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

Abstract

Electrocardiogram (ECG) denoising enhances the clarity of noisy signals while preserving or even improving diagnostic performance. Most existing single-lead denoising algorithms require a preliminary noise assessment across all 12 leads, discarding clean leads and denoising only the noisy leads. In this paper, a novel disentanglement learning denoising network is proposed for 12-lead wearable ECG that directly processes 12-lead ECG, denoising noisy leads while preserving clean leads. Specifically, the proposed network takes both raw ECG and its corresponding simulated noisy ECG as inputs, disentangling them into noise codes and signal content codes. To ensure consistency between the content codes from two inputs, a discriminator is introduced. Additionally, considering that clean leads within the same ECG can provide valuable information for denoising noisy leads, a lead encoder is designed to extract lead specific features from the raw ECG. A contrastive loss is then applied between the features of noisy and clean leads to optimize the model. The results demonstrate that our method achieves superior denoising performance across two different lead system test datasets. Furthermore, evaluations on an ST-segment change multi-label classification task indicate that the denoised ECG improve diagnostic AUC and AUPRC. Furthermore, our model can be used into remote wearable ECG diagnostic workflows, providing preliminary noise reduction to assist cardiologists in subsequent clinical assessments.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3315_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{ZhaYue_Contrastive_MICCAI2025,
        author = { Zhang, Yue and Zhao, Chenyu and Zhang, Wen and Wang, Jinliang and Guo, Jun and Yang, Wei and Feng, Qianjin},
        title = { { Contrastive Disentanglement Learning Framework for Multi-lead Wearable ECG Denoising } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15962},
        month = {September},
        page = {121 -- 130}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors propose a framework called CD-ECGNet for denoising 12-lead ECG signals and conduct experiments on both wearable ECG and clinical ECG datasets. The key contributions are:

    1. Development of a disentanglement learning approach using two encoders to separate noisy codes and content codes.
    2. Application of contrastive learning between clean lead features and noisy lead features to improve denoising performance.
    3. Implementation of a discriminator to ensure consistency between content codes extracted from original and simulated noisy inputs.
  • 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. Practical clinical application: Addresses a real-world problem in wearable ECG monitoring where noise affects leads unevenly
    2. Comprehensive evaluation: Tests performance on both wearable ECG dataset and standard clinical PTB-XL dataset
    3. Superior performance: Outperforms classical filtering methods and state-of-the-art deep learning approaches across multiple metrics
    4. Thorough ablation study: Demonstrates the contribution of each component (disentanglement learning, discriminator, contrastive loss)
  • 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. Insufficient explanation of evaluation metrics: No description of SSD, PRD, and other metrics presented in Tables 1 and 2
    2. Unclear noise simulation methodology: Limited details on how noise is added to the original signals beyond scaling amplitude with a random factor
    3. Inadequate noise profile analysis: No evaluation of performance under different noise types (baseline wander, electrode motion artifact, etc.)
    4. Dataset quality concerns: No discussion of how intrinsic noise in the original datasets affects evaluation
    5. Ambiguous lead classification: No clear explanation of how “clean” versus “noisy” leads are defined or annotated
    6. Questionable visual results: Figure 2(b) shows reduced amplitude of important waveforms (P wave, Q wave), suggesting potential loss of clinical information.
    7. Limited diagnostic validation: Evaluates improvement only on ST-segment change classification with marginal gains
    8. Diagram inconsistencies: Ambiguous arrows in Figure 1(a) need correction
    9. Unclear ablation study design: Does not specify whether disentanglement learning contains contrastive learning
    10. Contradictory statements: Section underneath equation 4 states “Dis enforce similarity between n and ns,” which contradicts section 2.2 and equation 8
  • 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.

  • 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 the paper presents a clear and novel approach to multi-lead ECG denoising, it lacks critical details about noise profiles, dataset preparation, annotation procedures, and proper interpretation of results. These shortcomings raise concerns about the validity and reproducibility of the findings. The authors should address these issues during rebuttal.

  • 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.

