List of Papers Browse by Subject Areas Author List
Abstract
Electrocardiogram (ECG) is a quintessential non-invasive modality for cardiovascular diagnosis. However, existing automated interpretation systems often rely on reconstructed ECG images, leading to a significant loss of fine-grained morphological details, or suffer from long-tail distribution and diagnostic drift in complex clinical scenarios. To address these issues, we propose ECG-CoCa, a unified framework that synergistically integrates contrastive and generative learning for direct interpretation of raw ECG signals. To capture essential ecg features, we design the Physiology-Informed Dual-Granularity Encoder (PI-DGE), which extracts multi-scale morphological details and global rhythmic patterns, providing both global contrastive alignment and token-level generative modeling.To mitigate long-tail bias, we develop Clinical-Concept Aware Reweighting (CCAR), which dynamically optimizes the generative objective by shifting from simple empirical risk to concept-rarity-informed reweighted risk. Furthermore, the Diagnostic-Context Memory (DCM) module bridges the semantic gap by injecting global diagnostic anchors into the multimodal decoder memory, ensuring clinical consistency in report generation.Results across diverse tasks demonstrate that ECG-CoCa effectively captures intricate ECG features, showcasing its versatility for automated diagnostic systems and its potential to enhance the efficiency of large-scale cardiovascular screening.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1094_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/yueyuedeshutu/ECG-CoCa
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuXin_ECGCoCa_MICCAI2026,
author = { Liu, Xinyue AND Lei, Yiming AND Li, Shiwei AND Zhu, Zhiyi AND Hu, Zhenghui AND Liu, Qingjie},
title = { { ECG-CoCa: Physiology-Informed Contrastive and Generative Framework for ECG Interpretation } },
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
This paper presents ECG-CoCa, a unified framework for automated ECG interpretation that combines contrastive and generative learning. The method is designed to capture both waveform morphology and rhythm patterns, address the imbalance of rare conditions, and improve the consistency of generated reports. Experiments on three benchmarks show strong overall performance.
- 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 is well motivated and addresses an important problem in automated ECG interpretation. 2.The proposed technical contributions are meaningful, especially the ECG report augmentation strategy and the augmentation of rare concepts. 3.The empirical evaluation is comprehensive and covers multiple benchmarks and tasks. 4.The paper is generally well written, and the figures are clear and informative.
- 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.A main concern is that the strong performance may be driven in part by the augmented text itself, rather than by better ECG representation learning. Since the original ECG reports are often very concise, heavily augmenting them may make retrieval-based evaluation substantially easier. 2.For this reason, it would be more convincing to directly evaluate the learned ECG representations, for example with KNN evaluation or linear probing. 3.An ablation study without text augmentation would also be important to clarify how much of the gain comes from the proposed framework versus the enriched reports. This analysis is currently missing. 4.For the zero-shot classification task, the experimental setting is not described clearly enough. The paper should explain how prompts are constructed and whether the data splits and evaluation protocol are aligned with those used by the compared baselines. 5.Several implementation details are unclear. For example, in Section 2.2, it is not clear how the clinical vocabulary is defined or how concepts are extracted from reports. The rationale for the two-stage training design in Figure 2 also needs further explanation, especially why the morphology module is frozen during the second stage.
- 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?
This paper has good technique contribution and comprehensive evaluation.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The authors propose ECG-CoCa, a unified foundation framework that synergistically integrates contrastive and generative learning for raw ECG interpretation , leveraging a physiology-informed dual-granularity encoder (PI-DGE) to capture multi-scale morphological and rhythmic features while employing clinical-concept aware reweighting (CCAR) and diagnostic-context memory (DCM) to mitigate long-tail bias and prevent diagnostic drift in report generation
- 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.Physiology-Informed Encoding (PI-DGE): Proposes a dual-granularity encoder that uses multi-scale convolutions to capture micro-morphological details like QRS complexes while employing Transformers to model macro-rhythmic patterns. 2.Clinical-Concept Aware Reweighting (CCAR): Introduces a rarity-informed reweighting strategy that dynamically prioritizes low-frequency pathological findings, effectively mitigating long-tail distribution bias in medical datasets. 3.Diagnostic-Context Memory (DCM): Designs a memory augmentation module that injects global diagnostic anchors into the decoder, forcing the generated text to align with physiological evidence and suppressing “diagnostic drift” or hallucinations.
