Pith. sign in

REVIEW 2 major objections 6 minor 28 references

Retinal image graphs rank which diabetes pathways drive eye disease

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-07 23:32 UTC pith:PQTSLTH5

load-bearing objection The stress-test finding is correct and is the central problem: S_X1234 is effectively constant (~1.0) across all pathways, so the cross-cohort priority score reduces to NHANES statistics alone. The retinal image component — the paper's headline contribution — does not actually participate in the ranking. the 2 major comments →

arxiv 2607.05204 v1 pith:PQTSLTH5 submitted 2026-07-06 cs.CV math.PRstat.CO

Causal-RetiGraph: Cross-Cohort Retinal Support and Same-Subject Pathway Analysis for Diabetic Retinopathy

Inam Ullah , Imran Razzak , Shoaib Jameel This is my paper
classification cs.CV math.PRstat.CO
keywords diabetic retinopathyretinal graph phenotypepathway prioritisationmediation analysisoculomicsNHANEScausal inferencevessel morphology
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces Causal-RetiGraph, a two-part framework that connects retinal image analysis with population health data to identify which systemic disease pathways matter most for diabetic retinopathy. The first part builds a retinal graph phenotype called X1234 from four evidence streams — vessel structure, lesion evidence, image embeddings, and vascular biomarkers — fused through a spatial branch (capturing where lesions sit relative to vessels) and a Jacobian branch (capturing how biomarker-sensitive the image embeddings are). This phenotype achieves 0.9055 binary DR accuracy and 0.9711 AUROC on held-out data. The second part anchors systemic analysis in NHANES survey data, where the same participants have both clinical measurements and retinal variables. The framework then combines the image-derived X1234 support (from an external retinal cohort) with NHANES exposure–DR associations to rank candidate disease pathways, while a separate same-subject retinal mediator family R* provides participant-level pathway summaries. The central design principle is that X1234 and R* are deliberately kept distinct: X1234 provides cross-cohort retinal support for pathway prioritisation, while R* provides same-subject mediation-style estimates. The paper finds that glycaemic (HbA1c, fasting glucose), renal (urine albumin, serum creatinine), and haemodynamic (pulse pressure, systolic blood pressure) axes dominate across both folds, while inflammatory pathways show weaker signals.

Core claim

The paper's central discovery is that a retinal graph phenotype combining lesion–vessel spatial geometry with embedding–biomarker sensitivity can serve as a structured intermediate layer between local retinal injury and systemic diabetic microvascular stress. When used to prioritise systemic pathways, this phenotype consistently elevates glycaemic–renal and glycaemic–haemodynamic pathway families above inflammatory ones, a pattern that replicates in the independent NHANES same-subject mediation analysis using R*. The recurrence of these three axes across two methodologically distinct folds — one image-derived and cross-cohort, one participant-level and same-subject — is the paper's main evid

What carries the argument

The X1234 retinal graph phenotype, constructed by fusing a spatial X12 branch (vessel maps × lesion evidence via pixel-wise interaction channels) with a Jacobian X34 branch (a differentiable mapper from image embeddings to biomarker representations, whose Jacobian matrix captures embedding–biomarker sensitivity). The cross-cohort prioritisation score P_j = norm{|β̂_j| × S_X1234_j × [−log₁₀(q_j)]} combines NHANES exposure–DR association strength with image-derived retinal support. The same-subject NHANES retinal mediator family R* = Rstar_family_pca1 enables participant-level mediation-style summaries.

Load-bearing premise

The cross-cohort prioritisation assumes that the retinal support score S_X1234, learned on an external retinal image cohort (APTOS), is informative about systemic pathways in NHANES participants whose retinal images were never processed through X1234. If the image-derived support does not transfer across these populations, the pathway ranking collapses to NHANES association strength alone.

