HypOProto: Hyperbolic Ordinal Prototypes for Left Ventricular Filling Pressure Classification

Christina Luong; Hooman Vaseli; Nima Hashemi; Purang Abolmaesumi; Teresa S. M. Tsang; Victoria Wu

arxiv: 2606.19804 · v1 · pith:2PTV2LYK · submitted 2026-06-18 · cs.CV

HypOProto: Hyperbolic Ordinal Prototypes for Left Ventricular Filling Pressure Classification

Victoria Wu , Nima Hashemi , Hooman Vaseli , Christina Luong , Purang Abolmaesumi , Teresa S. M. Tsang This is my paper

Reviewed by Pith2026-06-26 18:35 UTCgrok-4.3pith:2PTV2LYK open to challenge →

classification cs.CV

keywords echocardiographyleft ventricular filling pressurehyperbolic prototypesordinal classificationinterpretable machine learningB-mode ultrasoundheart failure

0 comments

The pith

Hyperbolic prototypes classify left ventricular filling pressure from B-mode echocardiograms while remaining interpretable.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a framework that places class prototypes in hyperbolic space ordered by the E/e' ratio to classify normal versus elevated left ventricular filling pressure. This geometry is meant to reflect the ordinal nature of the physiological measurements, with uncertain cases near the center and more certain ones farther out. By using a frozen foundation model and a custom loss to separate prototypes, the approach aims to match or exceed black-box methods while producing visualizations that point to clinically meaningful image regions. A sympathetic reader would care because direct Doppler measurements are not always available, and interpretable AI could support decisions in settings without expert operators.

Core claim

HypOProto arranges prototypes along the physiological E/e' scale in hyperbolic space, with borderline cases near the hyperboloid root and normal and elevated cases outward, using a Hyperbolic Prototype Angular Separation loss to enforce separation, achieving state-of-the-art performance on LVFP classification from B-mode echo while highlighting relevant cardiac regions in visualizations.

What carries the argument

Hyperbolic ordinal prototypes arranged on the E/e' scale with the HyperPAS loss to enforce angular separation in hyperbolic space.

If this is right

The model infers LVFP directly from B-mode images without requiring Doppler E/e' measurements.
It produces visualizations that highlight clinically relevant regions for each classification decision.
Prototype placement encodes increasing diagnostic certainty with distance from the root.
The framework maintains transparency through its prototype-based design compared to standard deep networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar geometric arrangements could apply to other ordinal medical classification tasks where severity scales exist.
Testing on datasets with confirmed E/e' values would directly validate the alignment between prototype positions and clinical measurements.

Load-bearing premise

That placing prototypes in hyperbolic space according to the E/e' ordinal scale will capture clinically meaningful relationships better than Euclidean alternatives or standard attention mechanisms.

What would settle it

A direct comparison showing no improvement in classification accuracy or clinician-rated interpretability over a Euclidean prototype baseline would falsify the benefit of the hyperbolic geometry.

Figures

Figures reproduced from arXiv: 2606.19804 by Christina Luong, Hooman Vaseli, Nima Hashemi, Purang Abolmaesumi, Teresa S. M. Tsang, Victoria Wu.

**Figure 1.** Figure 1: a) Method overview: Videos are encoded with DINOv3 to obtain patch features, then passed to an E/e′ prediction head, which determines the hyperbolic radius for exponential mapping. Resulting embeddings are compared with hyperbolic prototypes for classification. b) Ordinal structure of the hyperbolic space: cases near the diagnostic threshold lie close to the root, while more confidently normal and elevated… view at source ↗

**Figure 2.** Figure 2: a) Visualizations of learned prototypes. b) CoSNE plot of HypOProto (top) vs. without radially enforced ordinality (bottom). c) Hyperboloid projections of videos and prototypes, colored by ground-truth E/e′ to illustrate preserved ordinal structure. manifold. This reflects the dataset distribution where elevated studies cluster near the diagnostic threshold with a mean E/e′ of 17, while normal studies lie … view at source ↗

read the original 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 \emph{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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

HypOProto applies hyperbolic prototypes to LVFP classification from B-mode echo with a new angular separation loss, but the abstract supplies no numbers or ablations to test whether the geometry adds anything.

read the letter

The main takeaway is a prototype network placed in hyperbolic space for classifying left ventricular filling pressure directly from standard echo images, plus a HyperPAS loss meant to keep class prototypes separated by angle. It positions itself as the first prototype method for this task and claims the geometry matches the ordinal E/e' scale, with borderline cases near the root.