    The authors have addressed most of my previous concerns in their response. While the demonstrated improvement in downstream clinical diagnosis using the proposed denoising method appears modest, the overall concept and design of the framework for multi-lead ECG denoising are interesting and well-motivated. I recommend acceptance.



Review #2

  • Please describe the contribution of the paper

    The authors describe an approach to “disentangle” noise from ECG signals in order to improve the quality of collected data. The general approach to this problem is to recognize that all 12 channels are not fully independent, so all 12-leads are used in the denoising. The proposed approach attempts to focus on only denoising the “noisy” leads while preserving the “clean” signals. To accomplish the task, the approach combines the raw ECG (expected to have both signal and noise components) with a simulated noisy ECG. The model then extracts the clean ECG. Two datasets are used (one public and the other private). The approach is compared to other published methods. Performance is better than some methods and functionally equivalent to others. The method was also used to look at performance of the denoising algorithm on STC detection. Performance was similar in both cases.

  • 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 use of a simulated noisy signal in combination with contrastive learning is novel. Presumably, this will help drive the algorithm to fully “disentangle” the clear signal from the noisy signal. Another strength of the paper is that it compares performance to other published methods.

  • 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 details of the algorithm are vague, focusing on the conceptual framework rather than the implementation. While it is completely appropriate to use both public and private data, no details are provided regarding the demographics of the subjects in the private datasets. The diagnostic performance is roughly equivalent with and without the denoising. No analysis was completed to understand why it didn’t provide a more differentiating performance.

  • 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.

  • 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

    In order to better support the reader audiences, it is important to at least state the full name of the abbreviations (or preferably define them). SSD, PRD, CosSim, SNR, and RMSE are not defined. While they are likely known by some readers, not all readers will be aware of all of them.

  • 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?

    The overall recomendation is based on the lack of implementation details and similar performance to other methods. Presumably a deeper analysis clearly demonstrated a clinical advantage would sway readers to adopt this approach over others.

  • Reviewer confidence

    Very confident (4)

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

    N/A

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

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper proposes a novel deep learning framework for denoising 12-lead wearable ECG signals. Unlike most existing methods that require prior identification of noisy leads, proposed approach directly processes full 12-lead inputs. It employs disentanglement learning to separate noise and cardiac content representations, uses contrastive loss to leverage clean-noisy lead differences, and introduces a lead encoder for inter-lead feature extraction. Evaluations across wearable and clinical datasets show strong performance improvements over classical and deep learning baselines.

  • 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. Authors Combine contrastive learning with disentanglement in a multi-lead ECG setting—something relatively novel in ECG denoising research.
    2. Avoids pre identification of noisy leads and instead allows the model to learn from inter-lead relationships, which reduces manual effort for cardiologists.
    3. The authors’ ability to handle ECG signals that are entirely clean or entirely noisy demonstrates the robustness of their training mechanism.
    4. Shows superior quantitative results on both internal and external test sets, demonstrating generalizability.
    5. Demonstrates that denoising improves downstream diagnostic tasks which is crucial for real-world utility.
    6. Clearly demonstrates the value of each architectural component (disentanglement, discriminator, contrastive loss).
  • 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. The authors mention that 12-lead ECGs are common in clinical settings; however, these are typically recorded in controlled environments where noise is minimal. Given this, can the authors clarify whether the proposed method is primarily intended for wearable or ambulatory ECG scenarios rather than traditional clinical use? Additionally, since most consumer-grade wearable devices support only 1 to 5 leads, the authors should elaborate on the practical prevalence and clinical adoption of 12-lead wearable ECG systems, and clarify the specific deployment context for which this method is designed.

    2. The proposed method uses artificially added noise for training, which may not fully capture the variability of real-world wearable ECG noise. However, the authors acknowledge this domain mismatch. Since simulated noise is injected from an external dataset, I was wondering how authors ensure that the model doesn’t overfit the specific noise characteristics of that dataset.

    3. How do authors ensure that the noise and signal content codes are fully disentangled in the latent space? Is there any quantitative metric or validation method used to confirm effective separation?

    4. Why did authors opt for cosine similarity as the basis for contrastive loss instead of more advanced approaches like NT-Xent or InfoNCE?