- 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.Incremental Innovation (“Block-Building” Architecture): The framework appears to be a combination of existing architectural paradigms—specifically combining Contrastive Learning (inspired by CLIP ) and Generative Modeling (similar to HeartLLM ) 2.Unfair Baseline Comparison (Data Advantage): The evaluation lacks a strictly controlled comparison; while many baselines are typically trained only on specific subsets like MIMIC-IV-ECG, this work leverages a much larger multi-source corpus including PTB-XL, SPH, and CSN, making it difficult to discern if the performance gains are due to the architecture or simply the increased data volume 3.The writing needs to be improved to enhance readability.
- Please rate the clarity and organization of this paper
Poor
- 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.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
see above
- 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 authors claim in their rebuttal that all baseline methods were trained on the same datasets. However, while MERL and MELP were trained exclusively on MIMIC-ECG, this work additionally utilized PTB-XL, SPH, and CSN. Consequently, this leads to an unfair comparison.
Review #3
- Please describe the contribution of the paper
This paper proposes ECG‑CoCa, a unified learning framework for automated interpretation of raw electrocardiogram (ECG) signals. The method aims to address limitations in existing ECG interpretation systems, particularly the loss of fine‑grained morphological information introduced by image‑based ECG representations, as well as challenges arising from long‑tailed class distributions and diagnostic drift in real‑world clinical settings. The proposed framework integrates contrastive learning and generative learning within a single architecture. A key component is the Physiology‑Informed Dual‑Granularity Encoder (PI‑DGE), which is designed to capture both multi‑scale morphological characteristics and global rhythmic patterns from raw ECG signals. These representations support global contrastive alignment as well as token‑level generative modeling. To mitigate the effects of class imbalance, the authors introduce Clinical‑Concept Aware Reweighting (CCAR), which adapts the generative learning objective by incorporating concept rarity rather than relying solely on empirical risk. In addition, a Diagnostic‑Context Memory (DCM) module is proposed to inject global diagnostic context into the multimodal decoder, with the goal of improving semantic alignment and clinical consistency in generated diagnostic reports. Experimental results across multiple tasks suggest that ECG‑CoCa is able to capture detailed ECG features and demonstrates applicability across different automated diagnostic scenarios.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is generally well written and easy to follow. Methodology is clear and well articulated. The authors have conducted several experiments, including, cross-modal retrieval, zero-shot classification, and report generation, to demonstrate the utility of the framework.
- 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.
While I appreciate the authors effort in conducting such experiments, I’d printout the following evaluations that are missing, which would have help better justify the proposed contributions.
1.In the introduction, the authors argue that some of the existing discriminative paradigms lack physiological interpretability as they provide only isolated categorical labels. However, in the evaluations, beyond performance measure, I didn’t find any qualitative evidence that the proposed dual‑granularity representations align with clinically meaningful ECG features 2.Another critical limitation is the lack of systematic ablation study that disentangle the contributions of contrastive learning, generative modeling, and their interaction. Without this evaluation the authors cannot properly justify their claims that the combined formulation offers consistent gains over using either paradigm independently. 3.The paper also lacks an analysis of training and inference cost, especially given the multi‑component architecture. Such comparison is essential for the readers to properly evaluate the scalability and feasibility of this framework for large‑scale or real‑time clinical deployment.
- 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.
- 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.
(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?
The proposed framework is novel and achieves significant performance gains
- 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 successfully addressed my concerns on the contributions of contrastive learning, generative modeling, and their interaction. Also, they have commented on the computational complexity. Therefore, I am happy to recommend acceptance.
Author Feedback
We sincerely thank all reviewers for their insightful feedback. We appreciate that the reviewers recognized the strong motivation of ECG-CoCa (R1), its meaningful technical contributions (R1, R3), the core representation modules (R2), and the comprehensive multi-task evaluation (R1, R3). We are also grateful for their recognition of our clear methodology, well-written presentation, and practical utility (R1, R3). We will release all code and datasets upon publication to ensure reproducibility and benefit the community.
R1(1,2,3): Text Augmentation. Original ECG reports are often short and repetitive, leading to semantic collapse in contrastive learning. Text augmentation only adds objective physiological descriptors extracted from the ECG itself, without introducing extra diagnostic labels, thereby enriching supervision while preserving clinical consistency. For a fair comparison, all baseline methods were trained on the same augmented dataset, ensuring that the improvements stem from ECG-CoCa. Motivated by the reviewer’s suggestion, we further evaluated frozen ECG features via linear probing. ECG-CoCa achieves SOTA performance and remains competitive even without text augmentation. We appreciate this insightful advice and will consider more comprehensive representation-level analyses in future work.
