List of Papers Browse by Subject Areas Author List
Abstract
Heart murmur detection is critical for early diagnosis of cardiovascular conditions. However, existing automated methods typically operate at the recording level and lack interpretability regarding murmur timing and location, limiting their clinical utility. To address this gap, we propose the Physiology-Aligned Hierarchical Network (PHANet), a novel hierarchical framework that explicitly models heart sounds across multiple physiological scales: cardiac phases, cycles, and auscultation sites. With a patient-level supervision strategy, PHANet enables coarse inference of murmur-relevant phases and sites, improving interpretability. Experiments on the CirCor DigiScope data demonstrate that PHANet achieves state-of-the-art performance in murmur detection while providing clinically interpretable insights without fine-grained annotations.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1443_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{LinWen_PhysiologyAligned_MICCAI2026,
author = { Lin, Wenjun AND Guan, Rui AND Soon, Ren Jun AND Lei, Dapeng AND Chui, Chee Kong},
title = { { Physiology-Aligned Hierarchical Network for Heart Murmur Detection } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes PAHNet, a hierarchical deep learning framework for heart murmur detection that explicitly models cardiac physiology across phases, cycles, and auscultation sites. It introduces phase-aware and site-aware fusion modules, along with a patient-level supervision strategy, to enable murmur detection with coarse interpretability without requiring fine-grained annotations.
- 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 addresses an important limitation of existing methods by incorporating physiological structure into the model design, which is a meaningful step toward improving interpretability in heart murmur detection. The hierarchical formulation (phase, cycle, site) is conceptually well aligned with how clinicians analyze heart sounds.
The proposed fusion modules (PAFF and SAPF) provide a structured way to aggregate information across temporal and spatial dimensions, which is a reasonable and potentially effective design.
Another strength is the use of weak (patient-level) supervision to infer phase- and site-level relevance, which, if validated, could reduce reliance on costly fine-grained annotations.
The method achieves competitive performance on the CirCor DigiScope dataset, particularly in imbalance-aware metrics such as F1-score and AUPRC, which are important in clinical settings.
- 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 most significant limitation is the evaluation on a single dataset, which raises concerns about generalizability. Given variability in recording devices, environments, and patient populations, broader validation is necessary.
The novelty of the method is somewhat unclear. While the hierarchical design is well motivated, similar ideas of multi-scale modeling and attention-based aggregation have been explored in prior heart sound analysis and general audio classification literature (e. g. , hierarchical attention networks, multi-instance learning). The paper does not clearly distinguish its contribution from these prior approaches.
The claimed patient-level supervision strategy is insufficiently justified. It remains unclear how reliably the model can infer phase- and site-level relevance without explicit supervision, and no quantitative evaluation is provided to validate these interpretability claims.
Finally, while the method performs well on some metrics, it is not consistently the best across all evaluation criteria, which weakens the claim of clear superiority.
- 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?
I find the problem important and the proposed approach conceptually reasonable, particularly the attempt to incorporate physiological structure and improve interpretability under weak supervision. However, my confidence in the contribution is limited by several factors.
The evaluation is restricted to a single dataset, making it difficult to assess robustness and generalizability. The novelty relative to existing hierarchical and attention-based models is not clearly established, and the interpretability claims are not rigorously validated.
Overall, while the paper presents a promising direction, the current evidence is not sufficient to fully support its claims, leading to a borderline assessment.
- 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
Authors describe a deep learning method for Phonocardiogram analysis which contain the site, phase and patient level. They compare with existing solutions and provide an ablation study.
- 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.
. Including site, phase and patient level
- composite loss function
- comparison with existing solutions and ablation study
- 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.
- Dataset not comprehensively described, especially as it is a pediatric dataset which raises concerns about transfer capabilities
- no mentioning of ethical approval
- no explainability
- no details on dataset, e.g. class frequencies
- 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.
- 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
From the writing it is unclear, if CV has been done at the patient level. This raises concerns about data leakage. Consider reporting F2.Site weighting matrix not disclosed and clinically checked.
