pith. sign in

arxiv: 2606.19522 · v1 · pith:YYJBVETBnew · submitted 2026-06-17 · 💻 cs.AI

REVEAL++: Differentiable Phenotypic Grouping for Vision-Language Retinal Modeling of Alzheimer's Disease Risk

Ethan Elio Meidinger , Seowung Leem , Zeyun Zhao , Ruogu Fang This is my paper

Pith reviewed 2026-06-26 20:50 UTC · model grok-4.3

classification 💻 cs.AI
keywords Alzheimer's diseaseretinal imagingvision-language modelscontrastive learningphenotypic groupingdifferentiable weightingrisk predictionUK Biobank
0
0 comments X

The pith

A continuous differentiable weighting scheme for phenotypic similarity outperforms discrete groups in vision-language retinal models for Alzheimer's risk prediction.

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

The paper aims to show that phenotypic similarity in retinal and clinical data can be modeled as a learnable continuous signal rather than fixed discrete clusters. It derives soft multi-positive relationships from intra-modality embedding similarities and uses a soft-target contrastive objective to jointly optimize cross-modal alignment and structure. This is tested on UK Biobank retinal images for predicting incident Alzheimer's disease, where it beats both standard vision-language baselines and prior discrete grouping methods. A sympathetic reader would care because disease risk exists on a spectrum, so rigid groupings may lose information that graded supervision can retain.

Core claim

The central claim is that replacing hard group assignments with a differentiable weighting function derived from intra-modality embedding similarities in retinal images and risk profiles enables soft multi-positive pairs and a soft-target contrastive loss, which together produce better cross-modal representations and higher accuracy in predicting future Alzheimer's disease from fundus images than discrete phenotypic grouping approaches.

What carries the argument

Differentiable weighting function derived from intra-modality embedding similarities that defines soft multi-positive relationships through a continuous aggregation operator.

If this is right

  • Joint end-to-end learning of cross-modal alignment and phenotypic structure becomes possible without decoupled group formation.
  • Graded supervision matches the spectrum nature of disease risk instead of imposing rigid boundaries.
  • Consistent gains over discrete group-based contrastive learning and standard vision-language baselines occur on UK Biobank incident AD prediction tasks.
  • A more principled foundation is provided for population-scale modeling that combines retinal images with structured clinical risk narratives.

Where Pith is reading between the lines

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

  • The same continuous weighting idea could be tested on other eye-based biomarkers or additional neurodegenerative conditions where risk is graded rather than binary.
  • Integration with longitudinal follow-up data might reveal whether the learned weights track progression rates more closely than static clinical groupings.
  • If the weighting proves robust, it could reduce reliance on expert-defined thresholds for grouping patients in large biobank studies.
  • External validation on non-UK Biobank cohorts would be a direct next step to check whether the soft relationships generalize beyond the training distribution.

Load-bearing premise

Similarity weights computed from intra-modality embeddings accurately reflect the true spectrum of disease risk without circular dependence on the representations being learned.

What would settle it

An ablation on the same UK Biobank retinal dataset in which the continuous weighting is replaced by random weights or fixed discrete groups and the performance advantage disappears would falsify the benefit of the differentiable formulation.

Figures

Figures reproduced from arXiv: 2606.19522 by Ethan Elio Meidinger, Ruogu Fang, Seowung Leem, Zeyun Zhao.

Figure 1
Figure 1. Figure 1: Architecture of the proposed differentiable phenotypic weighting framework for group-aware contrastive learning. Image and text embeddings are aligned via a similarity-weighted multi-positive contrastive loss, where continuous phenotypic weights replace hard grouping to model the heterogeneous spectrum of Alzheimer’s disease risk late before clinical symptoms [6]. Advances in brain imaging and plasma-based… view at source ↗
read the original abstract

