Pith. sign in

REVIEW 3 major objections 3 minor 45 references

AMD-Mamba: A Phenotype-Aware Multi-Modal Framework for Robust AMD Prognosis

T0 review · 3 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AMD-Mamba, a multi-modal framework that fuses fundus images with genetic and demographic data under a severity-score metric-learning objective, produces a new biomarker that the authors find is among the most significant predictors of AMD…

desk verdict The submission is a shell: the abstract describes AMD-Mamba but the full text is an unrelated materials science paper, so there is nothing to review. read the letter →

arxiv 2508.02957 v1 pith:5ENXROB3 submitted 2025-08-04 eess.IV cs.CV

Puzhen Wu , Mingquan Lin , Qingyu Chen , Emily Y. Chew , Zhiyong Lu , Yifan Peng , Hexin Dong This is my paper
classification eess.IVcs.CV
keywords age-relatedmaculardegenerationAMDprognosismulti-modaldeeplearningmetricVisionMambabiomarkerdiscoveryfundusimaginggeneticriskfactors
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes AMD-Mamba, a multi-modal framework that predicts progression of age-related macular degeneration (AMD) by combining color fundus images, 52 genetic variants, and 3 socio-demographic variables. The key idea is to use the clinical AMD severity scale as prior knowledge in a metric-learning objective, so that learned image features reflect the ordinal severity of the disease rather than just a binary progression label. The model uses a Vision Mamba backbone to capture both local lesions and long-range structural changes, and fuses image features with clinical variables at multiple scales. The authors report that the resulting feature representation acts as a new biomarker that is one of the most significant predictors of AMD progression in the AREDS cohort, and that adding it to existing variables improves detection of high-risk patients at early stages. If correct, this would support earlier and more proactive clinical management of AMD.

What carries the argument

The load-bearing mechanism is the metric-learning strategy that treats the AMD severity scale score as supervised prior knowledge: the model is trained so that distances between learned feature representations mirror the ordinal severity ordering, not just the binary progression label. This is implemented with a Vision Mamba backbone that combines local image details (such as drusen) with long-range global features (such as vascular changes), and a multi-scale fusion module that merges the image representation with genetic variants and socio-demographic variables. The extracted feature representation is itself the new biomarker.

What would settle it

Train AMD-Mamba on an independent AMD cohort with fundus images and genetics, remove the severity-score supervision (or use held-out severity labels), and check whether the extracted biomarker remains a top predictor of progression; if its predictive value drops sharply once the severity score is controlled for, the central claim would be weakened.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a phenotype-aware metric-learning constraint—aligning learned image features with the ordinal AMD severity score—yields a prognostic biomarker from fundus images that carries independent predictive signal for progression to advanced AMD. This biomarker, extracted from the multi-modal AMD-Mamba model, is claimed to rank among the most significant predictors of progression in the AREDS cohort, and its inclusion alongside traditional variables produces improved identification of high-risk patients at early disease stages. The authors interpret this as evidence that jointly modeling imaging, genetics, and clinical phenotype in one framework can reveal progression patterns that conventional CNN-based, locally focused models miss.

Load-bearing premise

The results depend on the AMD severity scale scores used as metric-learning supervision being accurate and not already encoding the progression outcome; if those scores are noisy or leak the target, the reported biomarker significance and early-detection gains could be artifacts.

Editorial extensions

If this is right

  • If the biomarker is valid, a single fundus image from an early AMD patient could be used to estimate progression risk well before advanced symptoms appear.
  • The success of the metric-learning objective suggests that ordinal clinical severity scales can serve as effective supervision for deep learning in other progressive diseases.
  • The Vision Mamba backbone's ability to capture long-range vascular changes could generalize to other retinal conditions in which global structure matters, not just AMD.
  • The multi-modal design implies that genetic and demographic information are complementary to imaging features, supporting integrated risk models in routine practice.
  • Improved early detection of high-risk patients could enable earlier intervention and better targeting of clinical trials for treatments that slow AMD progression.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's evaluation is confined to the AREDS cohort, so the natural next step is to test the biomarker on an independent AMD dataset to confirm the severity-score supervision does not overfit to AREDS-specific grading.
  • A direct comparison between this ordinal metric-learning objective and a model trained directly on the progression label would isolate how much of the reported gain comes from the severity prior rather than the multi-modal architecture itself.
  • An ablation that removes each modality (images, genetics, demographics) in turn would clarify whether the imaging-derived biomarker adds unique predictive signal beyond standard genetic and demographic risk factors.
  • The same architecture could likely be adapted to other progressive eye diseases, such as diabetic retinopathy, but the ordinal clinical scale used as supervision would need to be defined anew for each disease.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The manuscript, as submitted, consists of an abstract claiming a novel multi-modal framework (AMD-Mamba) for age-related macular degeneration prognosis — including a metric-learning strategy using AMD severity scale scores, fusion of fundus images with genetic and socio-demographic variables, and evaluation on the AREDS dataset — followed by a full text that is entirely unrelated: it is a materials science paper on autonomous inorganic materials discovery via multi-agent physics-aware reasoning. No methods, equations, experimental protocols, result tables, code, or data for AMD-Mamba appear anywhere in the submission. The central claim of the abstract is therefore unsupported by the supplied document.