- 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?
The description of the used dataset is insufficient and data leakage could not be excluded from the writing. No explainability methods provided and no clinical plausibility check of the site weighting matrix. No code availability, not reproducible.
- 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.
some explanations have been sufficient but overall, iam not convinced, especially the explainability part of the answer is not clearly expressed
Review #3
- Please describe the contribution of the paper
The paper proposes a physiology-aligned hierarchical network (PAHNet) for heart murmur detection, which models phonocardiogram signals across multiple physiological levels, including cardiac phases, cycles, and auscultation sites. The method introduces phase-aware and site-aware fusion mechanisms, along with a patient-level supervision strategy, to enable interpretable and robust murmur detection using only patient-level labels.
- 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.Clinically grounded design: Models heart sounds using physiological hierarchy (phase → cycle → site). 2.Hierarchical framework: Integrates multi-level representations in a unified model (PAHNet). 3.Interpretable approach: Enables phase- and site-level relevance under patient-level supervision. 4.Attention-based fusion: PAFF and SAPF emphasize important phases and sites. 5.Physiology-guided weak supervision: PPC loss incorporates domain knowledge without fine-grained labels. 6.Solid evaluation: Competitive results with comprehensive ablation study.
- 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.Unclear supervision and interpretability claims: It is not clearly stated whether temporal and site-level annotations are used during training or only for evaluation, despite claims of no fine-grained supervision. The interpretability claim is therefore somewhat overstated.
2.Insufficient preprocessing details: The phase segmentation source and temporal standardization (padding/truncation strategy, masking) are not clearly specified, limiting reproducibility.
3.Unjustified design choices in fusion modules: The use of sigmoid-based weighting (instead of normalized attention), the role of the phase-level classifier, and the use of a Transformer over only four phases are not adequately motivated.
4.Strong assumptions in supervision strategy: The PPC loss enforces restrictive clinical priors (e.g., suppressing S1/S2, assuming murmur presence in specific phases), which may limit generalization to noisy or atypical cases.
5.Unclear evaluation protocol and potential data leakage: It is not specified whether cross-validation is performed at the patient level, despite multiple recordings per patient. Handling of variable numbers of auscultation sites is also not described.
6.Fairness of baseline comparison is unclear: It is not stated whether baselines are evaluated under consistent settings. The method does not achieve the best performance on the primary benchmark metric (weighted accuracy), and no statistical significance analysis is provided.
7.Ablation study lacks fine-grained analysis: While overall comprehensive, it does not isolate key design choices and includes configurations that change multiple factors simultaneously.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a clinically meaningful and well-motivated approach for heart murmur detection, with a hierarchical design that aligns with physiological structure and supports interpretable predictions. The proposed framework is conceptually sound and demonstrates competitive performance with a solid ablation study. However, some aspects of the methodology and evaluation require clarification and stronger justification. Overall, I consider this a promising contribution that is marginally above the acceptance threshold, leading to a 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
Author Feedback
We thank the reviewers for high-quality reviews. Below we provide point-to-point responses to the comments.
[R1] Single dataset We use CirCor DigiScope, the largest heart sound data. Crucially, it’s the only data providing phase and site-specific annotations required for evaluation. Robustness is verified via patient-level 5-fold cross-validation, preventing data leakage and ensuring generalizability.
[R1] Novelty PAHNet distinguishes itself by structurally aligning with cardiac physiology and localizing murmur-related phases and sites using only patient-level supervision. This provides clinical interpretability by identifying where and when a murmur occurs, which is not achieved by prior hierarchical or attention-based methods.
[R1] Patient-level supervision justification We provide quantitative validation of interpretability claims in Table2.Removing patient-level supervision (PPC loss, config2) degrades murmur detection, localization, site prediction, and phase classification performance.
[R1/R3] Not the best in all criteria While two models yield higher Accuracy, PAHNet achieves SOTA results in Precision, F1, AUC, and AUPRC and second-best in Recall. These metrics are more representative in unbalanced datasets.
