Intersectional Fairness in Vision-Language Models for Medical Image Disease Classification

Adam G. Dunn; Jinman Kim; Usman Naseem; Yupeng Zhang

REVIEW 5 major objections 5 minor 1 cited by

Intersectional Fairness in Vision-Language Models for Medical Image Disease Classification

T0 review · 5 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A new training loss equalizes diagnostic confidence across patient subgroups, nearly halving missed-diagnosis gaps in medical vision-language models.

desk verdict A sensible, potentially useful training-time fairness regularizer for medical VLMs, but the ophthalmology headline rests on an unreconciled test-set inconsistency and the code/supplements are missing; worth a careful referee only after those are fixed. read the letter →

arxiv 2512.15249 v2 pith:S6UDFEVT submitted 2025-12-17 cs.CV cs.AI

Yupeng Zhang , Adam G. Dunn , Usman Naseem , Jinman Kim This is my paper · ORCID

classification cs.CVcs.AI

keywords intersectionalfairnessvision-languagemodelsmedicalimageclassificationdiagnosticcertaintymaximummeandiscrepancymisseddiagnosisgapbiasmitigationdermatologyandophthalmology

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

Fine-tuning makes medical vision-language models more accurate on average while quietly widening gaps in how confidently they diagnose different intersectional patient groups—young women with skin lesions, older non-white men with glaucoma—leaving those groups in a 'grey zone' of borderline predictions. The paper argues that the right fairness target is not a shared feature space but the distribution of per-patient diagnostic certainty, and it introduces CMAC-MMD, a loss that penalizes differences in that distribution across subgroups using maximum mean discrepancy. In skin-lesion screening the method reduces the largest subgroup gap in true positive rate from 0.50 to 0.26 (AUC 0.94→0.97); in glaucoma screening it cuts the same gap from 0.41 to 0.31 (AUC 0.71→0.72). Fairness gains persist on an external skin-lesion dataset, and demographics are needed only during training, not at inference. If the claim holds, equitable performance in high-stakes screening can be improved without sacrificing accuracy or patient privacy.

What carries the argument

The load-bearing object is the per-sample alignment score, defined as a_i = S_{i,correct} − max_{j≠i} S_{i,j}, where S is the temperature-scaled cosine similarity between image and text embeddings produced by a contrastive vision-language model. A high positive score means the model decisively prefers the correct diagnosis; a score near zero places the patient in the paper's 'grey zone' of uncertainty. CMAC-MMD is the training loss that groups these scalars by intersectional subgroup and adds a maximum mean discrepancy penalty over all subgroup pairs to the standard contrastive objective, forcing the whole distribution of certainty—not just its mean—to be the same across subgroups. That dist

What would settle it

Train the same model with the same MMD objective but randomly permute subgroup assignments during training; if the ΔTPR reductions persist, the effect is generic regularization, not distributional fairness. More decisively, re-run the ophthalmology experiment using finer-grained race/ethnicity labels: if the true-positive-rate gap returns to the baseline value when subgroups are defined more precisely, then the binarized White/non-White partition—not the CMAC-MMD loss—was responsible for the apparent improvement.

Watch

Extended reading notes

Core claim

The paper's central claim is that intersectional diagnostic disparities in medical vision-language models are driven by a certainty gap: even when aggregate accuracy is similar, models are systematically less confident for some demographic intersections, and those borderline predictions translate into missed diagnoses. The proposed method, CMAC-MMD, turns this diagnosis into a training objective. For each image–text pair it computes an alignment score—the margin by which similarity to the correct diagnosis exceeds similarity to the most compelling alternative—and treats that scalar as the patient's diagnostic certainty. During fine-tuning, a maximum mean discrepancy penalty matches the full

Load-bearing premise

The load-bearing premise is that the demographic labels used to form the training subgroups—age bins, binary gender, and the White/non-White split in the eye cohort—are reliable enough to mark the axes of disparity that matter clinically; if that metadata is noisy or too coarse, the regularizer matches the wrong distributions and the reported gains will not transfer to real patients.

