNet, a generative model that generates a standard 12-lead electrocardiogram (ECG) from a single-lead ECG sourced from Lead I. Motivated the by hardware limitations of Internet of Things (IoT) devices, the authors focus on parameter efficiency.

  • 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.
    • Parameter Efficiency Motivation: The study is well motivated regarding model size, given the interest in wearable-based ECG measurements and concerns around computational and hardware resources in those settings.
    • Experiments demonstrate that the proposed method outperforms the baseline models.
  • 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.
    • Computational Motivation: The motivation for reconstruction to preserve privacy as outlined in the Introduction is not immediately clear. If a single lead ECG can be used to identify individuals, what guarantees are there that a reconstructed 12-lead ECG will not?
    • Clinical Motivation: The second paragraph introduction motivates the proposed mention by discussing several conditions that cannot be diagnosed by a single lead ECG. The authors explicitly mention myocardial infarction as an example; however, prior work such as Fatimah et al., 2021 (“Efficient detection of myocardial infarction from single lead ECG signal”) demonstrated that models using a single lead (in this case, lead II) can detect MI.
    • Proofreading: The authors should check their work for grammatical and spelling errors (e.g. Section 3.3 begins with a lowercase “w,” the “Params” column header in Table 3 does not mention the scale, etc.)
    • Metrics: The authors should justify their use of mean squared error and mean absolute error and exclusion of other signal processing metrics (e.g. peak signal to noise ratio) as well as domain-specific metrics such as timing and amplitude errors between peaks in the ECGs. Additionally, the authors should include distributional measures of performance rather than single point estimates.
  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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 — could be rejected, dependent on rebuttal

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

    While interesting and the results promising, the proposed method in the paper does not account for the full body of literature in single-lead ECG classification to justify the authors’ motivation for parameter efficient reconstruction.

  • 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 premise of the work is that single-lead ECG is now commonplace with wearable devices; however, 12-lead ECG is necessary for clinical diagnosis. As such, the investigators have developed a deep-learning reconstruction algorithm to synthesize a 12-lead ECG from a single-lead ECG signal. The processing pipeline splits the signal into frequency bands which are processed with the UNet-based “ECGModules”. The performance is compared with existing algorithms that accomplish the same task (EKGAN and ECG-Recover). Computational performance is better than the existing models; accuracy performance is similar. Automated disease classification is better than the other two models.

  • 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 strength of the paper is mostly in the technical implementation, which is more lightweight than existing algorithms. This is observable in the # of parameters used in the model, the GFLOPS, and the throughput performance.

  • 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 major weakness is that the premise of the work is that this approach can use ECG to complete a 12-lead clinical diagnosis. The analysis of disease prediction is (a) based on an algorithm which is not a clinical diagnostic tool (i.e. not approved for clinical diagnosis) rather than a cardiolgist’s review of the data, (b) ineffectively described in a brief 2 paragraph description, and (c) summerizes performance through simple averaging of the F1-scores for each class. The authors have the opportunity to complete a much more effective, deeper dive into the performance of the classification including an understanding of conditions where a synthesized signal performs well and where it does not.

  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/en/REVIEWER-GUIDELINES.html

    There is not statement about the source code. Some details of the mdoel are decsribed, but not all (like the selection of the frequency bands). The dataset are publically available.

  • 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 — could be accepted, dependent on rebuttal

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

    Given the premise of the work (i.e., the ability to generate a clinically useful 12-lead ECG signal from a single lead signal), more analysis and discussion should be focused on the clinical assessment. It would not be possible to complete a human evaluation study in the rebuttal, but an analysis that dives into the performance of the different classifications would be beneficial. Knowing the page limit, sections such as the ablation study could be compressed, given the limited findings of the ablation study (performance was nearly identical across conditions). From the perspective of “teaching the art” of building a lightweight but performant model, additional details could be provided to allow a reseracher to recreate the model, and an analysis of how others could gain efficiency in model development would be beneficial.

