Abstract

Echocardiography (echo) is a widely used imaging modality for assessing cardiac function, with Left Ventricular Filling Pressure (LVFP) serving as a critical physiological marker for conditions such as heart failure. Standard LVFP classification into normal vs elevated categories relies on the Doppler-derived E/e’ ratio, which is operator-dependent and often unavailable in resource-limited settings, motivating methods that infer LVFP directly from B-mode echo. Existing deep learning approaches achieve high performance but remain largely black-box, limiting clinical interpretability. We propose HypOProto, a hyperbolic, ordinal prototype-based framework for interpretable LVFP classification using a frozen, explainable foundation model backbone. HypOProto arranges prototypes along the physiological E/e’ scale, placing borderline cases near the hyperboloid root where small angular differences separate similar cases, while normal and elevated cases occupy outward positions reflecting increasing diagnostic certainty. This hyperbolic geometry encodes clinically meaningful ordinal relationships and improves interpretability. We also introduce a novel Hyperbolic Prototype Angular Separation (HyperPAS) loss, enforcing inter-class prototype separation in hyperbolic space. HypOProto achieves SOTA performance while maintaining transparency, and highlights clinically relevant regions in visualizations. This work represents the first prototype-based framework for LVFP classification in echo. Our code can be found at https://github.com/DeepRCL/HypOProto.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5916_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/DeepRCL/HypOProto

Link to the Dataset(s)

N/A

BibTex

@InProceedings{WuVic_HypOProto_MICCAI2026,
        author = { Wu, Victoria AND Hashemi, Nima AND Vaseli, Hooman AND Luong, Christina AND Abolmaesumi, Purang AND Tsang, Teresa S. M.},
        title = { { HypOProto: Hyperbolic Ordinal Prototypes for Left Ventricular Filling Pressure Classification } },
        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 introduces HypOProto, an interpretable framework that classifies Left Ventricular Filling Pressure (LVFP) from echocardiography videos by leveraging the unique geometry of Lorentzian hyperbolic space. Using a frozen DINOv3 backbone, the model maps feature embeddings and learnable prototypes such that the hyperbolic radius corresponds to clinical severity; borderline $E/e’$ cases reside near the origin while certain cases move outward, effectively encoding ordinal information. To ensure robust classification, the authors introduce a Hyperbolic Prototype Angular Separation (HyperPAS) loss that enforces class distinctness in the tangent space. Validated on a large-scale dataset of ~141K cines, HypOProto outperforms standard Euclidean and MLP-based baselines, offering both improved predictive performance and enhanced qualitative interpretability through prototype visualization.

  • 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 research addresses a high-impact clinical problem by framing LVFP classification as an ordinal task. By mapping clinical ambiguity near the hyperbolic root and severity toward the periphery, the authors provide a geometric intuition that aligns with the $E/e’$ threshold, effectively moving beyond traditional “black-box” approaches.

    The technical design is highly coherent, successfully integrating a frozen DINOv3 backbone with Lorentzian prototype learning. Using predicted $E/e’$ values to determine radial coordinates establishes a direct, interpretable link between the model’s geometry and established clinical metrics.

    The empirical validation is thorough, comparing the framework against a diverse set of baseline architectures across roughly 141K cines, providing confidence intervals. The inclusion of study-level metrics and a detailed discordance analysis adds a level of rigor often missing from prototype-based studies, providing a clear picture of failure modes.

    Qualitative evidence strongly reinforces the paper’s claims, with CoSNE visualizations and prototype activation maps demonstrating that the model correctly localizes features around the mitral valve. These results confirm that the hyperbolic geometry is a functional representation of physiological data rather than just a mathematical abstraction.

  • 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 work is very interesting and I really enjoyed reading it. Although there are some relevant works that have not been cited not discussed.

    First, HyperPAS mirrors existing hyperbolic supervised contrastive objectives, and the interpretation of radial distance representing ambiguity has already defined in [1]. The contribution should be reframed as an adaptation of these concepts to a prototype-based setting rather than a novel invention.