Editorial extensions

If this is right

In the dermatology test set, the largest true-positive-rate gap between intersectional subgroups falls 48% (0.50 to 0.26) while AUC rises from 0.94 to 0.97; in absolute terms, 18 of the 64 missed melanomas under standard training would be detected.
In glaucoma screening, ΔTPR falls 24% (0.41 to 0.31) at AUC 0.72 versus 0.71, with the largest relative reductions in missed diagnoses occurring among non-white subgroups.
On an external skin-lesion dataset, the fairness gain transfers: ΔTPR is 0.15 for CMAC-MMD versus 0.23 for standard training, a 35% relative reduction, with no significant loss in AUC under distribution shift.
CMAC-MMD is one of the few methods to satisfy the paper's pre-specified differential-fairness and intersectional-fairness criteria, indicating equity was achieved by improving disadvantaged subgroups rather than 'levelling down' all groups.
Demographics are required only for computing the training penalty; at inference the model takes the image alone, so deployment does not raise new privacy or data-availability requirements.

Reading between the lines

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

(Editorial inference) The reusable insight is not tied to contrastive vision-language models: any classifier that emits a scalar confidence could receive the same distribution-matching regularizer, so the method may transfer to conventional medical classifiers that currently lack fairness constraints.
(Editorial inference) If the certainty gap truly lives at the decision boundary rather than in representation space, fairness evaluations should report confidence calibration per subgroup, not only accuracy metrics; this is testable on existing models by comparing subgroup confidence distributions before and after any debiasing method.
(Editorial inference) A direct falsification is to train CMAC-MMD with randomly permuted subgroup labels; if the ΔTPR reductions persist, the effect comes from generic regularization rather than from matching the true demographic distributions. The authors' own caveat that binarized race is a simplification suggests the ophthalmology result should be re-checked with finer-grained ethnicity labels.

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.

Desk Editor's Note

read the letter

The thing to know: CMAC-MMD is a clean idea — regularize the distribution of margin-based certainty scores across intersectional subgroups with MMD during CLIP fine-tuning — and the reported results are stronger than what most fairness papers show: dermatology ΔTPR 0.50→0.26 with AUC up, ophthalmology ΔTPR 0.41→0.31, and external validation on BCN20000. But I would not take the ophthalmology result at face value yet: Section 2.1 defines the held-out test as n=2,032, while Table 5 sums to 1,626. That is a 406-sample gap, not a typo. If the metrics were computed on a filtered subset, the comparison is not against the pre-specified split. This has to be reconciled before the cross-domain claim holds. The dermatology count is off by one (1,989 vs 1,990), which I’d call a rounding or transcription slip; the BCN20000 discrepancy (~12,000 in the abstract vs 18,946 in Table 1) also needs a one-line explanation, since it could just mean a subset was used.

What is actually new: prior VLM fairness work like FairCLIP constrains single-attribute representations, whereas CMAC-MMD operates on the 1-D distribution of diagnostic certainty scores over intersections, and it does not need demographics at inference. The clinical framing — the “certainty gap” and missed-diagnosis disparity — is meaningful, and the two-domain evaluation with an external test is more than most fairness papers do. The subgroup-level tables are detailed, and the method does not appear to be “leveling down.”

Soft spots, in proportion: the MMD loss directly minimizes distributional distance between subgroup alignment scores, so any reported improvement in certainty-overlap is partly by construction. That is not fatal because the headline claims are about ΔTPR and AUC, which are not directly optimized, and the external BCN20000 result helps. The main hyperparameter λ_CMAC=0.5 is selected on validation — that is standard but worth flagging. The authors also acknowledge the White/Non-White binarization and the limits of demographic metadata; that is an honest limitation, not a hidden one.

The real barriers are reproducibility: code is withheld until acceptance, key derivations are deferred to an unavailable supplement, and the test-set inconsistency is unresolved. These are addressable, so I would send this to peer review, but I would not accept it as-is. The authors need to release code, provide the supplement, and reconcile the sample counts.

I’d bring it to a reading group once the authors post a revised version with the numbers cleaned up. Until then, I wouldn’t cite the ophthalmology result in my own work.