Significance. If the AMD-Mamba framework and biomarker were actually developed and validated as described, the claimed improvements in early AMD risk detection could be clinically valuable. However, because the submission contains no trace of the AMD work — neither a technical description nor any experimental results — the significance of the claimed contribution cannot be assessed. There are no machine-checked proofs, reproducible code, parameter-free derivations, or falsifiable predictions to evaluate. The only verifiable content is the unrelated materials science paper, which does not bear on the AMD claims.

major comments (3)
  1. [Full text (all sections)] The full text of arXiv:2508.02957 is not the AMD-Mamba paper; it is the materials science manuscript 'Autonomous Inorganic Materials Discovery via Multi-Agent Physics-Aware Scientific Reasoning' (arXiv:2508.02956). The abstract is the only AMD-related content in the submission. There are no methods, equations, experimental protocols, result tables, code, or data for AMD-Mamba, so the central claim that the proposed biomarker is 'one of the most significant biomarkers for the progression of AMD' is entirely unsupported.
  2. [Abstract (metrics and baselines)] The abstract reports experimental evaluation on AREDS with 45,818 photographs, 52 genetic variants, and 3 socio-demographic variables from 2,741 subjects, but it provides no quantitative metrics, no error bars, no ablation studies, and no baseline comparisons. Even if the intended manuscript had been submitted, the abstract alone cannot substantiate the claimed improvements in detecting high-risk AMD patients at early stages.
  3. [Abstract (metric learning and evaluation)] The described metric-learning strategy uses the AMD severity scale score as prior knowledge, while the biomarker is apparently evaluated on the same AREDS cohort. If progression is defined through the same severity scale, part of the reported biomarker significance could reduce to the training signal. This circularity concern cannot be investigated because the training objective and progression definition are not described anywhere in the submission.
minor comments (3)
  1. [Abstract (clarity of claims)] The phrase 'one of the most significant biomarkers' is vague and non-quantitative; the authors should report the actual statistical significance, hazard ratios, or C-statistics with confidence intervals.
  2. [Abstract (dataset description)] The abstract states that the AREDS dataset includes 45,818 color fundus photographs from 2,741 subjects, but does not clarify the number of eyes, follow-up duration, or how the severity scale was graded; these details are needed for any future reproduction.
  3. [Full text (mismatch)] The submission title and abstract refer to AMD-Mamba, but the body is a completely different paper; this appears to be a manuscript assembly error that must be corrected before any scientific review can proceed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity claim can be established: the AMD-Mamba full text is absent, so no derivation chain exists to reduce.

full rationale

The submission contains only the abstract of the AMD-Mamba paper; the accompanying full text is an unrelated materials-science manuscript ('Autonomous Inorganic Materials Discovery via Multi-Agent Physics-Aware Scientific Reasoning'). As a result, the AMD-Mamba methods, equations, training objectives, and AREDS result tables are not present in the submitted text. The only potential circularity suggested by the abstract is that the metric-learning strategy uses AMD severity scale scores as prior knowledge while the derived biomarker is said to be significant for AMD progression; however, the abstract never states that progression is defined by the severity scale, and without the methods and evaluation sections there is no Eq. X = Eq. Y reduction or fitted-parameter-renamed-as-prediction step that can be exhibited. Per the hard requirement that circularity be shown by quotation and explicit reduction rather than speculation, no circular step can be established. The observed mismatch between abstract and full text is a completeness and support failure, not an established circularity, and therefore receives a score of 0 on the circularity scale.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

Because the AMD methods are absent from the submitted full text, the ledger is reconstructed from the abstract. There are no visible fitted parameters or equations to audit. The main assumptions concern the reliability of the clinical severity labels and the sufficiency of the multimodal inputs. The invented entity is the model-derived biomarker, which has no independent evidence in this submission.