What would settle it

Compute X1234 directly on NHANES-linked retinal images and compare the resulting E→X1234→Y mediation estimates against the cross-cohort prioritisation ranking. If the rankings disagree substantially, the cross-cohort support score is not capturing transferable pathway-relevant retinal structure.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If X1234 generalises across cohorts, retinal image phenotypes could serve as a screening layer to prioritise which systemic disease pathways deserve expensive population-level mediation studies.
  • The glycaemic–renal and glycaemic–haemodynamic dominance suggests that anti-inflammatory interventions alone may be insufficient for DR prevention compared to glycaemic and blood-pressure control.
  • The Jacobian X34 branch — linking image embeddings to biomarker sensitivity — could be applied to other imaging biomarker pairs beyond retinal vascular morphology.
  • The explicit separation of cross-cohort image support from same-subject mediation provides a template for other domains where image-derived phenotypes and population health data come from different cohorts.

Where Pith is reading between the lines

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

  • The framework's value hinges on whether the X1234 retinal support score S_X1234_j actually captures pathway-relevant information that transfers from APTOS to NHANES. If S_X1234_j is roughly uniform across pathway families, the prioritisation score reduces to ranking by NHANES association strength alone, and the retinal image component adds no information.
  • The framework could be directly tested by computing X1234 on NHANES-linked retinal images (which exist for 2005–2008), enabling a true E→X1234→Y mediation analysis and removing the cross-cohort bridging assumption entirely.
  • The weak inflammatory pathway signals in R* may reflect NHANES variable availability rather than biological irrelevance; CRP is a coarse inflammatory marker compared to cytokine panels.
  • The spatial X12 branch's lesion–vessel interaction channels (L⊙A, L⊙R, L⊙V) could be decomposed further to test whether lesion proximity to arteries versus veins carries different prognostic meaning for systemic pathways.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper introduces Causal-RetiGraph, a two-fold framework for diabetic retinopathy (DR) analysis. The first fold constructs an interpretable retinal graph phenotype X1234 from four evidence streams (vessel maps, lesion evidence, image embeddings, AutoMorph biomarkers) via spatial (X12) and Jacobian (X34) branches, evaluated on APTOS 2019. The second fold uses NHANES 2005–2008 diabetic participants to estimate systemic exposure–DR associations (logistic regression with FDR correction), perform cross-cohort pathway prioritisation combining NHANES statistics with X1234-based retinal support (Eq. 7), and produce participant-level pathway summaries using a same-subject retinal mediator family R*. The paper identifies glycaemic, renal, and haemodynamic axes as dominant across both folds.

Significance. The paper's conceptual separation of X1234 (image-derived, cross-cohort support) from R* (same-subject NHANES mediator) is a thoughtful design choice that avoids overclaiming mediation from externally derived phenotypes. The retinal fold evaluation on APTOS is standard and the results (0.9711 AUROC binary, 0.8312 QWK graded) are credible. The NHANES association analysis with FDR correction and the participant-level R* pathway summaries (Fig. 4) provide useful epidemiological context. However, the paper's headline contribution — linking retinal image evidence to systemic pathway prioritisation via the cross-cohort priority score (Eq. 7) — is not demonstrated to function as claimed, as detailed below.