    There is a significant omission regarding Busemann-style prototype learning [2], which is a canonical approach for hyperbolic classification. The authors do not justify why they chose a geodesic formulation requiring an auxiliary regression target over the ideal boundary approach proposed in [2]. A discussion or baseline comparison is necessary to explain the trade-offs of using an externally supervised radial scalar versus the inherent geometric properties of the Busemann framework.

    The methodological advancement over existing works like HiProtoNet and HAPPI appears incremental. The paper lacks a direct head-to-head comparison with these very similar frameworks.

    The manuscript does not describe the angular initialization of the prototypes or the specific hyperparameter values for projection frequency and tolerance. Combined with missing mathematical definitions for log-maps and tangent-space similarity, significantly reduces the reproducibility of the proposed method.

    The statistical evidence supporting the model’s superiority is thin, as reported performance gains often fall within or near one standard deviation of the baselines. In some metrics, such as study-level balanced accuracy,the proposed method is actually outperformed by a standard MLP baseline. The absence of significance testing makes the “SOTA” claims difficult to substantiate given the high data imbalance and small number of training runs.

    [1] Alvaro Gonzalez-Jimenez, Simone Lionetti, Dena Bazazian, Philippe Gottfrois, Fabian Gröger, Marc Pouly, Alexander Navarini. Hyperbolic Metric Learning for Visual Outlier Detection. European Conference on Computer Vision ECCV 2024 BEW Workshop.

    [2] Mina Ghadimi Atigh, Martin Keller-Ressel, Pascal Mettes. Hyperbolic Busemann Learning with Ideal Prototypes. NeurIPS 2021

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    To improve the paper’s theoretical positioning, the methodology should re-contextualize the HyperPAS loss as a tailored adaptation of hyperbolic metric learning for medical prototypes rather than an entirely new objective. Adding a dedicated discussion that explicitly compares the chosen Lorentzian geometry against Busemann-style ideal prototypes would clarify why an $E/e’$ rather than an unconstrained hyperbolic radius is the superior design choice for encoding clinical severity.

    The reproducibility of the work can be significantly enhanced by including the logarithmic map and tangent-space functions while providing a centralized table for all hyperparameters, from curvature values to prototype initialization strategies. Detailing the prototype replacement schedule and its specific impact on training stability would transform the method into a transparent, verifiable framework for other researchers.

    The experimental section would be much more persuasive if it integrated direct regression diagnostics for the $E/e’$ prediction. Reporting $MAE$ or $R^2$ scores is essential to prove that the hyperbolic radius is actually capturing the intended clinical signal rather than just acting as a learned confidence score. Furthermore, replacing simple “without ordinality” ablations with head-to-head comparisons against established models like HiProtoNet using the same backbone would isolate the specific benefits of the proposed geometric modifications.

    The paper should be further motivated by highlighting the computational efficiency of hyperbolic embeddings for low-resource medical settings. Since hyperbolic space can represent hierarchical clinical features with minimal distortion even at very low dimensionality, it offers a significant advantage over Euclidean methodologies that require higher-dimensional vectors to achieve similar representational capacity. Strengthening this link to point-of-care echocardiography would broaden the paper’s impact and practical relevance.

    Figure 2 would benefit from showing each prototype’s assigned $E/e’$ value alongside its activation map, to let the reader visually verify that borderline prototypes attend to different regions than extreme-elevated prototypes. Also executing a pilot reader study where cardiologists grade the relevance of prototype activation maps would provide objective proof of clinical transparency.

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

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

    The paper addresses a significant clinical challenge through a compelling geometric framework that is competently executed and supported by promising qualitative results. While the coherence of the hyperbolic ordinal design offers clear clinical merit, the manuscript currently contains several technical and positioning weaknesses. I recommend a Weak Accept, contingent on the authors adding the missing citations, providing a direct discussion of Busemann prototypes, and clarifying the mathematical definitions and hyperparameters in their rebuttal.

  • 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 paper proposes a hyperbolic ordinality mechanism that organizes prototypes according to the physiological E/e′ scale to better capture graded disease progression and ambiguous cases; and it develops a novel Hyperbolic Prototype Angular Separation (HyperPAS) loss that enforces discriminative separation and diversity among prototypes in hyperbolic space.