assumptions (3)
  • domain assumption AMD severity scale score is a valid ordinal clinical phenotype suitable as metric-learning prior knowledge.
    Stated in the abstract's metric learning sentence; if the score is unreliable or inconsistently graded, the learned feature alignment may not reflect true disease progression.
  • domain assumption AREDS color fundus photographs, 52 genetic variants, and 3 socio-demographic variables contain sufficient signal to predict AMD progression.
    The abstract evaluates on these inputs but provides no feature analysis or baseline comparison in the submitted text.
  • ad hoc to paper Vision Mamba fusion of local and long-range image information improves prognosis over CNN backbones.
    Asserted in the abstract as a design advantage; no ablation study is present to support it.
invented entities (1)
  • New AMD progression biomarker derived from AMD-Mamba
    purpose: To flag high-risk AMD patients at early stages and improve prognosis when combined with existing variables.
    The abstract reports the biomarker is significant on AREDS, but no external validation, independent dataset, or prospective test is supplied; the submitted full text is an unrelated materials paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AMD-Mamba: A Phenotype-Aware Multi-Modal Framework for Robust AMD Prognosis." pith.science (2026). https://pith.science/paper/5ENXROB3

@misc{pith2026250802957,
  author       = {Pith},
  title        = {Pith review of: AMD-Mamba: A Phenotype-Aware Multi-Modal Framework for Robust AMD Prognosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ENXROB3}},
  note         = {Machine review of arXiv:2508.02957}
}
read the original abstract

Age-related macular degeneration (AMD) is a leading cause of irreversible vision loss, making effective prognosis crucial for timely intervention. In this work, we propose AMD-Mamba, a novel multi-modal framework for AMD prognosis, and further develop a new AMD biomarker. This framework integrates color fundus images with genetic variants and socio-demographic variables. At its core, AMD-Mamba introduces an innovative metric learning strategy that leverages AMD severity scale score as prior knowledge. This strategy allows the model to learn richer feature representations by aligning learned features with clinical phenotypes, thereby improving the capability of conventional prognosis methods in capturing disease progression patterns. In addition, unlike existing models that use traditional CNN backbones and focus primarily on local information, such as the presence of drusen, AMD-Mamba applies Vision Mamba and simultaneously fuses local and long-range global information, such as vascular changes. Furthermore, we enhance prediction performance through multi-scale fusion, combining image information with clinical variables at different resolutions. We evaluate AMD-Mamba on the AREDS dataset, which includes 45,818 color fundus photographs, 52 genetic variants, and 3 socio-demographic variables from 2,741 subjects. Our experimental results demonstrate that our proposed biomarker is one of the most significant biomarkers for the progression of AMD. Notably, combining this biomarker with other existing variables yields promising improvements in detecting high-risk AMD patients at early stages. These findings highlight the potential of our multi-modal framework to facilitate more precise and proactive management of AMD.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 33 canonical work pages

  1. [4]

    Theory-guided design of catalytic materials using scaling relationships and reactivity descriptors

    Zhao, Z.-J.et al. Theory-guided design of catalytic materials using scaling relationships and reactivity descriptors. Nature Reviews Materials4, 792–804 (2019)

    work page 2019

  2. [5]

    & Archer, L

    Zhao, Q., Stalin, S., Zhao, C.-Z. & Archer, L. A. Designing solid-state electrolytes for safe, energy-dense batteries. Nature Reviews Materials5, 229–252 (2020)

    work page 2020

  3. [6]

    E., Kirklin, S., Aykol, M., Meredig, B

    Saal, J. E., Kirklin, S., Aykol, M., Meredig, B. & Wolverton, C. Materials design and discovery with high-throughput density functional theory: the open quantum materials database (oqmd).Jom 65, 1501–1509 (2013)

    work page 2013

  4. [7]

    & Buehler, M

    Guo, K., Yang, Z., Yu, C.-H. & Buehler, M. J. Artificial intelligence and machine learning in design of mechanical materials. Materials Horizons8, 1153–1172 (2021)

    work page 2021

  5. [8]

    Graph neural networks for materials science and chemistry.Communications Materials 3, 93 (2022)

    Reiser, P.et al. Graph neural networks for materials science and chemistry.Communications Materials 3, 93 (2022)

    work page 2022

  6. [9]