    5. Given the multiple encoders and adversarial training, what is the inference time per 15s ECG segment, and is this feasible for real-time wearable applications?

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    This paper presents a novel, technically sound and well-motivated framework for to address practical issue in clinical setting. The manuscript is well written, and the experiments are thorough. The ability to handle entirely clean or noisy inputs also highlights the robustness of the model. I have a few questions regarding the use of simulated noise and deployment assumptions, these do not significantly detract from the contribution.

  • 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.

    I read the rebuttal and authors have addressed my concerns. I went through the other reviewers comments and concerns as well. I belive rebuttal also addressed those. I am satisfied with the proposed method and hence I am going with acceptance.



Author Feedback

We thank all reviewers for their constructive feedback and are glad they recognized the value of our work.

–R#123–: Details of Data: Our wearable ECG were collected from real-world users in China. The CONX CC1612 devices are widely used for daily health monitoring, self-checkups, and by rural doctors for assisted diagnosis. They are sold on major e-commerce platforms. Unlike smartwatches or Holter, wearable ECG are uploaded to the company’s server, where preliminary AI analysis and manual review are usually completed in 15 minutes. Our model runs on the server side and performs initial denoising before diagnosis.

–R#12–: Metrics: Due to space constraints, we will clarify some description. SSD measures the total squared error between the denoised and original ECG over the entire time. MAD measures the maximum absolute difference between the two signals. PRD measures the percentage of overall distortion in the denoised signal relative to the original. Cosine similarity evaluates the similarity in signal patterns by calculating the normalized dot product between two vectors.

–R#1–: 1. Noise Simulation: We simulate noise using the MIT-BIH NST database, which contains 3 half-hour real-world noise recordings. Electrodes were placed on limbs to avoid capturing visible ECG. Noise is added to our data at randomized SNR levels. 2. Data Annotation: All data were annotated by cardiologists. Leads with clear, visible waveforms and no diagnostic ambiguity were labeled as clean, while others were marked as noisy. For evaluation, we only use clean leads to calculate reconstruction performance. 3. Clarification on Results: Cardiologist feedback confirms that the the P wave in Fig. 2(b) is unaffected, while the Q shows a slight peak amplitude reduction, which does not affect the overall diagnosis; We use ST-segment changes in our diagnostic evaluation, as they are crucial indicators of early myocardial ischemia. We plan to extend our evaluation to include more arrhythmias in future work. 4. Clarification on Ablation: The ‘DL’ setting includes the leads contrastive loss design. 5. Clarification on Equations: To enforce similarity between n and ns, we use a min-max optimization strategy similar to GANs. The discriminator (Dis) maximizes its loss, while the encoder-decoder (E+D) minimizes it. Therefore, equations 8 and 9 are not contradictory.

–R#2–: 1. Demographics: The private dataset was collected between Dec. 2021 and Oct. 2022, covering 62.91% male and 37.09% female subjects. The age range from 7 to 108 years (mean: 54.7, median: 55). 2. Denoising Effectiveness: In many cases, the original noisy ECG cannot support clinical diagnosis. After denoising, the ECG showed clear visual improvement, enabling diagnosis and thus providing practical value for cardiologists. Thus, the practical benefit goes beyond metrics and shows clinical usefulness by recovering readable signals.

–R#3–: 1. Wearable Noise: We designed a combination of different ECG noise types and injected them at randomized SNR levels during training to avoid overfitting to specific noise patterns. Furthermore, the model’s success in real-world diagnosis further supports its generalization ability. 2. Disentanglement Validation: A clean ECG can be reconstructed solely from the disentangled signal content codes, which implicitly confirms effective separation. In future work, we plan to visualize the latent space using t-SNE to further assess the distinction between signal and noise embeddings. 3. Choice of Loss: Our lead contrastive loss is a variant of the NT-Xent/InfoNCE. While various similarity metrics can be used, we experimentally evaluated and found cosine similarity to be more effective in our framework. 4. Inference Time: Only the signal encoder and decoder are used during inference. Processing a 15-second 12-lead ECG on a CPU takes 0.24 seconds, which meets the needs for real-time applications.



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