  • 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 introduces a principled use of hyperbolic geometry to model ordinal LVFP labels, aligning the embedding space with the physiological ordering of the E/e′ scale. This is a meaningful innovation, as it naturally captures graded disease progression, unlike standard Euclidean approaches.

    2.The model enables case-based reasoning through learned prototypes, offering transparent and clinically relevant explanations. Integrating prototype learning with ordinal structure in hyperbolic space is both original and practically valuable for clinical interpretability.

    3.The proposed Hyperbolic Prototype Angular Separation loss enforces clear inter-class separation while preserving ordinal relationships. This improves prototype diversity and stability, addressing a common limitation in prototype-based models.

    4.By leveraging frozen DINOv3 features, the approach reduces computational cost while maintaining high performance, improving practicality and reproducibility.

    5.Prototype activation maps highlight relevant anatomical regions, and borderline cases are placed near the hyperbolic origin, reflecting diagnostic uncertainty and aligning well with clinical reasoning.

    6.Detailed Ablation studies clearly show that hyperbolic ordinality and the HyperPAS loss contribute directly to improved performance and representation quality.

  • 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 paper does not adequately explain how the spatiotemporal attention module operates or how it interacts with the learned embeddings. In particular, the process by which the model identifies k salient regions of interest (ROIs) remains unclear. This lack of clarity limits interpretability and reproducibility. 2.The model employs a complex combination of five loss functions, yet the rationale behind selecting specific components, particularly L1 loss and transformation consistency loss, is not provided. If mean absolute error (MAE) is already used for predicting E/e’, the additional role and necessity of L1 loss are unclear. Similarly, the purpose and contribution of the transformation consistency loss are not explained, raising concerns about redundancy and over-parameterization. 3.The dataset is highly imbalanced (approximately 85% normal vs. 15% elevated cases), which poses a risk of model bias and overfitting. Which measures have been taken to avoid it. 4.There is a mismatch between the framework diagram and the methodological description. While the text states that five loss functions are applied, the figure does not clearly reflect this design. It is also unclear whether these losses are optimized jointly, assigned to separate heads, or combined in a specific training strategy. This inconsistency creates confusion about the actual implementation and optimization pipeline.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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 framework presents a promising direction for interpretable, ordinal deep learning in cardiac imaging, though the strength of its clinical impact depends on the robustness of evaluation and clarity of methodological details.

  • 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 #3

  • Please describe the contribution of the paper

    This paper addresses the challenge of promoting clinical adoption of deep learning methods by enhancing their interpretability for clinical use. The authors propose HyppOProto, a hyperbolic, prototype-based framework for interpretable classification of Left Ventricular Filling Pressure (LVFP), built on a frozen and explainable foundation model backbone. Experimental results demonstrate state-of-the-art performance, particularly in identifying elevated LVFP, while preserving transparency through visualizations that highlight clinically relevant regions. The paper tackles an important problem by advancing non-invasive approaches for assessing cardiac function. The proposed method shows strong potential to support clinicians in diagnosing and managing heart conditions more effectively.

  • 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 clear and well-written and is thorough in their experiments.
    • The paper proposes a novel prototype-based echocardiography framework using frozen foundation model and is the first interpretable model for LFVP classification.
    • The proposed method HypOProto performs better than all baselines including an improvement in the elevated LVFP class which are the more challenging diagnostic boundary cases. Results also include qualitative visualizations that highlight anatomically relevant regions. The method’s ability to handle ambiguous cases near the diagnostic threshold is well-supported by both quantitative and qualitative result.
  • 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 overall novelty of the paper is more in terms of the application specific for LVFP classification in echocardiography as there have already been similar prototype-based and hyperbolic methods applied to related cardiac tasks.
    • The dataset is heavily imbalanced, with 85% normal and only 15% elevated LVFP cases, which may affect model calibation and sensitiyvity to minority classes
    • The lack of benchmark dataset for fair comparison makes it a bit difficult to judge generalizability as the current study relies on a private dataset from a single tertiary care center.
    • Authors may comment on whether using frozen DINOv3 encoder for feature extraction that is not specifically fine tuned on specific LVFP classification task may limit the models ability.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    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?