    Recent advances and applications of deep learning methods in materials science

    Choudhary, K.et al. Recent advances and applications of deep learning methods in materials science. npj Computational Materials8, 59 (2022)

    work page 2022

  7. [10]

    Scaling deep learning for materials discovery.Nature 624, 80–85 (2023)

    Merchant, A.et al. Scaling deep learning for materials discovery.Nature 624, 80–85 (2023)

    work page 2023

  8. [11]

    Jain, A. et al. Commentary: The materials project: A materials genome approach to accelerating materials innovation.APL materials1 (2013)

    work page 2013

Show all 45 references
  1. [12]

    The open quantum materials database (oqmd): assessing the accuracy of dft formation energies

    Kirklin, S.et al. The open quantum materials database (oqmd): assessing the accuracy of dft formation energies. npj Computational Materials1, 1–15 (2015)

    2015

  2. [13]

    Curtarolo, S. et al. Aflow: An automatic framework for high-throughput materials discovery. Computational Materials Science58, 218–226 (2012)

    2012

  3. [14]

    & Grossman, J

    Xie, T. & Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Physical review letters120, 145301 (2018)

    2018

  4. [15]

    Park, C. W. & Wolverton, C. Developing an improved crystal graph convolutional neural network framework for accelerated materials discovery.Physical Review Materials4, 063801 (2020)

    2020

  5. [16]

    & Jaakkola, T

    Xie, T., Fu, X., Ganea, O.-E., Barzilay, R. & Jaakkola, T. Crystal diffusion variational autoencoder for periodic material generation.arXiv preprint arXiv:2110.06197 (2021)

    2021 arXiv

  6. [17]

    An invertible crystallographic representation for general inverse design of inorganic crystals with targeted properties.Matter 5, 314–335 (2022)

    Ren, Z.et al. An invertible crystallographic representation for general inverse design of inorganic crystals with targeted properties.Matter 5, 314–335 (2022)

    2022

  7. [18]

    Physics guided deep learning for generative design of crystal materials with symmetry constraints

    Zhao, Y.et al. Physics guided deep learning for generative design of crystal materials with symmetry constraints. npj Computational Materials9, 38 (2023)

    2023

  8. [19]

    A generative model for inorganic materials design.Nature 1–3 (2025)

    Zeni, C.et al. A generative model for inorganic materials design.Nature 1–3 (2025)

    2025

  9. [20]

    & Ong, S

    Chen, C. & Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science2, 718–728 (2022)

    2022

  10. [21]

    Mace: Higherorderequivariantmessagepass- ing neural networks for fast and accurate force fields.Advances in neural information processing systems 35, 11423–11436 (2022)

    Batatia, I., Kovacs, D.P., Simm, G., Ortner, C.&Csányi, G. Mace: Higherorderequivariantmessagepass- ing neural networks for fast and accurate force fields.Advances in neural information processing systems 35, 11423–11436 (2022)

    2022

  11. [22]

    Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling

    Deng, B.et al. Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence5, 1031–1041 (2023)

    2023

  12. [23]

    A foundation model for atomistic materials chemistry.arXiv preprint arXiv:2401.00096 (2023)

    Batatia, I.et al. A foundation model for atomistic materials chemistry.arXiv preprint arXiv:2401.00096 (2023)

    2023 arXiv

  13. [24]

    Barroso-Luque, L. et al. Open materials 2024 (omat24) inorganic materials dataset and models. arXiv preprint arXiv:2410.12771 (2024)

    2024 arXiv

  14. [25]

    Mattersim: A deep learning atomistic model across elements, temperatures and pressures

    Yang, H.et al. Mattersim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967 (2024)

    2024 arXiv

  15. [26]

    Emergent abilities of large language models.arXiv preprint arXiv:2206.07682 (2022)

    Wei, J.et al. Emergent abilities of large language models.arXiv preprint arXiv:2206.07682 (2022)

    2022 arXiv

  16. [27]

    Sparks of artificial general intelligence: Early experiments with GPT-4 (2023)

    Bubeck, S.et al. Sparks of artificial general intelligence: Early experiments with GPT-4 (2023). URL https://arxiv.org/abs/2303.12712. 2303.12712

    2023 arXiv

  17. [28]

    Buehler, M. J. Generative retrieval-augmented ontologic graph and multiagent strategies for interpretive large language model-based materials design.ACS Engineering Au4, 241–277 (2024)

    2024

  18. [29]