The retina offers a noninvasive window into neurodegenerative disease, capturing subtle structural patterns associated with a risk of future cognitive decline. Vision-language alignment frameworks such as REVEAL have shown that pairing retinal fundus images with structured clinical risk narratives improves early prediction of Alzheimer's disease (AD). A key design choice in these approaches is the use of phenotypic grouping, where individuals with similar risk profiles are treated as multi-positive pairs during contrastive learning. However, existing methods operationalize phenotypic similarity as a discrete construct, relying on hard group assignments that impose rigid supervision and decouple group formation from representation learning. We propose a continuous formulation of phenotypic structure within contrastive learning. Rather than assigning samples to fixed clusters, we model inter-subject similarity as a differentiable weighting function derived from intra-modality embedding similarities in both retinal images and risk profiles. These weights define soft multi-positive relationships through a continuous aggregation operator, enabling graded supervision that reflects the spectrum nature of disease risk. We further introduce a soft-target contrastive objective that jointly learns cross-modal alignment and phenotypic structure in an end-to-end manner. Evaluated on UK Biobank retinal imaging data for incident AD prediction, the proposed framework consistently outperforms discrete group-based contrastive learning and standard vision-language baselines. By treating phenotypic similarity as a learnable, continuous signal rather than a fixed grouping rule, our approach provides a principled and robust foundation for population-scale neurodegenerative risk modeling from multi-modal retinal and clinical data.

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.

Referee Report

2 major / 1 minor

Summary. The paper proposes REVEAL++, an extension of vision-language retinal modeling for Alzheimer's disease (AD) risk prediction. It replaces discrete phenotypic grouping with a continuous, differentiable weighting function derived from intra-modality embedding similarities in retinal images and risk profiles. These weights enable soft multi-positive relationships via a continuous aggregation operator and a soft-target contrastive objective that jointly optimizes cross-modal alignment and phenotypic structure in an end-to-end manner. The framework is evaluated on UK Biobank retinal imaging data for incident AD prediction and is claimed to consistently outperform discrete group-based contrastive learning and standard vision-language baselines.

Significance. If the central claims hold after addressing the circularity concern, the continuous formulation of phenotypic similarity would represent a meaningful technical advance in contrastive learning for medical multimodal data. It could better capture the spectrum nature of neurodegenerative risk, enabling more flexible supervision than hard clustering and supporting improved population-scale early AD prediction from retinal images paired with clinical narratives. The end-to-end differentiability is a clear strength relative to prior discrete grouping approaches.

major comments (2)
  1. [Abstract] Abstract (third paragraph): the weighting function is defined directly from intra-modality embedding similarities produced by the encoders that are jointly optimized by the soft-target contrastive loss. No stop-gradient, separate pre-training stage, or external similarity source is described, so the operator can amplify structure already present in the current representations. This creates a load-bearing risk that reported gains over discrete baselines reflect self-reinforcement rather than independent capture of the AD risk spectrum.
  2. [Abstract] Abstract (final paragraph): the claim of consistent outperformance on UK Biobank incident AD prediction is stated without any quantitative metrics, statistical tests, baseline details, or confound handling. This prevents assessment of whether the evaluation supports the central claim that the continuous formulation improves upon discrete grouping.
minor comments (1)
  1. The abstract would be strengthened by including at least one key performance number (e.g., AUC or hazard ratio with confidence interval) to ground the outperformance statement.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. Below we provide point-by-point responses to the two major comments, indicating where we agree revisions are warranted and where we offer substantive clarification while preserving the core technical contribution.

read point-by-point responses

  1. Referee: [Abstract] Abstract (third paragraph): the weighting function is defined directly from intra-modality embedding similarities produced by the encoders that are jointly optimized by the soft-target contrastive loss. No stop-gradient, separate pre-training stage, or external similarity source is described, so the operator can amplify structure already present in the current representations. This creates a load-bearing risk that reported gains over discrete baselines reflect self-reinforcement rather than independent capture of the AD risk spectrum.

    Authors: The joint optimization is intentional: it allows phenotypic weights to adapt to the evolving representations under the supervision of the cross-modal contrastive loss, which incorporates independent clinical narrative information. This design captures the continuous spectrum of AD risk more flexibly than static discrete groups. While we acknowledge the referee's concern about potential self-reinforcement, the soft-target objective ties the intra-modality similarities to the downstream prediction task, providing an external anchor. We will add a dedicated paragraph in the methods and discussion sections clarifying this rationale and include an ablation with stop-gradient on the weighting function to demonstrate the value of end-to-end learning. revision: partial

  2. Referee: [Abstract] Abstract (final paragraph): the claim of consistent outperformance on UK Biobank incident AD prediction is stated without any quantitative metrics, statistical tests, baseline details, or confound handling. This prevents assessment of whether the evaluation supports the central claim that the continuous formulation improves upon discrete grouping.