major comments (2)
  1. §III-G, Eq. (7): The cross-cohort priority score P_j = norm{|β̂_j| × S_X1234_j × [-log10(q_j)]} is the paper's central novel contribution, but S_X1234_j is never formally defined. The text only describes it qualitatively as 'reflecting whether the retinal graph phenotype contains lesion–vascular or biomarker-sensitive evidence consistent with the pathway family.' Without an operational definition, the score is not reproducible.
  2. Table IV: Back-computing from the reported values shows that S_X1234_j ≈ 1.0 for every pathway, meaning the retinal support component is vacuous to the ranking. For example: HbA1c gives |0.7260| × [-log10(1.16e-14)] = 0.7260 × 13.935 = 10.117, matching the reported priority of 10.1186. Pulse pressure: 0.4533 × 5.541 = 2.512 vs. 2.5114. Fasting glucose: 0.4425 × 3.204 = 1.418 vs. 1.4179. Urine albumin: 0.6062 × 2.095 = 1.270 vs. 1.2701. SBP: 0.2993 × 3.204 = 0.959 vs. 0.9588. In every case, the priority score equals |β̂_j| × [-log10(q_j)] to within rounding, so S_X1234_j ≡ 1. If S_X1234 is constant across pathways, the priority order is entirely determined by NHANES β and FDR values, and the 'retinal support' contributes nothing to the ranking. This directly undermines the paper's central claim that 'retinal graph phenotypes can help prioritise systemic pathways in DR.' The authors must (
minor comments (6)
  1. §III-G, Eq. (7): The 'norm' operator is unspecified. Is it min-max normalisation, z-score, or something else? This should be stated explicitly.
  2. Table I: The entry for E reads 'ENHANES' — appears to be a typo for 'E NHANES'.
  3. §III-H, Eq. (8): R* = Rstar_family_pca1 is introduced without explaining what variables comprise the 'Rstar family' before PCA. The reader cannot assess construct validity without knowing the input variables.
  4. Fig. 4: The pathway labels use NHANES variable codes (LBXGH, LBXGLU, URXUMA, etc.) without a legend mapping codes to clinical variables. A lookup table would improve readability.
  5. §VI: The fourth limitation mentions 'the nonlinear smooth branch is exploratory because the linear branch currently gives the more stable pattern,' but no nonlinear smooth branch results are reported in the main text. Either report these results or remove the limitation.
  6. §III-B: The curated APTOS subset of 2,910 images from the original ~3,662 should note the exclusion criteria.

Simulated Author's Rebuttal

2 responses · 0 unresolved

The referee raises two major comments, both concerning the cross-cohort priority score P_j in Eq. (7): (1) S_X1234_j is not formally defined, and (2) back-computation from Table IV shows S_X1234_j ≈ 1.0 for all pathways, making the retinal support component vacuous. We acknowledge that both points are correct. The manuscript currently lacks an operational definition of S_X1234_j, and the reported priority scores are effectively determined by NHANES β and FDR values alone. We will revise the manuscript to provide an explicit formula for S_X1234_j, recompute priority scores with pathway-varying values, and add a sensitivity analysis showing the ranking's dependence on retinal support. If pathway-varying S_X1234_j values do not change the ranking, we will honestly state that the current retinal support acts as a binary plausibility filter rather than a ranking differentiator, and adjust claims accordingly.

read point-by-point responses

  1. Referee: §III-G, Eq. (7): The cross-cohort priority score P_j = norm{|β̂_j| × S_X1234_j × [-log10(q_j)]} is the paper's central novel contribution, but S_X1234_j is never formally defined. The text only describes it qualitatively as 'reflecting whether the retinal graph phenotype contains lesion–vascular or biomarker-sensitive evidence consistent with the pathway family.' Without an operational definition, the score is not reproducible.

    Authors: The referee is correct. S_X1234_j is not formally defined in the current manuscript; it is described only qualitatively. This is a genuine gap that must be addressed for reproducibility. We will add an explicit operational definition in the revised manuscript. Specifically, we will define S_X1234_j as a composite score computed from the X1234 phenotype evidence: (i) the FDR-corrected grade-trend association strength for lesion–vascular features (from Fig. 2a) mapped to each pathway family, and (ii) the FDR-corrected lesion–biomarker coupling strength (from Fig. 2b) for biomarkers relevant to each pathway family. The formula, mapping table from X1234 features to pathway families, and the resulting per-pathway S_X1234_j values will be reported in full. revision: yes

  2. Referee: Table IV: Back-computing from the reported values shows that S_X1234_j ≈ 1.0 for every pathway, meaning the retinal support component is vacuous to the ranking. [...] In every case, the priority score equals |β̂_j| × [-log10(q_j)] to within rounding, so S_X1234_j ≡ 1. If S_X1234 is constant across pathways, the priority order is entirely determined by NHANES β and FDR values, and the 'retinal support' contributes nothing to the ranking. This directly undermines the paper's central claim that 'retinal graph phenotypes can help prioritise systemic pathways in DR.'