[R2] Dataset description & Ethical approval We use publicly available CirCor DigiScope data (widely used in prior works). It comprises 942 patients, with a murmur distribution of 179 present, 695 absent, and 68 unknown. To prevent leakage, 5-fold cross validation is split at the patient level. As a secondary analysis of de-identified data, this study is exempt from additional ethical approval. We will add more dataset details in the final version.
[R2] Explainability PAHNet is designed with intrinsic explainability. Unlike black-box models that require post-hoc explanation tools, our PAFF/SAPF modules ensure decision-making is explicitly grounded in cardiac phases and sites. Removing these modules (Table2) leads to a significant performance drop, suggesting that PAHNet captures physiologically meaningful features.
[R3] Unclear supervision and interpretability claims Training uses only patient-level labels. Temporal and site-level annotations are used for evaluation only.
[R3] Preprocessing details Phase segmentation followed the official dataset reference. We use 20-cycle consecutive cropping (randomized in training, first 20 in testing). Zero-padding and binary masking are applied to missing cycles or sites.
[R3] Unjustified design choices in fusion modules Sigmoid allows concurrent high weights across sites, essential as murmurs often propagate across multiple sites simultaneously. Phase-level classifier functions as a content-based detector for identifying murmur-like phases. Transformer captures contextual dependencies across phases.
[R3] Strong assumptions in supervision strategy Murmurs are associated with specific phases (systole/diastole) as they arise from abnormal blood flow turbulence. PPC loss leverages this prior to focus on diagnostic phases, enhancing robustness against noise. Ablation study confirms this boosts performance compared to unconstrained learning.
[R3] Unclear evaluation protocol and potential data leakage Data splitting is performed at patient level to prevent leakage. To handle variable numbers of sites, we use zero-padding for missing sites and apply a masking mechanism to ensure that only available sites contribute to diagnosis.
[R3] Fairness of baseline comparison is unclear All baselines were trained using hyperparameters specified in original papers and evaluated under the same data split. PAHNet demonstrated consistent improvements over all baselines across 5-fold cross-validation.
[R3] Ablation study lacks fine-grained analysis While certain factors change together due to functional interdependencies (e.g., removing Phase-level classifier in config4 disables PPC loss), our 7 ablation settings isolate the modular impact of each major innovation (PPC loss, PAFF, SAPF).
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.
The paper proposes PAHNet, a physiology-aligned hierarchical network for heart murmur detection from phonocardiograms. Its main strength is the clinically motivated hierarchical design for multi-level representation integration, but the novelty from a technical point of view is considered limited by the reviewers. Another claim is the patient-level supervision strategy, which is questioned by R1 and R3: it is unclear if the model can reliably infer phase- and site-level relevance, as quantitative evaluation is missing, and this introduces a restrictive clinical prior. All reviewers raised questions about the claimed interpretability. R2 also raised multiple questions about the dataset specification. In the rebuttal, please address the concerns on novelty, rationale and verification of patient-level supervision strategy, and clarify the interpretability claim.
- 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 paper proposes a physiology-aligned hierarchical network for heart murmur detection from phonocardiograms. This is a relatively minority topic for MICCAI. The rebuttal provides some clarification for explainability (with reference to Table 2), which remains unconvincing for R2.However, the clinical motivation, hierarchical design, and imbalance-aware performance make the work marginally above the acceptance threshold. Therefore I recommend acceptance.
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 paper addresses a clinically meaningful problem and proposes a physiology-aligned framework for heart murmur detection with a well-motivated hierarchical design. Reviewers generally agreed that the idea is interesting and clinically relevant, particularly the integration of phase-, cycle-, and site-level modeling under patient-level supervision. The main concerns were related to novelty, interpretability claims, dataset details, and the justification of the supervision strategy. The rebuttal clarified several important points, including the patient-level split to avoid leakage, the role of the supervision strategy, and the intended form of intrinsic interpretability.