    Paper is clear and well-written. The proposed method achieves SOTA performance and performs well even for challenging elevated LVFP classes, while providing transparent and anatomically meaningful visualizations and robust handling of ambiguous cases. Experimental results and ablation studies are comprehensive. Overall, the combination of interpretability and strong empirical results make this paper have strong potential for real-world impact.

  • 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 #4

  • Please describe the contribution of the paper

    This paper proposes HypOProto, an interpretable prototype-based framework for classifying left ventricular filling pressure from B-mode echocardiography videos. The central contribution is to treat LVFP classification not simply as a binary visual recognition problem, but as an ordinal physiological problem governed by the continuous E/e′ scale. To this end, the method uses frozen DINOv3 echo features and maps both video representations and prototypes into a Lorentzian hyperbolic space, where the radial coordinate is controlled by a predicted E/e′ value and is intended to reflect distance from the clinical decision boundary. The paper further introduces Hyperbolic Prototype Angular Separation loss to encourage class-specific prototype directions in hyperbolic space, while retaining prototype-based visual explanations. In practical terms, the work seeks to combine three desirable properties that are often treated separately in echo AI: reduced training cost via frozen foundation features, case-based interpretability via prototypes, and a physiologically meaningful ordinal structure via hyperbolic geometry.

    A more subtle contribution is that the paper does not merely add hyperbolic geometry as a generic representation trick. Its design is motivated by a clinically meaningful ambiguity pattern: cases near the E/e′ threshold are harder and should occupy a region where angular separation is more discriminative, whereas more diagnostically certain cases should move outward. This is an interesting conceptual bridge between geometric representation learning and the clinical structure of diastolic assessment.

  • 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 main strength of the paper is the physiological specificity of the modeling idea. Many prototype-based medical imaging papers use prototypes only as visual exemplars after a standard classification pipeline has already been defined. Here, the prototype space is explicitly organized around the E/e′ continuum, which gives the model a clinically interpretable axis rather than a purely data-driven latent space. This is a meaningful design choice for LVFP because the binary normal/elevated label is only a discretization of an underlying continuous measurement. The paper therefore has a stronger clinical inductive bias than a conventional black-box video classifier or a generic Euclidean prototype network.

    Another strength is the computationally pragmatic use of frozen foundation-model features. Echocardiography videos are expensive to train on end-to-end, especially when labels are noisy, imbalanced, and derived from clinical reports. By using precomputed DINOv3 features and training a lighter prototype-based head, the paper presents a potentially more deployable approach for institutions that cannot afford repeated full video-model training. This also makes the comparison between black-box performance and interpretable performance more interesting, because the proposed method does not rely on brute-force end-to-end optimization to obtain competitive results.

    The elevated LVFP class result is also noteworthy. The reported improvement in elevated-class F1 is clinically more important than a small gain in overall accuracy, because elevated LVFP is the minority and clinically actionable class. The cine-level and study-level results suggest that the proposed formulation improves the harder minority-class decision rather than merely benefiting from the dominant normal class. This is consistent with the stated motivation that borderline and mildly elevated cases require a representation that allocates more discriminative capacity near the decision boundary. The paper’s failure analysis is also valuable because it acknowledges that discordance between report-derived LVFP labels and E/e′ thresholds is a real limitation rather than hiding it behind aggregate metrics.

    The visual interpretation is more convincing than many heatmap-only explanations. The prototype activation maps focus on anatomically plausible regions around the mitral inflow area, and the CoSNE/hyperboloid visualizations attempt to relate prototype position to physiological severity rather than simply showing colorful embeddings. Although I have concerns about whether some of this structure is partially imposed by construction, the effort to connect visualization, prototype identity, and clinical measurement is a strength.