    Authors: We agree that the abstract would be strengthened by including concrete quantitative support. In the revised version we will incorporate key performance metrics (e.g., AUC improvements over discrete baselines), mention of statistical testing, and a brief note on confound handling while remaining within abstract length limits. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation remains self-contained

full rationale

The provided abstract describes an end-to-end framework in which intra-modality embedding similarities are used to define differentiable weights for a soft-target contrastive loss, but no equations are shown that reduce the final AD prediction performance or phenotypic structure to the input embeddings by construction. The reference to prior REVEAL work is background context only and does not serve as a load-bearing uniqueness theorem or ansatz for the central continuous formulation. No fitted parameter is renamed as an independent prediction, and no self-citation chain is invoked to forbid alternatives. The method is presented as a joint optimization whose outputs are evaluated against external UK Biobank incident AD labels, keeping the derivation independent of its own fitted values.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no equations or implementation details; no free parameters, axioms, or invented entities can be identified with certainty.

pith-pipeline@v0.9.1-grok · 5801 in / 1075 out tokens · 26875 ms · 2026-06-26T20:50:20.264963+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 14 canonical work pages

  1. [1]

    Optuna: A next-generation hyperparameter optimization framework

    Akiba, T., Sano, S., Yanase, T., Ohta, T., Koyama, M.: Optuna: A next- generation hyperparameter optimization framework. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Min- ing (KDD ’19). pp. 2623–2631. ACM, New York, NY, USA (2019).https: //doi.org/10.1145/3292500.3330701

    work page doi:10.1145/3292500.3330701 2019

  2. [2]

    https://doi.org/10.1007/s11910-025-01451-5

    Aktan Süzgün, M., Tang, Q., Stefani, A.: Sleep abnormalities and risk of alzheimer’sdisease.CurrentNeurologyandNeuroscienceReports25(1), 67(2025). https://doi.org/10.1007/s11910-025-01451-5

    work page doi:10.1007/s11910-025-01451-5 2025

  3. [3]

    Journal of Neuroinflammation21(1), 309 (2024).https://doi.org/10

    Banna, H., Slayo, M., Armitage, J., Del Rosal, B., Vocale, L., Spencer, S.: Imag- ing the eye as a window to brain health: frontier approaches and future direc- tions. Journal of Neuroinflammation21(1), 309 (2024).https://doi.org/10. 1186/s12974-024-03304-3

    2024

  4. [4]

    The Lancet Regional Health – Western Pacific64, 101743 (2025)

    Bueno Lopez, C., Iona, A., Avery, D., Turnbull, I., Yang, L., Du, H., Chen, Y., Zhang, N., Chen, J., Pei, P., Lv, J., Yu, C., Sun, D., Li, L., Bennett, D., van Duijn, C., Clarke, R., Chen, Z., Bragg, F.: Cardiometabolic health and risk of dementia and brain atrophy: a community-based prospective cohort study of 0.5 million adults in china. The Lancet Regi...

    work page doi:10.1016/j.lanwpc.2025.101743 2025

  5. [5]

    The UK Biobank resource with deep phenotyping and genomic data

    Bycroft, C., Freeman, C., Petkova, D., Band, G., Elliott, L.T., Sharp, K., Motyer, A., Vukcevic, D., Delaneau, O., O’Connell, J., Cortes, A., Welsh, S., Young, A., Effingham, M., McVean, G., Leslie, S., Allen, N., Donnelly, P., Marchini, J.: The uk biobank resource with deep phenotyping and genomic data. Nature562(7726), 203–209 (2018).https://doi.org/10....

    work page doi:10.1038/s41586-018-0579-z 2018

  6. [6]

    Online learning: A comprehensive survey

    Chow, K.H.M., Abel, T.: Neurodevelopmental origins of age-related neurodegen- erative diseases. eBioMedicine124, 106151 (2026).https://doi.org/10.1016/j. ebiom.2026.106151

    work page doi:10.1016/j 2026

  7. [7]

    arXiv preprint arXiv:2405.14137 (2024)