Referee Report

5 major / 5 minor

Summary. The paper proposes CMAC-MMD, a training-time regularizer for CLIP-based architectures that penalizes the Maximum Mean Discrepancy (MMD) between per-sample alignment-score (diagnostic certainty) distributions across intersectional demographic subgroups. The method is evaluated on dermatology (HAM10000, internal; BCN20000, external) and glaucoma detection (Harvard-FairVLMed), and compared against ERM, resampling, reweighting, GroupDRO, DANN/CDANN, and FairCLIP. The headline claims are that CMAC-MMD reduces the missed-diagnosis gap ΔTPR while maintaining or improving AUC: dermatology ΔTPR 0.50→0.26 with AUC 0.94→0.97; ophthalmology ΔTPR 0.41→0.31 with AUC 0.71→0.72; external validation shows ΔTPR 0.15 vs 0.23. A central methodological novelty is that demographic attributes are used only at training time, not at inference.

Significance. If the empirical claims hold, this is a valuable contribution to a timely problem: it addresses intersectional fairness, not just single-attribute parity, and it does so at the decision-certainty level rather than through feature-space debiasing. The paper includes external validation, multiple fairness baselines, and quantifies clinical impact in terms of prevented false negatives. However, the current manuscript contains unresolved inconsistencies in the definition of the test sets and in the relationship between the metric being regularized and the certainty metric used for analysis. These inconsistencies currently block confidence in the headline numerical results. The central idea is plausible and the direction is promising, but the evidence as presented is not yet reliable.

major comments (5)

[§2.1, §3.3, Table 5] The ophthalmology held-out test set is defined in §2.1 as n=2,032, but the subgroup sizes in Table 5 sum to 1,626 and the clinical-impact calculations use n=1,626. No filtering or exclusion step is described anywhere in the manuscript. Since the headline ΔTPR reduction (0.41→0.31), the AUC values, and the '28 prevented missed diagnoses' are all computed on this test set, the primary endpoint is not measured against the pre-specified held-out split. This is a load-bearing internal inconsistency. The authors must either reconcile the counts or explicitly report the inclusion/exclusion criteria and recompute all affected statistics.
[§2.1, Table 3] The dermatology test set is stated as n=1,989, but Table 3 lists subgroup sizes that sum to 1,990. Moreover, the training/validation/test split sizes sum to 5,968+1,989+1,989=9,946, not 10,015. These discrepancies suggest an unreported filtering step. Because the dermatology results (ΔTPR, AUC, false-negative counts) are central to the paper, the sample accounting must be reconciled for both cohorts.
[Abstract, §2.1, §3.4, Table 1] BCN20000 is described as '~12,000 images' in the abstract, §2.1, and §3.4, but Table 1 lists the dataset as containing 18,946 images. If the external-validation experiments used a labeled subset of approximately 12,000 images, the construction of that subset must be described. External validation is a central part of the generalisability claim, so the denominator and any filtering criteria need to be explicit.
[§2.2 Eq. (1), §2.3 Eqs. (3)–(4)] The paper defines two different quantities as 'diagnostic certainty'. Eq. (1) defines c_i as a softmax-calibrated probability over diagnostic classes, and Eq. (2) measures the certainty gap as a difference in mean c_i. In contrast, the CMAC-MMD loss in Eq. (4) regularizes the distribution of the alignment score a_i = S_ii − max_{j≠i} S_ij from Eq. (3), which is a margin, not a probability. The zone-of-uncertainty and KDE analyses in §3.1 use probabilities, while the training objective uses margins. The paper does not reconcile these two operationalizations. Since the core claim is that CMAC-MMD 'directly regularises the distribution of diagnostic certainty scores', the relationship between a_i and c_i must be clarified. If a_i is intended as a proxy, the justification and any re-analysis using c_i should be provided.
[§2.3 Eq. (4), §3.2–3.3] Because the MMD loss directly minimizes the distributional distance between subgroup alignment-score distributions, any improvement in alignment-score distributional overlap is partly by construction. The paper's headline outcomes (ΔTPR and AUC) are not directly optimized, which is the appropriate focus, but the manuscript also presents improvements in 'certainty gap' and 'uncertain predictions' as evidence. These certainty metrics should be clearly separated from the optimized quantity, and any claims about improved certainty-distribution alignment should acknowledge that this alignment is the explicit training target.

minor comments (5)