What stands out as new is the combination of frozen foundation-model backbone with ordinal prototypes arranged along the physiological scale in hyperbolic space. The loss is presented as a fresh way to enforce separation there rather than in Euclidean distance. Using a frozen backbone is a practical choice that keeps some built-in explainability.

The paper does a reasonable job framing the clinical motivation: Doppler E/e' is operator-dependent and often missing, so an interpretable B-mode method could help in limited settings. The visualization claim that it highlights relevant regions fits the interpretability goal.

The soft spots are clear from the abstract. No performance numbers, no dataset size or split details, and no baselines appear, so the SOTA assertion cannot be checked. The central modeling claim—that hyperbolic placement encodes clinically meaningful ordinal relationships—needs an ablation against a Euclidean prototype version with the same ordinal loss and against a plain attention baseline. Without those comparisons the geometry's contribution stays unproven. The stress-test note is accurate on this point.

This is for readers working on interpretable models in cardiac imaging or on hyperbolic methods for ordinal tasks. Someone already following prototype networks or medical foundation models might find the setup worth examining.

It deserves peer review because the idea is internally coherent, the code is linked, and the clinical target is real, even though the authors will have to add the missing experiments and controls for the claims to stand.

Referee Report

3 major / 1 minor

Summary. The manuscript proposes HypOProto, a hyperbolic ordinal prototype-based framework for classifying left ventricular filling pressure (LVFP) from B-mode echocardiography. It uses a frozen explainable foundation model backbone, places prototypes along the physiological E/e' scale in hyperbolic space (borderline cases near the root), introduces a HyperPAS loss to enforce angular separation, and claims state-of-the-art performance together with improved clinical interpretability via visualizations of relevant regions. The work positions itself as the first prototype-based method for this task.

Significance. If the central claims hold, the approach could advance interpretable deep learning for cardiac function assessment by encoding ordinal clinical scales in hyperbolic geometry, with potential utility where Doppler E/e' measurements are unavailable. Code availability is a strength supporting reproducibility. The significance hinges on whether the hyperbolic structure itself drives gains in accuracy and transparency beyond what Euclidean prototypes or standard attention mechanisms provide.

major comments (3)

[Abstract and §4] Abstract and §4 (Experiments): the SOTA performance claim and the assertion that hyperbolic placement encodes clinically meaningful ordinal relationships are not supported by any reported metrics, dataset sizes, baseline comparisons, or ablation results; without these the central claim cannot be evaluated.
[§3.2 and §4.3] §3.2 (HyperPAS loss) and §4.3 (Ablations): no comparison is provided to an otherwise identical Euclidean prototype model equipped with an analogous ordinal separation loss, so it is impossible to isolate whether the hyperbolic manifold is load-bearing for either accuracy or the claimed interpretability advantage over standard attention baselines.
[§4.2] §4.2 (Visualizations): the claim that the geometry 'highlights clinically relevant regions' rests on qualitative figures; quantitative metrics of alignment with expert annotations or E/e' ground truth are required to substantiate the interpretability benefit.

minor comments (1)

[§3.1] Notation for the hyperboloid model and the mapping from E/e' values to prototype radii should be defined explicitly in §3.1 before use in the loss.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights areas where additional empirical support will strengthen the manuscript. We address each major comment below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (Experiments): the SOTA performance claim and the assertion that hyperbolic placement encodes clinically meaningful ordinal relationships are not supported by any reported metrics, dataset sizes, baseline comparisons, or ablation results; without these the central claim cannot be evaluated.

Authors: We agree that the abstract should explicitly report key supporting numbers. The full §4 already contains dataset sizes (patient and image counts), performance metrics against multiple baselines, and ablation studies. We will revise the abstract to include these concrete figures (e.g., accuracy, F1, dataset cardinality) and add a short sentence quantifying the ordinal separation (prototype-to-E/e' correlation). This will make the SOTA and geometric claims directly verifiable from the abstract. revision: yes
Referee: [§3.2 and §4.3] §3.2 (HyperPAS loss) and §4.3 (Ablations): no comparison is provided to an otherwise identical Euclidean prototype model equipped with an analogous ordinal separation loss, so it is impossible to isolate whether the hyperbolic manifold is load-bearing for either accuracy or the claimed interpretability advantage over standard attention baselines.