  • 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

    The paper “Parameter-Efficient 12-Lead ECG Reconstruction from a Single Lead” introduces mEcgNet, a novel deep learning model designed to reconstruct 12-lead ECG signals from a single lead-I input, addressing a critical need in cardiovascular monitoring using wearable IoT devices. As these devices often only capture lead-I signals, analyzing complex conditions like bundle branch blocks and myocardial infarction, which require full 12-lead data, becomes challenging. mEcgNet utilizes a modular architecture to enhance parameter efficiency and accuracy by dividing the lead-I signal into multiple frequency segments. This frequency-based segmentation allows the model to focus on key characteristics of the ECG signal, thereby improving the overall reconstruction quality. The experimental results show that mEcgNet significantly reduces the number of parameters and inference time by approximately 23.1× and 5.4×, respectively, compared to existing state-of-the-art models. Moreover, it achieves around a 22.1% reduction in reconstruction errors, demonstrating its effectiveness and efficiency for real-time ECG monitoring on resource-constrained devices. In conclusion, mEcgNet stands out for its lightweight design, making it suitable for deployment on wearable devices while ensuring accurate and timely ECG reconstruction, with potential future work exploring federated learning for enhanced privacy and efficiency in ECG data processing.

  • 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 introduction of mEcgNet represents a significant advancement in ECG reconstruction techniques. By employing a modular architecture that segments the lead-I ECG signal into multiple frequency components, the model captures essential diagnostic information more effectively than traditional single-model approaches.
    • mEcgNet achieves a remarkable reduction in the number of parameters—approximately 23.1 times fewer than existing state-of-the-art models. This reduction is crucial for deployment on wearable devices, which often have limited computational resources.
    • The model demonstrates a 22.1% reduction in mean squared error (MSE) in ECG reconstruction compared to existing methods, indicating its high accuracy in reconstructing 12-lead ECG signals from a single lead input.
    • With an inference time approximately 5.4 times faster than traditional models
    • By enabling the reconstruction of 12-lead ECG signals from a single lead, mEcgNet addresses a critical need for accurate cardiovascular diagnostics, particularly for complex conditions that require detailed waveform analysis.
    • The authors validate mEcgNet using two public datasets (PTB-XL and Chapman-Shaoxing), ensuring that the results are reliable and reproducible. The comprehensive evaluation strengthens the credibility of the findings.
  • 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 introduces the mEcgNet model; however, the concept of using modular architectures for signal reconstruction is not entirely new. Similar ideas have been explored in previous works, such as EKGAN and ECGrecover. A deeper discussion on how mEcgNet uniquely improves upon these existing approaches would be helpful. Related Work is missing
    • While the paper demonstrates the technical performance of mEcgNet on public datasets, it lacks real-world clinical validation, particularly concerning ECG signals that exhibit pathological patterns. A differentiation in clinical data, especially with respect to various cardiovascular diseases, would be beneficial. Specifically, it would be helpful to assess how well the reconstruction performs with ECG signals that show abnormalities associated with conditions such as bundle branch blocks or myocardial infarction. Understanding the model’s effectiveness in reconstructing these pathological patterns is crucial, as these conditions often manifest as specific changes in the QRS complex and T-wave morphology, which are critical for accurate diagnosis. By including a discussion of the model’s performance on these pathological ECGs, the paper would provide a clearer picture of its clinical applicability and robustness in real-world scenarios.
    • The results presented in the paper rely on specific datasets (PTB-XL and Chapman-Shaoxing). Without additional validation on independent datasets, concerns about overfitting may arise. The paper should address this by providing cross-validation results or testing on a wider range of datasets to ensure generalizability.
    • While the paper compares mEcgNet to EKGAN and ECGrecover, it does not adequately discuss the differences in architecture and methodology that contribute to the performance improvements. Additionally, comparisons with more recent models or different types of architectures could provide a more comprehensive evaluation of mEcgNet’s strengths and weaknesses.
    • The choice of segmenting the ECG signal into three frequency components is based on the authors’ findings, but the rationale for this specific segmentation could be more thoroughly justified. It would be beneficial to explore whether different segmentations yield better results or if the chosen segmentation is optimal across various ECG datasets.
  • 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 mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    • The introduction of mEcgNet represents a significant advancement in the reconstruction of 12-lead ECG signals from a single lead, specifically addressing the challenges faced by wearable IoT devices. The modular architecture and frequency-based segment partitioning are novel features that enhance parameter efficiency and reconstruction accuracy.
    • The experimental results demonstrate significant improvements, achieving approximately 23.1 times fewer parameters and 5.4 times faster inference time compared to existing state-of-the-art models like EKGAN and ECGrecover. Additionally, the model exhibits a 22.1% reduction in reconstruction errors, highlighting its effectiveness and efficiency.
    • The paper is well-structured and clearly articulates the methodology, experimental setup, and results. The use of figures and tables to illustrate findings enhances understanding and makes the content accessible to readers.
    • The paper addresses a critical need for accurate ECG reconstruction in the context of cardiovascular monitoring. However, the lack of clinical validation, particularly concerning the model’s performance on pathological ECG patterns, is a notable weakness. This aspect is crucial for the practical application of the model in clinical settings.
    • Although the results are promising, the reliance on specific datasets raises concerns about overfitting. Additional validation on independent datasets would strengthen the findings and ensure the model’s generalizability.
  • 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.