[§3.2] The DEOdds improvement in dermatology is reported with Wilcoxon W=1.0, p=0.0625, which does not reach the pre-specified 0.05 threshold. The sentence 'representing a 60% improvement' should be accompanied by an explicit statement of non-significance, or the significance threshold should be reconsidered.
[§3.1, Table 5] In the zero-shot CLIP example, the subgroups 'White Male 60+' and 'White Female 0-60' are both reported as n=110, but Table 5 lists their test-set sizes as 472 and 108, respectively. These numbers should be corrected or explained.
[§3.4, Table 6] The text reports ΔAUC=-0.01 for CMAC-MMD versus ERM on BCN20000, while Table 6's 'Absolute change' row shows -0.02 (presumably for DPD) and -0.08 (for ΔTPR). The table layout and text should be aligned to avoid ambiguity about which metric the absolute changes refer to.
[§2.2] The uncertainty zone [0.40, 0.60] is said to be chosen 'based on statistical analysis', but the analysis is not described. Please specify how these boundaries were derived.
[Supplementary references] Several key results, including gradient derivations and the λ_CMAC sensitivity analysis, are referenced as Supplementary Appendices D and Fig. A1 but are not included in the provided manuscript. Please ensure the supplementary material is available for review.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the MMD regularizer optimizes alignment-score distributions, while the headline ΔTPR/AUC outcomes are separately measured on held-out and external data.

full rationale

The paper's central claim is that CMAC-MMD reduces the intersectional missed-diagnosis gap (ΔTPR) without sacrificing AUC. The training loss (Eq. 4-5) directly penalizes MMD between per-subgroup distributions of the alignment score a_i (Eq. 3), but the reported success metrics are defined on binary true-positive rates (Eq. 6) and AUC, both evaluated on strictly held-out test data and on the external BCN20000 cohort. These are not equal to the training objective by construction: matching marginal alignment-score distributions does not logically force equality of conditional TPRs across subgroups, and the external result cannot be created by a HAM10000 training loss. No fitted parameter is relabeled as a prediction: λ_CMAC=0.5 is validation-selected and the test metrics are held out; Eq. 7 is an algebraic restatement of TPR differences, not a new prediction. No uniqueness theorem is imported, and the baselines (FairCLIP, DANN, etc.) are empirical comparisons rather than definitional constraints. The paper also acknowledges limitations (binarized race, retrospective design, algorithmic intervention not upstream causes), which are not circular. I flag two non-circular concerns: (i) the ophthalmology test denominator is inconsistent (§2.1 states n=2,032; Table 5 subgroup counts sum to 1,626), and dermatology similarly states n=1,989 while Table 3 sums to 1,990; this is a correctness/reporting issue that should be reconciled but is not a circularity. (ii) Reference [41], used to describe DANN, may be a self-citation, but it is not load-bearing because the DANN comparison is empirical. Overall, the derivation chain is self-contained for its headline claim.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The ledger is dominated by modeling choices and dataset assumptions rather than fitted physical constants. The main free parameters are the regularization weight, the uncertainty-zone threshold, and the unspecified MMD kernel bandwidth; no new physical entities are introduced. The most consequential assumptions are that the demographic metadata are reliable and that equalizing margin-score distributions yields equitable TPR.

free parameters (3)

λ_CMAC = 0.5
Fairness regularization weight in Eq. 5, selected on the validation set (Sec 2.4); controls the trade-off between AUC and ΔTPR, so the headline results are partially tuned.
Uncertainty zone boundaries = [0.40, 0.60]
Defined 'based on statistical analysis' in Sec 2.2; the choice affects the reported certainty-gap quantification, though not the primary AUC/ΔTPR outcomes.
MMD RBF kernel bandwidth = not reported
Required to compute MMD in Eq. 4; not specified in the main text, leaving the exact training objective underdetermined.

assumptions (5)

domain assumption CLIP cosine-similarity scores between an image and class-text embeddings are a valid diagnostic certainty measure.
Used to define c_i in Eq. 1 and a_i in Eq. 3; if softmax similarities are poorly calibrated, the regularized quantity may not capture clinical confidence.
domain assumption MMD between 1-D alignment-score distributions is a meaningful fairness objective.
Eq. 4 uses MMD to enforce overlap of score distributions; this assumes distributional overlap in margin scores translates into equitable true-positive rates.
domain assumption Demographic metadata in the datasets are accurate and the chosen subgroup definitions are clinically meaningful.
Section 2.1 constructs subgroups from age, gender, and race; errors or coarse binarization would misdirect the regularization.
domain assumption Held-out test sets were not used in model or hyperparameter selection.
Asserted in Secs 2.4–2.5, but the unexplained ophthalmology test-size discrepancy (2,032 vs 1,626) makes this assumption harder to audit.
domain assumption The external BCN20000 subset used for validation is well-defined and histopathology-confirmed.
Table 1 lists 18,946 BCN20000 images while the text uses 'approximately 12,000'; the exact external-validation subset is not specified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Intersectional Fairness in Vision-Language Models for Medical Image Disease Classification." pith.science (2026). https://pith.science/paper/S6UDFEVT