    Du, J., Guo, J., Zhang, W., Yang, S., Liu, H., Li, H., Wang, N.: Ret-clip: A retinal image foundation model pre-trained with clinical diagnostic reports. arXiv preprint arXiv:2405.14137 (2024)

    arXiv 2024

  8. [8]

    Global Advances in Health and Medicine2(5), 38–43 (September 2013).https: //doi.org/10.7453/gahmj.2013.008

    Gagnier, J.J., Kienle, G., Altman, D.G., Moher, D., Sox, H., Riley, D., Group, C.: Thecareguidelines:Consensus-basedclinicalcasereportingguidelinedevelopment. Global Advances in Health and Medicine2(5), 38–43 (September 2013).https: //doi.org/10.7453/gahmj.2013.008

    work page doi:10.7453/gahmj.2013.008 2013

  9. [9]

    arXiv preprint arXiv:2407.21783 (2024),https://arxiv.org/abs/2407.21783

    Grattafiori, A., Dubey, A., Jauhri, A., et al.: The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024),https://arxiv.org/abs/2407.21783

    Pith/arXiv arXiv 2024

  10. [10]

    JAR Life13, 1–21 (January 2024)

    Hayden, K.M., Mielke, M.M., Evans, J.K., Neiberg, R., Molina-Henry, D., Culkin, M.,Marcovina,S.,Johnson,K.C.,Carmichael,O.T.,Rapp,S.R.,Sachs,B.C.,Ding, J., Shappell, H., Wagenknecht, L., Luchsinger, J.A., Espeland, M.A.: Association between modifiable risk factors and levels of blood-based biomarkers of alzheimer’s and related dementias in the look ahead ...

    work page doi:10.14283/jarlife.2024.1 2024

  11. [11]

    Alzheimer’s Research & Therapy16, 238 (October 2024).https: //doi.org/10.1186/s13195-024-01602-9

    Huszár, Z., Solomon, A., Engh, M.A., Koszovácz, V., Terebessy, T., Molnár, Z., Hegyi, P., Horváth, A., Mangialasche, F., Kivipelto, M., Csukly, G.: Association of modifiable risk factors with progression to dementia in relation to amyloid and tau pathology. Alzheimer’s Research & Therapy16, 238 (October 2024).https: //doi.org/10.1186/s13195-024-01602-9

    work page doi:10.1186/s13195-024-01602-9 2024

  12. [12]

    Jack, C.R.J., Bennett, D.A., Blennow, K., Carrillo, M.C., Dunn, B., Haeberlein, S.B., Holtzman, D.M., Jagust, W., Jessen, F., Karlawish, J., Liu, E., Molinuevo, 10 Meidinger et al. J.L., Montine, T., Phelps, C., Rankin, K.P., Rowe, C.C., Scheltens, P., Siemers, E., Snyder, H.M., Sperling, R.: Nia-aa research framework: Toward a biological definition of al...

    work page doi:10.1016/j.jalz.2018.02.018 2018

  13. [13]

    In: Medical Imaging with Deep Learning (2026),https: //openreview.net/pdf?id=aOKAXRHXVw, accepted by MIDL 2026

    Leem, S., Gu, L., You, C., Gong, K., Fang, R.: REVEAL: Multimodal vision– language alignment of retinal morphometry and clinical risks for incident AD and dementia prediction. In: Medical Imaging with Deep Learning (2026),https: //openreview.net/pdf?id=aOKAXRHXVw, accepted by MIDL 2026. Proceedings of Machine Learning Research (PMLR)

    2026

  14. [14]

    (eds.): Preventing Cognitive Decline and Dementia: A Way Forward

    Leshner, A.I., Landis, S., Stroud, C., Downey, A. (eds.): Preventing Cognitive Decline and Dementia: A Way Forward. National Academies Press, Washington, DC (September 2017).https://doi.org/10.17226/24782

    work page doi:10.17226/24782 2017

  15. [15]

    arXiv preprint arXiv:2303.07240 (2023)