    The author has responded to the reviewers’ feedback and made suggestions for improvement. Clinical validations and comparisons with other models have been conducted, which supports the relevance of the research. The author demonstrates a willingness to include additional information and explanations to enhance the clarity of the work.




Author Feedback

  1. Lack of clinical validation[R1,2] We believe that we validated our method clinically in two ways. First, we evaluated the reconstruction accuracy by comparing our reconstructed 12-lead ECG with hospital-collected datasets (PTB-XL and Shaoxing), which means the correctness validation in clinical settings (§3.2). Second, we use our reconstructed ECG signals as input to a cardiovascular disease classification model [20] and show that our classification results are comparable (5.7% difference in §3.3) to the ground truth (verified by cardiologists). We continue to do more clinical validation.
  2. Insufficient classification results and explanation—only simple average F1-score[R1,2] We apologize. In fact, we did measure F1-scores for all six diseases (1dAVb, RBBB, LBBB, SB, AF, ST), but due to page limit, we resorted to report only the average. For each disease, mEcgNet outperformed EKGAN and ECGrecover. For example, mEcgNet’s F1-scores for RBBB and LBBB are 1.88× and 12.5× better, respectively. This is because mEcgNet does more accurate reconstruction—up to 22.1% lower error (§3.2). As suggested by the reviewer, we will shorten the ablation study to include these results and explanation.
  3. Insufficient explanation of how mEcgNet improves upon EKGAN and ECGrecover[R2] EKGAN uses two generators that consist of a 5-layer U-Net and a 5-layer autoencoder, resulting in 26.1M params, 2.06 GFLOPs. ECGrecover uses one 5-layer U-Net, 6.2M, 4.59 GFLOPs. In contrast, mEcgNet decomposes signals into three frequency bands, each processed by a 3-layer EcgModule (1.1M, 0.84 GFLOPs). This architecture differentiates mEcgNet from EKGAN and ECGrecover, and furthermore it improves accuracy with lower cost. We will add this explanation as much as the page permits.
  4. Generalizability evaluation missed—cross-validation results, different dataset[R2,3] In fact, we did perform the 5-fold cross-validation but omitted them due to page limit. It is our oversight. Note that EKGAN and ECGrecover did not perform cross-validation in their papers. Our result (MSE std.) on PTB-XL is 0.005 (mEcgNet); on Shaoxing, 0.005. Also, the MAE std. on PTB-XL 0.007; on Shaoxing, 0.007. These results confirm that mEcgNet achieves high generalizability across folds.
  5. Justify the choice of three frequency bands and explore alternatives[R2] The design of three frequency bands is from our observation of the physiological properties of ECG signals. We observe that the waves can be grouped in frequency bands such that P/T waves are 0.5–10 Hz, QRS complex 8–40 Hz, and anomaly waves >40 Hz. We will clarify this in the paper.
  6. Overlooked literature—myocardial infarction (MI) classification from a single lead[Fatimah, Biomed. Signal Process. Control., 2021][R3] Yes, Fatimah et al. detects MI, but please note that it cannot localize MI (e.g., inferior, anterior, or lateral MI). It has been reported that 1) a full 12-lead ECG is necessary for localization[Xiong, Front. Cardiovasc. Med., 2022], 2) detecting ‘anterior’ MI requires leads V1, V2, V3, and V4[Gholipour, Ther. Adv. Cardiovasc. Dis., 2025]. Moreover, diseases such as RBBB, LBBB, and Pericarditis require the full ECGs for diagnosis. We will add the paper mentioned by the reviewer to the reference.
  7. Grammatical errors (e.g., incorrect lowercase use)[R3] We have carefully reviewed the paper and corrected grammatical errors.
  8. Justify the use of MSE/MAE only. Why not signal processing metrics and domain-specific metrics (timing and amplitude)[R2,3] Yes, we agree. Before the paper submission, we did measure the cosine similarity and the timing errors of the normalized QRS complex. Due to page limit, however, we reported only MSE and MAE as they are commonly used in signal reconstruction tasks[Lan, WACV, 2023; Toda, ISCIT, 2021]. Our results are that mEcgNet achieves ~30.4% and ~20% better normalized QRS timing errors than EKGAN and ECGrecover. We will include these results by shortening the ablation study.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top