@misc{pith2026251215249,
  author       = {Pith},
  title        = {Pith review of: Intersectional Fairness in Vision-Language Models for Medical Image Disease Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6UDFEVT}},
  note         = {Machine review of arXiv:2512.15249}
}

abstract

Medical artificial intelligence (AI) systems, particularly multimodal vision-language models (VLM), often exhibit intersectional biases where models are systematically less confident in diagnosing marginalised patient subgroups. Such bias can lead to higher rates of inaccurate and missed diagnoses due to demographically skewed data and divergent distributions of diagnostic certainty. Current fairness interventions frequently fail to address these gaps or compromise overall diagnostic performance to achieve statistical parity among the subgroups. In this study, we developed Cross-Modal Alignment Consistency (CMAC-MMD), a training framework that standardises diagnostic certainty across intersectional patient subgroups. Unlike traditional debiasing methods, this approach equalises the model's decision confidence without requiring sensitive demographic data during clinical inference. We evaluated this approach using 10,015 skin lesion images (HAM10000) with external validation on 12,000 images (BCN20000), and 10,000 fundus images for glaucoma detection (Harvard-FairVLMed), stratifying performance by intersectional age, gender, and race attributes. In the dermatology cohort, the proposed method reduced the overall intersectional missed diagnosis gap (difference in True Positive Rate, $\Delta$TPR) from 0.50 to 0.26 while improving the overall Area Under the Curve (AUC) from 0.94 to 0.97 compared to standard training. Similarly, for glaucoma screening, the method reduced $\Delta$TPR from 0.41 to 0.31, achieving a better AUC of 0.72 (vs. 0.71 baseline). This establishes a scalable framework for developing high-stakes clinical decision support systems that are both accurate and can perform equitably across diverse patient subgroups, ensuring reliable performance without increasing privacy risks.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Intersectional Disentangling of Temporal and Acquisition Bias in Fetal Ultrasound
cs.LG 2026-05 unverdicted novelty 6.0 of 10

The headline disentangling of scan-delivery and pixel-spacing effects is absent from the paper's body, which reports different results.

Reference graph

Works this paper leans on

54 extracted references · 2 linked inside Pith · cited by 1 Pith paper

[1]

M., Parker, D

Hu, S., Soza-Vento, R. M., Parker, D. F. & Kirsner, R. S. Comparison of stage at diagnosis of melanoma among Hispanic, black, and white patients in Miami-Dade County, Florida.Arch. Dermatol.142, 704–708 (2006)

2006
[2]

M., Tsai, S., Gittleman, H., Barnholtz-Sloan, J

Dawes, S. M., Tsai, S., Gittleman, H., Barnholtz-Sloan, J. S. & Bordeaux, J. S. Racial disparities in melanoma survival.J. Am. Acad. Dermatol.75, 983–991 (2016)

2016
[3]

M.et al.Racial variations in the prevalence of primary open-angle glaucoma: The Baltimore Eye Survey.JAMA266, 369–374 (1991)

Tielsch, J. M.et al.Racial variations in the prevalence of primary open-angle glaucoma: The Baltimore Eye Survey.JAMA266, 369–374 (1991)

1991
[4]

Varma, R.et al.Prevalence of open-angle glaucoma and ocular hypertension in Latinos: The Los Angeles Latino Eye Study.Ophthalmology111, 1439–1448 (2004)

2004
[5]

Sommer, A.et al.Racial differences in the cause-specific prevalence of blindness in east Baltimore.N. Engl. J. Med.325, 1412–1417 (1991)

1991
[6]

Allison, K., Patel, D. G. & Greene, L. Racial and ethnic disparities in primary open-angle glaucoma clinical trials: A systematic review and meta-analysis.JAMA Netw. Open4, e218348 (2021)

2021
[7]

A.et al.The prevalence of glaucoma in a population-based study of Hispanic subjects: Proyecto VER.Arch