  • 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 important weakness is that the physiological target is not fully disentangled from the mechanism used to organize the representation. The method predicts E/e′ and then uses this predicted value to determine the hyperbolic radius, while the task label itself is derived from clinical reports related to LVFP. This creates an interpretability ambiguity: when the hyperboloid visualization shows a smooth relationship between radial position and ground-truth E/e′, it is not entirely clear how much of this is a learned discovery and how much is a consequence of explicitly forcing the radius to be a function of predicted E/e′. In other words, the most visually appealing part of the representation may be partly guaranteed by the formulation. The paper would be stronger if it separated “geometry imposed by the design” from “structure learned from visual evidence.” This is especially important because the method’s clinical claim depends on the representation being meaningful rather than merely mechanically aligned with a supervised scalar.

    The comparison with baselines is useful but not entirely satisfying. The Akerman et al. model is described as a fully trained closed-source model, but the paper compares an implementation of it on a related but not identical task. Since Akerman et al. target HFpEF detection rather than LVFP classification, the comparison is clinically adjacent rather than task-identical. In addition, the frozen DINOv3 encoder is described as trained on 10,000 multi-view echo studies, but the paper does not sufficiently clarify whether these studies overlap with the private LVFP dataset or with the same institution. Without this information, it is difficult to judge whether the foundation representation is truly external or benefits from institutional/domain overlap.

    The ablation study does not fully support the claimed role of HyperPAS. Removing HyperPAS yields slightly higher balanced accuracy than the full model, while the full model mainly improves prototype sparsity/diversity trade-offs. This does not invalidate the method, but it means HyperPAS is more an interpretability/prototype-organization regularizer than a clear predictive-performance contributor. The current wording sometimes presents it as broadly improving the method; a more precise interpretation would be that it sacrifices or stabilizes some classification behavior in exchange for more desirable prototype geometry. The paper should acknowledge this trade-off more explicitly.

    There are also several issues with presentation and reproducibility. The abbreviation “E2E” appears in the result tables but is not defined, which may confuse readers unsure whether it refers to end-to-end training or another property. The capitalization of DINO/Dino is inconsistent across the paper, for example, between the main text/tables and the ablation table. The manuscript also does not fully specify several implementation choices needed for reproduction. The loss weights λ1 to λ5 are introduced, but not all of their values are clearly specified; the text mentions additional weights for MAE and HyperPAS and states that other weights largely follow ProtoASNet, which is not sufficiently self-contained. It is also not fully explicit which components are trainable beyond the E/e′ head, prototypes, attention module, and other parameters. These issues are not fatal, but they reduce reproducibility and make the method harder to audit.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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

    This is an interesting and clinically motivated paper, and I appreciate that the method tries to make the geometry of the representation correspond to the physiology of the task. My main suggestion is to be more precise about what is learned versus what is imposed. Since the radius is explicitly determined by the predicted E/e′ value, the alignment between radial position and E/e′ should not be presented as entirely emergent evidence of clinical structure. A clearer discussion of this point would make the interpretability claim more credible rather than weaker.

    I also encourage the authors to refine the clinical framing of discordant cases. These cases should not only be described as label noise; they may be exactly the cases where a visually interpretable model would be most clinically valuable. If the model fails strongly on patients whose report-derived LVFP label and E/e′ threshold disagree, then the method may be learning a thresholded E/e′ proxy rather than a broader LVFP phenotype. This distinction matters for clinical deployment.

    Finally, please define E2E in the tables, standardize DINO/Dino capitalization, and make the implementation section more self-contained. In particular, all loss weights, curvature/radius parameters, details of prototype updates, and trainable modules should be specified explicitly.

  • 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 recommend Weak Accept because the paper has a genuinely interesting idea and addresses a clinically meaningful problem with a thoughtful combination of frozen foundation features, prototype-based interpretability, and ordinal hyperbolic geometry. The work is not merely an incremental replacement of Euclidean embeddings with hyperbolic embeddings; it also seeks to align the model’s geometry with the physiological structure of E/e′ and LVFP classification. The reported results are competitive, especially for the elevated LVFP class, and the visualizations are clinically plausible.