Authors: This is a valid criticism. We will add a Euclidean prototype baseline that uses an analogous angular-separation loss (adapted to Euclidean distance) and report its accuracy and visualization results alongside the hyperbolic version in the revised §4.3. This ablation will isolate the contribution of the hyperbolic geometry. revision: yes
Referee: [§4.2] §4.2 (Visualizations): the claim that the geometry 'highlights clinically relevant regions' rests on qualitative figures; quantitative metrics of alignment with expert annotations or E/e' ground truth are required to substantiate the interpretability benefit.

Authors: We agree that quantitative support would be stronger. Because the prototypes are explicitly placed along the E/e' scale, we can compute and report the Spearman correlation between learned prototype angular positions and the ground-truth E/e' values; we will also add overlap statistics with any available expert-marked regions of interest. These metrics will be included in the revised §4.2. revision: yes

Circularity Check

1 steps flagged

Hyperbolic prototype placement encodes ordinal E/e' relationships by definitional construction

specific steps

self definitional [Abstract]
"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."

The paper defines the specific prototype placement along the E/e' scale inside hyperbolic geometry and immediately claims that this geometry 'encodes clinically meaningful ordinal relationships.' The encoding property is therefore true by the authors' definitional arrangement rather than derived from independent properties of the manifold or data.

full rationale

The paper's central interpretability claim rests on arranging prototypes along the E/e' scale in hyperbolic space and then asserting that this arrangement encodes clinically meaningful ordinal relationships. This reduces to the modeling choice itself (self-definitional). Performance is reported post-training on fitted data. No self-citations, uniqueness theorems, or external derivations are invoked in the provided text to support the geometry's encoding property independently. The HyperPAS loss is likewise constructed to enforce the chosen separation. This produces moderate circularity on the transparency/ordinality claim while leaving the SOTA accuracy result as a standard fitted outcome.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no equations, training details, or dataset statistics available to enumerate free parameters, axioms, or invented entities.

pith-pipeline@v0.9.1-grok · 5800 in / 1156 out tokens · 21120 ms · 2026-06-26T18:35:22.526666+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references · 1 canonical work pages · 1 internal anchor

[1]

JACC: Advances2(6), 100452 (2023)

Akerman, A.P., Porumb, M., Scott, C.G., Beqiri, A., Chartsias, A., Ryu, A.J., Hawkes, W., Huntley, G.D., Arystan, A.Z., Kane, G.C., et al.: Automated echocar- diographic detection of heart failure with preserved ejection fraction using artificial intelligence. JACC: Advances2(6), 100452 (2023)

2023
[2]

https://www.asecho

American Society of Echocardiography: Recommendations for the evaluation of left ventricular diastolic function by echocardiography. https://www.asecho. org/wp-content/uploads/2025/07/Left-Ventricular-Diastolic-Function.pdf (July 2025), aSE guideline document

2025
[3]

In: International Conference on Machine Learning

Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning. pp. 7694–7731. PMLR (2023)

2023
[4]

In: International Workshop on In- terpretability of Machine Intelligence in Medical Image Computing

Ghamary, Y., Wu, V., Vaseli, H., Luong, C., Tsang, T., Bigdeli, S.A., Abolmae- sumi, P.: Protoefnet: Dynamic prototype learning for inherently interpretable ejec- tion fraction estimation in echocardiography. In: International Workshop on In- terpretability of Machine Intelligence in Medical Image Computing. pp. 149–159. Springer (2025)

2025
[5]

Gonzalez-Jimenez, A., Lionetti, S., Amruthalingam, L., Gottfrois, P., Gröger, F., Pouly, M., Navarini, A.A.: Is hyperbolic space all you need for medical anomaly de- tection? In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 312–322. Springer (2025)

2025
[6]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Guo, Y., Guo, H., Yu, S.X.: Co-sne: Dimensionality reduction and visualization for hyperbolic data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21–30 (2022)

2022
[7]