    K., Knowles, T

    Yang, Z., Yorke, S. K., Knowles, T. P. & Buehler, M. J. Learning the rules of peptide self-assembly through data mining with large language models.Science Advances11, eadv1971 (2025). 12 Autonomous Inorganic Materials Discovery via Multi-Agent Physics-Aware Scientific Reasoning

    2025

  19. [30]

    Buehler, M. J. Accelerating scientific discovery with generative knowledge extraction, graph-based representation, and multimodal intelligent graph reasoning.Machine Learning: Science and Technology (2024). URL http://iopscience.iop.org/article/10.1088/2632-2153/ad7228

    2024 doi

  20. [31]

    & Buehler, M

    Ghafarollahi, A. & Buehler, M. J. Sciagents: Automating scientific discovery through bioinspired multi-agent intelligent graph reasoning.Advanced Materials 2413523 (2024)

    2024

  21. [32]

    Ni, B., Kaplan, D. L. & Buehler, M. J. Forcegen: End-to-end de novo protein generation based on nonlinear mechanical unfolding responses using a language diffusion model.Science Advances10, eadl4000 (2024)

    2024

  22. [33]

    Buehler, M. J. Melm, a generative pretrained language modeling framework that solves forward and inverse mechanics problems.Journal of the Mechanics and Physics of Solids181, 105454 (2023)

    2023

  23. [34]

    Buehler, M. J. Cephalo: Multi-modal vision-language models for bio-inspired materials analysis and design. Advanced Functional Materials34, 2409531 (2024)

    2024

  24. [35]

    Large language models for material property predictions: elastic constant tensor prediction and materials design.Digital Discovery4, 1625–1638 (2025)

    Liu, S.et al. Large language models for material property predictions: elastic constant tensor prediction and materials design.Digital Discovery4, 1625–1638 (2025)

    2025

  25. [36]

    & Ceder, G

    Hafner, J., Wolverton, C. & Ceder, G. Toward computational materials design: the impact of density functional theory on materials research.MRS bulletin31, 659–668 (2006)

    2006

  26. [37]

    & Persson, K

    Jain, A., Shin, Y. & Persson, K. A. Computational predictions of energy materials using density functional theory.Nature Reviews Materials1, 1–13 (2016)

    2016

  27. [38]

    Szymanski, N. J.et al. An autonomous laboratory for the accelerated synthesis of novel materials. Nature 624, 86–91 (2023)

    2023

  28. [39]

    Self-driving laboratories for chemistry and materials science.Chemical Reviews124, 9633–9732 (2024)

    Tom, G.et al. Self-driving laboratories for chemistry and materials science.Chemical Reviews124, 9633–9732 (2024)

    2024

  29. [40]

    Lu, C. et al. The AI scientist: Towards fully automated open-ended scientific discovery. arXiv preprint arXiv:2408.06292 (2024)

    2024 arXiv

  30. [41]

    Bran, A.et al

    M. Bran, A.et al. Augmenting large language models with chemistry tools.Nature Machine Intelligence 6, 525–535 (2024)

    2024

  31. [42]

    & Buehler, M

    Ghafarollahi, A. & Buehler, M. J. ProtAgents: protein discovery via large language model multi-agent collaborations combining physics and machine learning.Digital Discovery3, 1389–1409 (2024)

    2024

  32. [43]

    & Buehler, M

    Ghafarollahi, A. & Buehler, M. J. Sparks: Multi-agent artificial intelligence model discovers protein design principles. arXiv preprint arXiv:2504.19017 (2025)

    2025 arXiv

  33. [44]

    & Buehler, M

    Ghafarollahi, A. & Buehler, M. J. Automating alloy design and discovery with physics-aware multimodal multiagent ai.Proceedings of the National Academy of Sciences122, e2414074122 (2025)

    2025

  34. [45]

    Horton, M. K. et al. Accelerated data-driven materials science with the materials project. Nature Materials 1–11 (2025)

    2025

  35. [46]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    Achiam, J.et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    2023 arXiv

  36. [47]

    & the AG2 Community

    Wang, C., Wu, Q. & the AG2 Community. Ag2: Open-source agentos for ai agents (2024). URL https://github.com/ag2ai/ag2. Available at https://docs.ag2.ai/

    2024

  37. [48]

    Mattergen implementation (2024)

    CZeni, C.et al. Mattergen implementation (2024). URLhttps://github.com/microsoft/mattergen. Available at https://github.com/microsoft/mattergen. 13 AI-generated document for Task 1 AI-generated document for Task 1 AI-generated document for Task 1 AI-generated document for Task...

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.