R1(4,5): Details of Implementation. Following prior works such as MERL and ECG-Chat, zero-shot prompts and the CCAR vocabulary are derived from standardized ECG diagnostic terminology, using clinically equivalent synonyms to reduce abbreviation ambiguity. For CCAR, clinical concepts are extracted by matching normalized report text to a predefined ECG concept vocabulary. These concept frequencies are then used to compute rarity-aware sample weights for generative training.
R1(6): Morphology Module. Following common practice in MLLM alignment stage, the morphology module is frozen during Stage 2 to preserve stable features and prevent feature drift. Unfreezing it led to a 5% drop in zero-shot ACC, indicating that freezing this module is critical for maintaining robust and transferable ECG representations.
R2(1): Novelty and Effectiveness of ECG-CoCa. ECG-CoCa is not a simple combination of contrastive and generative learning, but a unified framework that jointly models ECG representation, cross-modal alignment, and report generation. Unlike CLIP-style methods focusing on global alignment, ECG-CoCa preserves fine-grained morphology and rhythm dependencies via PI-DGE. In contrast to generative models such as HeartLLM, our CCAR and DCM further mitigate long-tail bias and semantic drift. Consistent improvements across Tables 1–3 further support the effectiveness of this unified design.
R1(4)&R2(2,3): Details of Training Data. All compared methods were trained on the same datasets and identical patient-level/official splits to avoid data leakage, ensuring that the gains arise from the proposed architecture rather than additional training data.
R3(1): Interpretability. We validated the learned representations through linear probing on PTB-XL, where the features showed clear separability for both morphology- and rhythm-related categories, supporting the physiological relevance of the proposed dual-granularity representations. We appreciate the reviewer’s suggestion on more qualitative interpretability analyses and will consider additional visualizations and case studies in future work.
R3(2,3): Training Objectives and Efficiency. We separately evaluated contrastive-only, generative-only, and joint training. The unified objective consistently achieved the best overall performance, indicating that contrastive and generative learning are complementary rather than merely additive. In terms of efficiency, joint training introduces only moderate overhead (~44 min/epoch) compared with contrastive-only (~30) and generative-only (~38). We appreciate the reviewer’s suggestion and will consider more comprehensive analyses in future work.
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.
Though the ECG interpretation pipeline the authors proposed show some merits, there exist some concerns about the novelty of the method and the fairness of the evaluation, which need to be solved in the rebuttal phase.
- 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 has solved my concerns very well.
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 author’s response addresses the reviewers’ concerns.
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 main weaknesses concern the evaluation protocol and the extent to which the reported gains can be attributed to the proposed model rather than to the enriched textual supervision and training setup. R1 raised a substantive concern that augmented ECG text may make retrieval and generation easier, and requested representation-level evaluation and ablations without text augmentation. The rebuttal clarifies that text augmentation uses objective physiological descriptors and states that baselines were trained under the same augmented-data setting. It also reports additional linear-probing and no-augmentation analyses. These clarifications reduce the concern, but the additional analyses are not part of the original submission and therefore cannot be treated as the primary basis for acceptance. The submitted paper would still benefit from a clearer description of prompt construction, clinical concept vocabulary construction, baseline retraining, and the exact role of augmented or template-based reports. R2 maintained a reject recommendation after rebuttal, mainly due to concern about unfair baseline comparison and possible data advantage. This concern was considered carefully. The rebuttal explicitly states that all compared methods were trained on the same datasets and identical patient-level or official splits. While this clarification is important, the original paper does not make the baseline retraining protocol sufficiently explicit, so some uncertainty remains. However, R2’s post-rebuttal justification focuses almost exclusively on this issue and does not fully engage with the broader methodological contributions, the submitted ablations, or the positive assessments from R1 and R3.Therefore, the remaining concern is considered a limitation of reporting and experimental transparency rather than a clear fatal flaw. R3 initially requested stronger evidence on the interaction between contrastive and generative learning, interpretability, and computational cost. The rebuttal addressed these points by clarifying separate training-objective comparisons and computational overhead. R3 judged these responses sufficient and maintained an accept recommendation. Some of these details should be added to the camera-ready version to improve reproducibility and transparency. Overall, the paper has meaningful methodological content, a relevant clinical signal-processing application, and a broad empirical evaluation. The novelty is not entirely radical, as the framework builds on established contrastive and generative paradigms, but the ECG-specific integration of dual-granularity raw-signal encoding, concept-aware generation, and diagnostic memory is sufficiently coherent and supported by results.