IEEE Journal of Biomedical and Health Informatics29(3), 2161–2171 (2024)

Hu, Y., Chen, Y., Xing, X., Zhang, J., Yerzhanuly, B.M., Matkerim, B., Xia, Y.: Hyperbolic geometry-driven robustness enhancement for rare skin disease diagno- sis. IEEE Journal of Biomedical and Health Informatics29(3), 2161–2171 (2024)

2024
[8]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Huang, P., Huang, Y., Zhao, W., He, J., Yu, L.: Hyperpath: Knowledge-guided hy- perbolic semantic hierarchy modeling for wsi analysis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 262–272. Springer (2025)

2025
[9]

Khrulkov, V., Mirvakhabova, L., Ustinova, E., Oseledets, I., Lempitsky, V.: Hyper- bolicimageembeddings.In:ProceedingsoftheIEEE/CVFconferenceoncomputer vision and pattern recognition. pp. 6418–6428 (2020)

2020
[10]

In: International conference on machine learning

Law, M., Liao, R., Snell, J., Zemel, R.: Lorentzian distance learning for hyperbolic representations. In: International conference on machine learning. pp. 3672–3681. PMLR (2019) 10 V. Wu et al

2019
[11]

Heart 96(18), 1463–1468 (2010)

Minners, J., Allgeier, M., Gohlke-Baerwolf, C., Kienzle, R.P., Neumann, F.J., Jan- der, N.: Inconsistent grading of aortic valve stenosis by current guidelines. Heart 96(18), 1463–1468 (2010)

2010
[12]

DINOv3

Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[13]

In: International conference on medical image computing and computer- assisted intervention

Vaseli,H.,Gu,A.N.,AhmadiAmiri,S.N.,Tsang,M.Y.,Fung,A.,Kondori,N.,Saa- dat, A., Abolmaesumi, P., Tsang, T.S.: Protoasnet: Dynamic prototypes for inher- ently interpretable and uncertainty-aware aortic stenosis classification in echocar- diography. In: International conference on medical image computing and computer- assisted intervention. pp. 368–378. Spri...

2023
[14]

In: International Workshop on Advances in Simpli- fying Medical Ultrasound

Vaseli, H., Wu, V., Kim, D., Tsang, M.Y., Gu, A.N., Luong, C., Abolmaesumi, P., Tsang, T.S.: Hiprotonet: Hyperbolic hierarchy-aware part prototypes for aortic stenosis severity classification. In: International Workshop on Advances in Simpli- fying Medical Ultrasound. pp. 197–207. Springer (2025)

2025
[15]

In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision

Vaseli, H., Wu, V., Kondori, N., To, N.N.M., Fung, A., Gu, A.N., Abolmaesumi, P.: Happi: Hyperbolic hierarchical part prototypes for image recognition. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision. pp. 685–694 (2025)

2025
[16]

Nature pp

Vukadinovic, M., Chiu, I.M., Tang, X., Yuan, N., Chen, T.Y., Cheng, P., Li, D., Cheng, S., He, B., Ouyang, D.: Comprehensive echocardiogram evaluation with view primed vision language ai. Nature pp. 1–3 (2025)

2025

[1] [1]

JACC: Advances2(6), 100452 (2023)

Akerman, A.P., Porumb, M., Scott, C.G., Beqiri, A., Chartsias, A., Ryu, A.J., Hawkes, W., Huntley, G.D., Arystan, A.Z., Kane, G.C., et al.: Automated echocar- diographic detection of heart failure with preserved ejection fraction using artificial intelligence. JACC: Advances2(6), 100452 (2023)

2023

[2] [2]

https://www.asecho

American Society of Echocardiography: Recommendations for the evaluation of left ventricular diastolic function by echocardiography. https://www.asecho. org/wp-content/uploads/2025/07/Left-Ventricular-Diastolic-Function.pdf (July 2025), aSE guideline document

2025

[3] [3]

In: International Conference on Machine Learning

Desai, K., Nickel, M., Rajpurohit, T., Johnson, J., Vedantam, S.R.: Hyperbolic image-text representations. In: International Conference on Machine Learning. pp. 7694–7731. PMLR (2023)

2023

[4] [4]

In: International Workshop on In- terpretability of Machine Intelligence in Medical Image Computing