    Authors: The referee's back-computation is correct. In the current manuscript, S_X1234_j was effectively set to 1.0 for all pathways, meaning the priority ranking in Table IV is entirely determined by NHANES β and FDR values. We acknowledge this honestly. The retinal support component as currently implemented does not differentiate between pathways in the ranking. This is a substantive issue that the revision must address. We will take the following steps: (1) Compute pathway-varying S_X1234_j values using the operational definition described above, derived from the grade-trend and lesion–biomarker association evidence in Fig. 2. (2) Recompute Table IV with these non-constant S_X1234_j values and report whether the ranking changes. (3) Add a sensitivity analysis comparing rankings with and without retinal support. (4) If the ranking does not change meaningfully, we will honestly state that X1234 currently functions as a binary plausibility filter (confirming that all five pathways have retinal evidence) rather than a ranking differentiator, and we will revise the paper's claims accordingly. The abstract and conclusion will be modified to accurately reflect what the retinal support does and does not contribute. We will not claim that retinal graph phenotypes 'help prioritise' pathways if the data show that the ranking is driven entirely by NHANES statistics. revision: yes

Circularity Check

1 steps flagged

The cross-cohort priority score (Eq. 7) reduces to NHANES statistics alone because S_X1234 is effectively constant (~1.0) across all pathways, making the retinal support component vacuous to the ranking.

specific steps
  1. renaming known result [Eq. 7 (§III-G) and Table IV]
    "Pj = norm{|β̂j| × S_X1234_j × [−log10(qj)]} ... where β̂j is the NHANES exposure–DR coefficient, qj is the FDR-adjusted association value, and S_X1234_j is the retinal support score derived from the externally evaluated image-based phenotype. The support score reflects whether the retinal graph phenotype contains lesion–vascular or biomarker-sensitive evidence consistent with the pathway family."

    The skeptic attack demonstrates by back-computation from Table IV that S_X1234_j ≈ 1.0 for every pathway. For HbA1c: |0.7260| × [-log10(1.16e-14)] = 0.7260 × 13.935 = 10.117, matching the reported priority of 10.1186. For pulse pressure: 0.4533 × 5.542 = 2.512 vs 2.5114. For fasting glucose: 0.4425 × 3.204 = 1.418 vs 1.4179. For urine albumin: 0.6062 × 2.095 = 1.270 vs 1.2701. For SBP: 0.2993 × 3.204 = 0.959 vs 0.9588. In every case, the priority score equals |β̂j| × [-log10(qj)] to within rounding, meaning S_X1234_j ≡ 1. If S_X1234 is constant across pathways, it contributes nothing to the ranking — the priority order is entirely determined by NHANES β and FDR values. The paper's headline contribution — 'retinal graph phenotypes can help prioritise systemic pathways in DR' — is not in the

full rationale

The paper's central claim is that retinal graph phenotypes (X1234) provide 'structured retinal support for cross-cohort pathway prioritisation.' This is operationalised through Eq. 7, where S_X1234_j is supposed to encode pathway-specific retinal evidence. However, back-computation from Table IV reveals S_X1234_j is effectively 1.0 for all five pathways, meaning the priority ranking is entirely determined by NHANES association statistics (β̂j and qj). The retinal support component is vacuous to the ranking. The paper never specifies how S_X1234_j is computed numerically — §III-G only describes it qualitatively as 'reflecting whether the retinal graph phenotype contains lesion–vascular or biomarker-sensitive evidence consistent with the pathway family.' This is a renaming of a known result: the NHANES association ranking is relabelled as a 'cross-cohort retinal-supported priority' without the retinal component contributing to the ordering. The R* participant-level mediation analysis (Fig. 4) does remain independent of this issue, but the cross-cohort prioritisation — the paper's headline contribution — reduces to NHANES statistics alone.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 3 invented entities