    Lin,W.,Zhao,Z.,Zhang,X.,Wu,C.,Zhang,Y.,Wang,Y.,Xie,W.:Pmc-clip:Con- trastive language-image pre-training using biomedical documents. arXiv preprint arXiv:2303.07240 (2023)

    arXiv 2023

  16. [16]

    The Lancet404(10452), 572–628 (August 2024).https://doi.org/10.1016/S0140-6736(24)01296-0

    Livingston, G., Huntley, J., Liu, K.Y., Costafreda, S.G., Selbæk, G., Alladi, S., Ames, D., Banerjee, S., Burns, A., Brayne, C., Fox, N.C., Ferri, C.P., Gitlin, L.N., Howard, R., Kales, H.C., Kivimäki, M., Larson, E.B., Nakasujja, N., Rockwood, K., Samus, Q., Shirai, K., Singh-Manoux, A., Schneider, L.S., Walsh, S., Yao, Y., Sommerlad, A., Mukadam, N.: De...

    work page doi:10.1016/s0140-6736(24)01296-0 2024

  17. [17]

    arXiv preprint arXiv:2405.11793 (2024)

    Wu, R., Zhang, C., Zhang, J., Zhou, Y., Zhou, T., Fu, H.: Mm-retinal: Knowledge-enhanced foundational pretraining with fundus image-text expertise. arXiv preprint arXiv:2405.11793 (2024)

    arXiv 2024

  18. [18]

    Journal of Neuroscience Nursing55(3), 103–109 (June 2023).https://doi.org/10.1097/JNN.0000000000000705

    Xiong,J.,Bhimani,R.,Carney-Anderson,L.:Reviewofriskfactorsassociatedwith biomarkers for alzheimer disease. Journal of Neuroscience Nursing55(3), 103–109 (June 2023).https://doi.org/10.1097/JNN.0000000000000705

    work page doi:10.1097/jnn.0000000000000705 2023

  19. [19]

    npj Digital Medicine5(1), 1–9 (December 2022).https://doi.org/10.1038/s41746-022-00742-2

    Yang, X., Chen, A., PourNejatian, N., Shin, H.C., Smith, K.E., Parisien, C., Com- pas, C., Martin, C., Costa, A.B., Flores, M.G., Zhang, Y., Magoc, T., Harle, C.A., Lipori, G., Mitchell, D.A., Hogan, W.R., Shenkman, E.A., Bian, J., Wu, Y.: A large language model for electronic health records. npj Digital Medicine5(1), 1–9 (December 2022).https://doi.org/1...

    work page doi:10.1038/s41746-022-00742-2 2022

  20. [20]

    arXiv preprint arXiv:2303.00915 (2025)

    Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., Wong, C., Tupini, A., Wang, Y., Mazzola, M., Shukla, S., Liden, L., Gao, J., Crabtree, A., Piening, B., Bifulco, C., Lungren, M.P., Naumann, T., Wang, S., Poon, H.: Biomedclip: A multimodal biomedical foundation model pretrained from scientific image-t...

    Pith/arXiv arXiv 2025

  21. [21]

    Nature622(7981), 156–163 (October 2023).https://doi.org/10.1038/ s41586-023-06555-x

    Zhou, Y., Chia, M.A., Wagner, S.K., Ayhan, M.S., Williamson, D.J., Struyven, R.R., Liu, T., Xu, M., Lozano, M.G., Woodward-Court, P., Kihara, Y., Alt- mann, A., Lee, A.Y., Topol, E.J., Denniston, A.K., Alexander, D.C., Keane, P.A.: A foundation model for generalizable disease detection from retinal im- ages. Nature622(7981), 156–163 (October 2023).https:/...

    2023

  22. [22]

    Translational Vision Science & Technology11(7), 12 (July 2022).https://doi.org/10.1167/tvst

    Zhou, Y., Wagner, S.K., Chia, M.A., Zhao, A., Woodward-Court, P., Xu, M., Struyven, R., Alexander, D.C., Keane, P.A.: Automorph: Automated retinal vas- cular morphology quantification via a deep learning pipeline. Translational Vision Science & Technology11(7), 12 (July 2022).https://doi.org/10.1167/tvst. 11.7.12,https://doi.org/10.1167/tvst.11.7.12

    work page doi:10.1167/tvst 2022