    At the same time, my enthusiasm is limited by several conceptual and reproducibility concerns. The strongest interpretability claim depends on a representation whose radial structure is partly imposed by the predicted E/e′ value, so the paper should be more careful about claiming that the learned hyperbolic structure independently discovers ordinal physiology. The target definition is also somewhat ambiguous when report-derived labels and E/e′ thresholds disagree, and the failure analysis suggests that this is not a rare edge case. In addition, the ablation study shows that the full model is not always the best in terms of balanced accuracy, so the benefit of HyperPAS should be framed more as a prototype organization than as an unambiguous predictive improvement. Finally, several implementation details remain underspecified.

    Overall, I find the paper promising and above the acceptance threshold, but I would expect the rebuttal or final version to clarify the clinical target, disentangle imposed versus learned geometry, and improve reproducibility details.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 sincerely thank reviewers for their thoughtful and constructive feedback. We are encouraged that the reviewers recognized the clinical motivation and potential of our work. We agree that the manuscript can be strengthened based on reviewer comments, and we will address these in the revision.

Novelty and Positioning (R1, R3): We agree that our contribution is best framed as an adaptation of hyperbolic learning to a prototype-based clinical setting rather than an entirely new objective. We acknowledge that radial distance encoding ambiguity was previously established, and that HyperPAS shares similarity with hyperbolic contrastive losses. We clarify that: (1) HyperPAS operates exclusively on learnable prototypes, not on all sample pairs, and (2) the radius is not implicitly learned but explicitly anchored to predicted E/e’, enforcing clinically meaningful ordinal structure. We will revise the manuscript to cite the papers mentioned by R1, position the method relative to prior hyperbolic contrastive methods, HiProtoNet, and HAPPI, and clarify our key contribution as the explicit physiological organization of prototypes.

Busemann Prototypes and Geometry Design (R1, R4): In Busemann-style classification, prototypes are positioned at the ideal boundary and class structure emerges from unconstrained geometric properties alone. While elegant, this offers no guarantee that the learned representation aligns with clinically established measurements. Our externally supervised radius enforces fidelity to the E/e’ threshold (14), making the geometry auditable by clinicians and encoding ordinal “distance from the diagnostic boundary,” which boundary-positioned prototypes (maximally committed to a class direction) cannot capture. We will discuss this trade-off and the rationale for clinically imposed vs. purely learned geometry in the revised manuscript.

Experimental Clarifications and Statistical Significance (R1, R4): We acknowledge that some gains are modest compared to reported baselines. Our goal is to balance predictive performance with interpretability and physiologically meaningful representation learning. Although EchoPrime+MLP achieves slightly higher study-level balanced accuracy, it is not explainable. We nevertheless believe the observed improvements are meaningful given the large scale of our dataset and the gains observed across metrics. To address the substantial class imbalance (85% normal / 15% elevated), we employed class-balanced sampling during training, and reported balanced accuracy and per-class F1 rather than standard accuracy. We will make this more explicit in the revised manuscript.

HyperPAS Interpretation (R4): We agree that HyperPAS primarily acts as a prototype organization regularizer rather than a pure performance enhancer. We will revise the manuscript to explicitly state that HyperPAS improves prototype organization, even when classification gains are modest.

Reproducibility and Methodological Clarity (R1, R2, R4): We will expand implementation details and clarify loss optimization and the spatiotemporal attention/ROI selection process. While several components follow prior work, we will explicitly reference the corresponding papers.

Dataset and Foundation Encoder (R3, R4): We agree that external validation is crucial. While public echo datasets with LVFP labels have been scarce, we recognize the recent release of MIMIC-IV-ECHO as a candidate for future benchmarking. We will clarify this limitation and look forward to validating on the external dataset. We will also clarify that the frozen DINOv3 encoder was not fine-tuned on the LVFP task itself, and discuss both the benefits and limitations of relying on pretrained echo representations.

Finally, we will address presentation issues noted by the reviewers, including defining “E2E,” standardizing DINO capitalization, and improving figure clarity. We believe these revisions will strengthen the manuscript and improve clarity and reproducibility.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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 four reviewers have, together, provided a very thorough review, and while the consensus is for marginal acceptance, this is unanimous. The contribution will be improved by considering these careful reviews.



back to top