The ledger captures the key fitted parameters (attention weights, mapper parameters) and underspecified constructs (S_X1234, R*). The axioms highlight the reliance on Grad-CAM as lesion evidence and the cross-cohort generalisation assumption. The invented entities include the core X1234 phenotype (which has independent evaluation) and the priority score P_j and R* (which lack independent validation).

free parameters (4)
  • Attention weights alpha_12, alpha_34 = Learned
    Fitted during training of the fused phenotype (Eq. 5).
  • S_X1234 (retinal support score) = Not specified
    Used in Eq. 7 for cross-cohort prioritisation but its exact derivation from X1234 is not specified.
  • Rstar_family_pca1 = Not specified
    The NHANES retinal mediator family R* is defined as this PCA component, but the input variables and PCA construction are not detailed.
  • Mapper f_theta parameters = Learned
    The differentiable mapper from embedding to biomarker representation (Eq. 3) is learned.
axioms (3)
  • domain assumption Grad-CAM heatmaps provide meaningful lesion evidence for DR.
    The paper uses Grad-CAM as X2 lesion evidence, acknowledging it is 'weak DR-XAI-style evidence rather than expert lesion segmentation' (Section VI).
  • ad hoc to paper Retinal features derived from APTOS generalize to NHANES participants.
    The cross-cohort prioritisation (Step 7) assumes S_X1234 from APTOS is relevant to NHANES pathways despite no same-subject data.
  • domain assumption Cross-sectional NHANES data can support mediation-style summaries.
    The participant-level R* pathway summaries (Section III.H) use cross-sectional data for TE/NDE/NIE estimation, which the paper acknowledges as a limitation.
invented entities (3)
  • X1234 retinal graph phenotype independent evidence
    purpose: Fused retinal representation for DR grading and pathway support.
    Evaluated on held-out APTOS test set (Table II, III) with falsifiable performance metrics.
  • Cross-cohort priority score P_j no independent evidence
    purpose: Ranking candidate exposure-retina-DR pathways.
    Defined in Eq. 7 as a heuristic combination; no external validation that this score ranks pathways correctly.
  • R* (Rstar_family_pca1) no independent evidence
    purpose: Same-subject NHANES retinal mediator family for pathway summaries.
    Construction is underspecified; no independent validation that this PCA component captures retinal mediation.

pith-pipeline@v1.1.0-glm · 16413 in / 2308 out tokens · 147046 ms · 2026-07-07T23:32:44.132492+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Causal-RetiGraph: Cross-Cohort Retinal Support and Same-Subject Pathway Analysis for Diabetic Retinopathy." pith.science (2026). https://pith.science/paper/PQTSLTH5

@misc{pith2026260705204,
  author       = {Pith},
  title        = {Pith review of: Causal-RetiGraph: Cross-Cohort Retinal Support and Same-Subject Pathway Analysis for Diabetic Retinopathy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PQTSLTH5}},
  note         = {Machine review of arXiv:2607.05204}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Diabetic retinopathy (DR) is a local retinal lesion process and a visible manifestation of systemic microvascular injury. Modern retinal AI can grade images accurately, but often leaves unanswered how local lesion evidence, retinal vascular structure, and systemic disease pathways are connected. This paper introduces \emph{Causal-RetiGraph}, a compact biomedical informatics framework that links retinal graph phenotypes with NHANES-anchored pathway modelling. The retinal-image fold constructs an interpretable $X1234$ phenotype from vessel maps, lesion evidence, image embeddings, and AutoMorph biomarkers through spatial $X_{12}$ and Jacobian $X_{34}$ branches. The NHANES fold models systemic exposures, covariates, a same-subject retinal mediator family $R^*$, and downstream outcome families. $X1234$ is used for retinal support and pathway prioritisation, while $R^*$ is used for participant-level pathway summaries. On the retinal fold, $X1234$ achieves 0.9055 binary DR accuracy and 0.9711 AUROC, with graded DR QWK of 0.8312. The results show that lesion and biomarker streams improve contextual retinal representation under scarce and imbalanced data. In NHANES, HbA1c, urine albumin, pulse pressure, fasting glucose, and systolic blood pressure are the strongest binary DR anchors. Participant-level pathway analysis identifies glycaemic--renal and glycaemic--haemodynamic pathways as the clearest mediator-style signals. These results suggest that retinal graph phenotypes can help prioritise systemic pathways in DR while preserving the distinction between image-derived support and same-subject mediation.