Quigley, H. A.et al.The prevalence of glaucoma in a population-based study of Hispanic subjects: Proyecto VER.Arch. Ophthalmol.119, 1819–1826 (2001)

2001
[8]

D., Lavin, P

Abr` amoff, M. D., Lavin, P. T., Birch, M., Shah, N. & Folk, J. C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices.npj Digit. Med.1, 39 (2018)

2018

Show all 54 references

[9]

P., Kadakia, K

Venkatesh, K. P., Kadakia, K. T. & Gilbert, S. Learnings from the first AI-enabled skin cancer device for primary care authorized by FDA.npj Digit. Med.7, 156 (2024)

2024
[10]

EyeArt AI eye screening system: Real-world implementation in Federally Qualified Health Centers.Eyenuk Technical Report(2024)

Eyenuk Inc. EyeArt AI eye screening system: Real-world implementation in Federally Qualified Health Centers.Eyenuk Technical Report(2024). https://www.eyenuk.com. 18

2024
[11]

Adv.8, eabq6147 (2022)

Daneshjou, R.et al.Disparities in dermatology AI performance on a diverse, curated clinical image set.Sci. Adv.8, eabq6147 (2022)

2022
[12]

IEEE/CVF Conf

Groh, M.et al.Evaluating deep neural networks trained on clinical images in dermatology with the Fitzpatrick 17k dataset.Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. 2021, 1820–1828 (2021)

2021
[13]

Med.30, 573–583 (2024)

Groh, M.et al.Deep learning-aided decision support for diagnosis of skin disease across skin tones.Nat. Med.30, 573–583 (2024)

2024
[14]

Radford, A.et al.Learning transferable visual models from natural language supervision. Proc. Mach. Learn. Res.139, 8748–8763 (2021)

2021
[15]

& Hoi, S

Li, J., Li, D., Savarese, S. & Hoi, S. BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models.Proc. Mach. Learn. Res.202, 19730–19742 (2023)

2023
[16]

Zhang, S.et al.BiomedCLIP: A multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs.NEJM AI1(2024)

2024
[17]

Notes Comput

Lin, W.et al.PMC-CLIP: Contrastive language-image pre-training using biomedical documents.Lect. Notes Comput. Sci.14221, 525–536 (2023)

2023
[18]

& Sun, J

Wang, Z., Wu, Z., Agarwal, D. & Sun, J. MedCLIP: Contrastive learning from unpaired medical images and text.Proc. Conf. Empir. Methods Nat. Lang. Process.2022, 3876–3887 (2022)

2022
[19]

& de Melo, G

Eslami, S., Meinel, C. & de Melo, G. PubMedCLIP: How much does CLIP benefit visual ques- tion answering in the medical domain?Proc. Conf. Eur. Chapter Assoc. Comput. Linguist. 2023, 1181–1193 (2023)

2023
[20]

Zhang, K.et al.A generalist vision–language foundation model for diverse biomedical tasks. Nat. Med.30, 3129–3141 (2024)

2024
[21]

Adv.11, eadq3609 (2025)

Yang, Y.et al.Demographic bias of expert-level vision-language foundation models in medical imaging.Sci. Adv.11, eadq3609 (2025)

2025
[22]

Med.8, 360 (2025)

Liu, M.et al.A scoping review and evidence gap analysis of clinical AI fairness.npj Digit. Med.8, 360 (2025)

2025
[23]

& Cheng, L

Gohar, U. & Cheng, L. A survey on intersectional fairness in machine learning: Notions, mitigation, and challenges.Proc. Int. Jt. Conf. Artif. Intell.32, 6580–6589 (2023)

2023
[24]

D., Joshi, S., Mazwi, M

McCradden, M. D., Joshi, S., Mazwi, M. & Anderson, J. A. Ethical limitations of algorithmic fairness solutions in health care machine learning.Lancet Digit. Health2, e221–e223 (2020)

2020
[25]

Y.et al.Ethical machine learning in healthcare.Annu

Chen, I. Y.et al.Ethical machine learning in healthcare.Annu. Rev. Biomed. Data Sci.4, 123–144 (2021)

2021
[26]

R., Islam, R., Keya, K

Foulds, J. R., Islam, R., Keya, K. N. & Pan, S. An intersectional definition of fairness.Proc. IEEE Int. Conf. Data Eng.2020, 1918–1921 (2020)