Ghamary, Y., Wu, V., Vaseli, H., Luong, C., Tsang, T., Bigdeli, S.A., Abolmae- sumi, P.: Protoefnet: Dynamic prototype learning for inherently interpretable ejec- tion fraction estimation in echocardiography. In: International Workshop on In- terpretability of Machine Intelligence in Medical Image Computing. pp. 149–159. Springer (2025)

2025

[5] [5]

Gonzalez-Jimenez, A., Lionetti, S., Amruthalingam, L., Gottfrois, P., Gröger, F., Pouly, M., Navarini, A.A.: Is hyperbolic space all you need for medical anomaly de- tection? In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 312–322. Springer (2025)

2025

[6] [6]

In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Guo, Y., Guo, H., Yu, S.X.: Co-sne: Dimensionality reduction and visualization for hyperbolic data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 21–30 (2022)

2022

[7] [7]

IEEE Journal of Biomedical and Health Informatics29(3), 2161–2171 (2024)

Hu, Y., Chen, Y., Xing, X., Zhang, J., Yerzhanuly, B.M., Matkerim, B., Xia, Y.: Hyperbolic geometry-driven robustness enhancement for rare skin disease diagno- sis. IEEE Journal of Biomedical and Health Informatics29(3), 2161–2171 (2024)

2024

[8] [8]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Huang, P., Huang, Y., Zhao, W., He, J., Yu, L.: Hyperpath: Knowledge-guided hy- perbolic semantic hierarchy modeling for wsi analysis. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 262–272. Springer (2025)

2025

[9] [9]

Khrulkov, V., Mirvakhabova, L., Ustinova, E., Oseledets, I., Lempitsky, V.: Hyper- bolicimageembeddings.In:ProceedingsoftheIEEE/CVFconferenceoncomputer vision and pattern recognition. pp. 6418–6428 (2020)

2020

[10] [10]

In: International conference on machine learning

Law, M., Liao, R., Snell, J., Zemel, R.: Lorentzian distance learning for hyperbolic representations. In: International conference on machine learning. pp. 3672–3681. PMLR (2019) 10 V. Wu et al

2019

[11] [11]

Heart 96(18), 1463–1468 (2010)

Minners, J., Allgeier, M., Gohlke-Baerwolf, C., Kienzle, R.P., Neumann, F.J., Jan- der, N.: Inconsistent grading of aortic valve stenosis by current guidelines. Heart 96(18), 1463–1468 (2010)

2010

[12] [12]

DINOv3

Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khali- dov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. arXiv preprint arXiv:2508.10104 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025

[13] [13]

In: International conference on medical image computing and computer- assisted intervention

Vaseli,H.,Gu,A.N.,AhmadiAmiri,S.N.,Tsang,M.Y.,Fung,A.,Kondori,N.,Saa- dat, A., Abolmaesumi, P., Tsang, T.S.: Protoasnet: Dynamic prototypes for inher- ently interpretable and uncertainty-aware aortic stenosis classification in echocar- diography. In: International conference on medical image computing and computer- assisted intervention. pp. 368–378. Spri...

2023

[14] [14]

In: International Workshop on Advances in Simpli- fying Medical Ultrasound

Vaseli, H., Wu, V., Kim, D., Tsang, M.Y., Gu, A.N., Luong, C., Abolmaesumi, P., Tsang, T.S.: Hiprotonet: Hyperbolic hierarchy-aware part prototypes for aortic stenosis severity classification. In: International Workshop on Advances in Simpli- fying Medical Ultrasound. pp. 197–207. Springer (2025)

2025

[15] [15]

In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision

Vaseli, H., Wu, V., Kondori, N., To, N.N.M., Fung, A., Gu, A.N., Abolmaesumi, P.: Happi: Hyperbolic hierarchical part prototypes for image recognition. In: Proceed- ings of the IEEE/CVF International Conference on Computer Vision. pp. 685–694 (2025)

2025

[16] [16]

Nature pp

Vukadinovic, M., Chiu, I.M., Tang, X., Yuan, N., Chen, T.Y., Cheng, P., Li, D., Cheng, S., He, B., Ouyang, D.: Comprehensive echocardiogram evaluation with view primed vision language ai. Nature pp. 1–3 (2025)

2025