Figures

Figures reproduced from arXiv: 2607.05204 by Imran Razzak, Inam Ullah, Shoaib Jameel.

Figure 1
Figure 1. Figure 1: Causal-RetiGraph workflow. Steps 1–4 construct and evaluate the image-derived retinal graph phenotype [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Retinal support evidence for X1234. (a) FDR-corrected grade-trend analysis showing severity-sensitive graph, lesion and biomarker components across DR grades. Positive associations indicate features increasing with DR severity, while negative associations indicate features decreasing with severity. (b) FDR-corrected lesion–biomarker associations showing that weakly supervised lesion evidence is coupled wit… view at source ↗
Figure 3
Figure 3. Figure 3: NHANES binary DR association anchor in the 884-row diabetes subset. Model 1 is crude, Model 2 adjusts for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Participant-level NHANES pathway summaries split by outcome family. The mediator is the same-subject NHANES [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

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

28 extracted references · 28 canonical work pages · 2 internal anchors

  1. [1]

    Diabetes mellitus: Classification, mediators, and complications; a gate to identify potential targets for the development of new effective treatments,

    S. A. Antar, N. A. Ashour, M. Sharaky, M. Khattab, N. A. Ashour, R. T. Zaid, E. J. Roh, A. Elkamhawy, and A. A. Al-Karmalawy, “Diabetes mellitus: Classification, mediators, and complications; a gate to identify potential targets for the development of new effective treatments,” Biomedicine & Pharmacotherapy, vol. 168, p. 115734, 2023

    work page 2023

  2. [2]

    Dia- betes mellitus as a multisystem disease: understanding subtypes, complications, and the link with steatotic liver diseases in humans,

    A. Giannakogeorgou, M. Roden, and K. Pafili, “Dia- betes mellitus as a multisystem disease: understanding subtypes, complications, and the link with steatotic liver diseases in humans,”Hormones, vol. 25, no. 1, pp. 61– 80, 2026

    work page 2026

  3. [3]

    Diabetes,

    World Health Organization, “Diabetes,” https: //www.who.int/news-room/fact-sheets/detail/diabetes, Nov. 2024, fact sheet. Accessed 2026-07-03

    work page 2024

  4. [4]

    Vascular complications of diabetes: A narrative review,

    Y . Lu, W. Wang, J. Liu, M. Xie, Q. Liu, and S. Li, “Vascular complications of diabetes: A narrative review,” Medicine, vol. 102, no. 40, p. e35285, 2023

    work page 2023

  5. [5]

    Cardiovascular complications of diabetes: From mi- crovascular to macrovascular pathways,

    M. Zakir, N. Ahuja, M. A. Surksha, R. Sachdev, Y . Kalariya, M. Nasir, M. Kashif, F. Shahzeen, A. Tayyab, M. S. M. Khan, M. Junejo, F. M. Kumar, G. Varrassi, S. Kumar, M. Khatri, and T. Mohamad, “Cardiovascular complications of diabetes: From mi- crovascular to macrovascular pathways,”Cureus, vol. 15, no. 9, p. e45835, 2023

    work page 2023

  6. [6]