2020
[27]

& Keller, M

Maheshwari, G., Bellet, A., Denis, P. & Keller, M. Fair without leveling down: A new intersec- tional fairness definition.Proc. Conf. Empir. Methods Nat. Lang. Process.2023, 2335–2354 (2023)

2023
[28]

A., Echeveste, R

Ricci Lara, M. A., Echeveste, R. & Ferrante, E. Addressing fairness in artificial intelligence for medical imaging.Nat. Commun.13, 4581 (2022)

2022
[29]

& Kittler, H

Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Sci. Data5, 180161 (2018). 19

2018
[30]

Massive biomolecular shifts occur in our 40s and 60s, Stanford Medicine study finds.Stanford Medicine News(2024)

Stanford Medicine. Massive biomolecular shifts occur in our 40s and 60s, Stanford Medicine study finds.Stanford Medicine News(2024). https://med.stanford.edu/news/all-news/2024/ 08/massive-biomolecular-shifts-occur-in-our-40s-and-60s--stanford-m.html

2024
[31]

Perez, C. H. BCN20000: Dermoscopic lesions in the wild.Sci. Data11, 641 (2024)

2024
[32]

IEEE/CVF Conf

Luo, Y.et al.FairCLIP: Harnessing fairness in vision-language learning.Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.2024, 12289–12301 (2024)

2024
[33]

& Jiang, B

Zhang, N., Wang, J., Li, Y. & Jiang, B. Prevalence of primary open angle glaucoma in the last 20 years: A meta-analysis and systematic review.Sci. Rep.11, 13762 (2021)

2021
[34]

W., Katabi, D

Yang, Y., Zhang, H., Gichoya, J. W., Katabi, D. & Ghassemi, M. The limits of fair medical imaging AI in real-world generalization.Nat. Med.30, 2838–2848 (2024)

2024
[35]

& Crowley, M

Ghojogh, B., Ghodsi, A., Karray, F. & Crowley, M. Reproducing kernel hilbert space, mercer’s theorem, eigenfunctions, nystr¨ om method, and use of kernels in machine learning: Tutorial and survey.arXiv preprint arXiv:2106.08443(2021)

2021 arXiv
[36]

Thurnhofer-Hemsi, K., L´ opez-Rubio, E., Molina-Cabello, M. A. & Najarian, K. Radial basis function kernel optimization for Support Vector Machine classifiers.arXiv preprint arXiv:2007.08233(2020)

2007 arXiv
[37]

M., Rasch, M

Gretton, A., Borgwardt, K. M., Rasch, M. J., Scholkopf, B. & Smola, A. A kernel two-sample test.J. Mach. Learn. Res.13, 723–773 (2012)

2012
[38]

Bagui, S. & Li, K. Resampling imbalanced data for network intrusion detection datasets.J. Big Data8, 6 (2021)

2021
[39]

& Urtasun, R

Ren, M., Zeng, W., Yang, B. & Urtasun, R. Learning to reweight examples for robust deep learning.Proc. Mach. Learn. Res.80, 4334–4343 (2018)

2018
[40]

W., Hashimoto, T

Sagawa, S., Koh, P. W., Hashimoto, T. B. & Liang, P. Distributionally robust neural networks. Proc. Int. Conf. Learn. Represent.2020(2020)

2020
[41]

& Ruan, W

Zhang, Y., Zhang, T., Mu, R., Huang, X. & Ruan, W. Towards fairness-aware adversarial learning.Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit.2024, 28156–28165 (2024)

2024
[42]

& Jordan, M

Long, M., Cao, Z., Wang, J. & Jordan, M. I. Conditional adversarial domain adaptation.Adv. Neural Inf. Process. Syst.31, 1640–1650 (2018)

2018
[43]

R., DeLong, D

DeLong, E. R., DeLong, D. M. & Clarke-Pearson, D. L. Comparing the areas under two or more correlated receiver operating characteristic curves: A nonparametric approach.Biometrics 44, 837–845 (1988)

1988
[44]

Individual comparisons by ranking methods.Biometrics Bull.1, 80–83 (1945)

Wilcoxon, F. Individual comparisons by ranking methods.Biometrics Bull.1, 80–83 (1945)

1945
[45]