    Un- derstanding the clinical relationship between diabetic retinopathy, nephropathy, and neuropathy: A comprehen- sive review,

    A. Kulkarni, A. R. Thool, and S. Daigavane, “Un- derstanding the clinical relationship between diabetic retinopathy, nephropathy, and neuropathy: A comprehen- sive review,”Cureus, vol. 16, no. 3, p. e56674, 2024

    work page 2024

  7. [7]

    The pathophysiological mechanisms underlying diabetic retinopathy,

    L. Wei, X. Sun, C. Fan, R. Li, S. Zhou, and H. Yu, “The pathophysiological mechanisms underlying diabetic retinopathy,”Frontiers in Cell and Developmental Biol- ogy, vol. 10, p. 963615, 2022

    work page 2022

  8. [8]

    Oculomics: Current concepts and evidence,

    Z. Zhu, Y . Wang, Z. Qi, W. Hu, X. Zhang, S. K. Wagner, Y . Wang, A. R. Ran, J. Ong, E. Waisberget al., “Oculomics: Current concepts and evidence,”Progress in Retinal and Eye Research, vol. 106, p. 101350, 2025

    work page 2025

  9. [9]

    Stan- dardization and clinical applications of retinal imag- ing biomarkers for cardiovascular disease: A roadmap from an NHLBI workshop,

    E. Y . Chew, S. A. Burns, A. G. Abrahamet al., “Stan- dardization and clinical applications of retinal imag- ing biomarkers for cardiovascular disease: A roadmap from an NHLBI workshop,”Nature Reviews Cardiology, vol. 22, no. 1, pp. 47–63, 2025

    work page 2025

  10. [10]

    Advances in structural and functional retinal imaging and biomarkers for early detection of diabetic retinopathy,

    Z. Zhang, C. Deng, and Y . M. Paulus, “Advances in structural and functional retinal imaging and biomarkers for early detection of diabetic retinopathy,”Biomedicines, vol. 12, no. 7, p. 1405, 2024

    work page 2024

  11. [11]

    Development and valida- tion of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs,

    V . Gulshan, L. Peng, M. Coram, M. C. Stumpe, D. Wu, A. Narayanaswamy, S. Venugopalan, K. Widner, T. Madams, J. Cuadroset al., “Development and valida- tion of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs,”JAMA, vol. 316, no. 22, pp. 2402–2410, 2016

    work page 2016

  12. [12]

    Deep learning for diabetic retinopathy analysis: A review, research challenges, and future directions,

    M. W. Nadeem, H. G. Goh, V . Ponnusamy, I. Andonovic, M. A. Khan, and M. Hussain, “Deep learning for diabetic retinopathy analysis: A review, research challenges, and future directions,”Sensors, vol. 22, no. 18, p. 6780, 2022

    work page 2022

  13. [13]

    Managing Diabetic Retinopathy with Deep Learning: A Data Centric Overview

    S. Dey, Z. Khan, T. A. PramodKumar, B. U. Shankar, A. K. Dhara, R. Rajalakshmi, R. Raman, and S. Mitra, “Managing diabetic retinopathy with deep learning: A data centric overview,”arXiv preprint arXiv:2604.02448, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  14. [14]

    Bag of Tricks for Developing Diabetic Retinopathy Analysis Framework to Overcome Data Scarcity

    G. Kwon, E. Kim, S. Kim, S. Bak, M. Kim, and J. Kim, “Bag of tricks for developing diabetic retinopathy analy- sis framework to overcome data scarcity,”arXiv preprint arXiv:2210.09558, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  15. [15]

    Deciphering the impact of diversity in cnn-based en- sembles on overcoming data imbalance and scarcity in medical datasets: A case study on diabetic retinopathy,

    Inamullah, S. Hassan, S. B. Belhaouari, and I. Amin, “Deciphering the impact of diversity in cnn-based en- sembles on overcoming data imbalance and scarcity in medical datasets: A case study on diabetic retinopathy,” Informatics in Medicine Unlocked, vol. 49, p. 101557, 2024

    work page 2024

  16. [16]

    An inherently interpretable ai model improves screening speed and accuracy for early diabetic retinopathy,

    K. Djoumessi, Z. Huang, L. K ¨uhlewein, A. Rickmann, N. Simon, L. M. Koch, and P. Berens, “An inherently interpretable ai model improves screening speed and accuracy for early diabetic retinopathy,”PLOS Digital Health, vol. 4, no. 5, p. e0000831, 2025

    work page 2025

  17. [17]

    Grad-CAM: Visual explana- tions from deep networks via gradient-based localiza- tion,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-CAM: Visual explana- tions from deep networks via gradient-based localiza- tion,” inProceedings of the IEEE International Confer- ence on Computer Vision, 2017, pp. 618–626

    work page 2017

  18. [18]

    Deep image mining for diabetic retinopathy screening,

    G. Quellec, K. Charri `ere, Y . Boudi, B. Cochener, and M. Lamard, “Deep image mining for diabetic retinopathy screening,”Medical Image Analysis, vol. 39, pp. 178– 193, 2017

    work page 2017

  19. [19]

    The false hope of current approaches to explainable artificial intelligence in health care,

    M. Ghassemi, L. Oakden-Rayner, and A. L. Beam, “The false hope of current approaches to explainable artificial intelligence in health care,”The lancet digital health, vol. 3, no. 11, pp. e745–e750, 2021

    work page 2021

  20. [20]

    AutoMorph: Automated retinal vascular morphology quantification via a deep learning pipeline,

    Y . Zhou, S. K. Wagner, M. A. Chia, A. Zhao, P. Woodward-Court, M. Xu, R. R. Struyven, D. C. Alexander, and P. A. Keane, “AutoMorph: Automated retinal vascular morphology quantification via a deep learning pipeline,”Translational Vision Science & Tech- nology, vol. 11, no. 7, p. 12, 2022

    work page 2022

  21. [21]

    Lesion- based contrastive learning for diabetic retinopathy grad- ing from fundus images,

    Y . Huang, L. Lin, P. Cheng, J. Lyu, and X. Tang, “Lesion- based contrastive learning for diabetic retinopathy grad- ing from fundus images,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2021. Springer, 2021, pp. 113–123

    work page 2021

  22. [22]

    National health and nutrition examination survey: Questionnaires, datasets, and related documentation,

    National Center for Health Statistics, “National health and nutrition examination survey: Questionnaires, datasets, and related documentation,” https://wwwn.cdc.gov/nchs/nhanes/, 2026, accessed July 2026

    work page 2026

  23. [23]

    Automated identification of di- abetic retinopathy using deep learning,

    R. Gargeya and T. Leng, “Automated identification of di- abetic retinopathy using deep learning,”Ophthalmology, vol. 124, no. 7, pp. 962–969, 2017

    work page 2017

  24. [24]

    Pearl,Causality: Models, Reasoning, and Inference

    J. Pearl,Causality: Models, Reasoning, and Inference. Cambridge University Press, 2009

    work page 2009

  25. [25]

    A general approach to causal mediation analysis,

    K. Imai, L. Keele, and D. Tingley, “A general approach to causal mediation analysis,”Psychological Methods, vol. 15, no. 4, pp. 309–334, 2010

    work page 2010

  26. [26]

    T. J. Hastie and R. J. Tibshirani,Generalized Additive Models. Chapman and Hall, 1990

    work page 1990

  27. [27]

    S. N. Wood,Generalized Additive Models: An Introduc- tion with R. CRC Press, 2017

    work page 2017

  28. [28]

    Aptos 2019 blindness detection,

    Karthik, Maggie, and S. Dane, “Aptos 2019 blindness detection,” https://kaggle.com/competitions/ aptos2019-blindness-detection, 2019, kaggle

    work page 2019