Pacheco, A. G. C.et al.PAD-UFES-20: A skin lesion dataset composed of patient data and clinical images collected from smartphones.Data Brief32, 106221 (2020)

2020
[46]

& Liu, H

Li, L., Xu, M., Wang, X., Jiang, L. & Liu, H. Attention based glaucoma detection: A large- scale database and CNN model.Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. 2019, 10571–10580 (2019)

2019
[47]

Data9, 291 (2022)

Kovalyk, O.et al.PAPILA: Dataset with fundus images and clinical data of both eyes of the same patient for glaucoma assessment.Sci. Data9, 291 (2022)

2022
[48]

& Ovreiu, E

Ovreiu, S., Paraschiv, E.-A. & Ovreiu, E. Deep learning & digital fundus images: Glaucoma detection using DenseNet.Proc. Int. Conf. Electron. Comput. Artif. Intell.2021, 1–4 (2021)

2021
[49]

Zhang, Z.et al.ORIGA-light: An online retinal fundus image database for glaucoma analysis and research.Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.2010, 3065–3068 (2010). 20

2010
[50]

& Hamarneh, G

Abhishek, K., Jain, A. & Hamarneh, G. Investigating the quality of DermaMNIST and Fitzpatrick17k dermatological image datasets.Sci. Data12, 196 (2025)

2025
[51]

Xu, J.et al.Algorithmic fairness in computational medicine.EBioMedicine84, 104250 (2022)

2022
[52]

Dermatol.145, 1369–1374 (2009)

Hu, S.et al.Disparity in melanoma: a trend analysis of melanoma incidence and stage at diagnosis among whites, Hispanics, and blacks in Florida.Arch. Dermatol.145, 1369–1374 (2009)

2009
[53]

& Daneshvar, M

Sagona, M., Dai, T., Macis, M. & Daneshvar, M. Trust in AI-assisted health systems and AI’s trust in humans.npj Health Syst.2, 10 (2025)

2025
[54]

Drukker, K.et al.Toward fairness in artificial intelligence for medical image analysis: Iden- tification and mitigation of potential biases in the roadmap from data collection to model deployment.J. Med. Imaging10, 061104 (2023). Supplementary Information This article has a su...

2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

M., Parker, D

Hu, S., Soza-Vento, R. M., Parker, D. F. & Kirsner, R. S. Comparison of stage at diagnosis of melanoma among Hispanic, black, and white patients in Miami-Dade County, Florida.Arch. Dermatol.142, 704–708 (2006)

2006

[2] [2]

M., Tsai, S., Gittleman, H., Barnholtz-Sloan, J

Dawes, S. M., Tsai, S., Gittleman, H., Barnholtz-Sloan, J. S. & Bordeaux, J. S. Racial disparities in melanoma survival.J. Am. Acad. Dermatol.75, 983–991 (2016)

2016

[3] [3]

M.et al.Racial variations in the prevalence of primary open-angle glaucoma: The Baltimore Eye Survey.JAMA266, 369–374 (1991)

Tielsch, J. M.et al.Racial variations in the prevalence of primary open-angle glaucoma: The Baltimore Eye Survey.JAMA266, 369–374 (1991)

1991

[4] [4]

Varma, R.et al.Prevalence of open-angle glaucoma and ocular hypertension in Latinos: The Los Angeles Latino Eye Study.Ophthalmology111, 1439–1448 (2004)

2004

[5] [5]

Sommer, A.et al.Racial differences in the cause-specific prevalence of blindness in east Baltimore.N. Engl. J. Med.325, 1412–1417 (1991)

1991

[6] [6]

Allison, K., Patel, D. G. & Greene, L. Racial and ethnic disparities in primary open-angle glaucoma clinical trials: A systematic review and meta-analysis.JAMA Netw. Open4, e218348 (2021)

2021

[7] [7]

A.et al.The prevalence of glaucoma in a population-based study of Hispanic subjects: Proyecto VER.Arch

Quigley, H. A.et al.The prevalence of glaucoma in a population-based study of Hispanic subjects: Proyecto VER.Arch. Ophthalmol.119, 1819–1826 (2001)

2001

[8] [8]

D., Lavin, P

Abr` amoff, M. D., Lavin, P. T., Birch, M., Shah, N. & Folk, J. C. Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices.npj Digit. Med.1, 39 (2018)

2018