Mammography Foundation Models for Opportunistic Prediction of Major Adverse Cardiovascular Events
Paula Feldman, Nusrat Binta Nizam, Sunwoo Kwak, Batuhan Karaman, Katerina Dodelzon, Mert Sabuncu
eess.IV · cs.CV · cs.LG
Quick verdict
Mammography foundation models predict 5-year MACE (AUROC 0.82) from pixels alone, per single-site study; add-on value over clinical variables untested.
Summary
The paper asks whether mammography foundation models, pretrained only for breast-imaging tasks, can predict 5-year major adverse cardiovascular events (MACE) from screening mammograms without cardiovascular labels or BAC annotation. From 28,746 women screened at Weill Cornell Medicine (2014-2019), the authors build a cohort of 22,497 women linked to EHR outcomes, with 500 events (2.22%), defined by ICD-9/10 codes for MI, ischemic stroke, and acute heart failure, and excluding events within 180 days of imaging to limit leakage. Frozen Mammo-CLIP and Mammo-FM encoders supply view embeddings that are mean-pooled and fed to a small MLP. On a 5,287-patient test set (118 events), AUROC is 0.823 (CLIP) and 0.822 (FM) versus 0.765 for age-only and 0.859 for a tabular clinical model; top-20% predicted risk captures 68.6% of events at 7.7% precision. Among true negatives, BAC-positive women receive 2.5-2.6-fold higher predicted risk, and Grad-CAM maps highlight vessel-like structures. Limitations: single institution; ICD-coded, non-adjudicated outcomes; no combined image-plus-clinical model despite the 'complementary' claims; interpretability is qualitative only; and the tabular baseline's missing-data handling and the BAC-extraction method are unspecified.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Mammography foundation models pretrained for breast tasks transfer to 5-year MACE prediction from screening mammograms alone, reaching AUROC 0.823 (Mammo-CLIP) and 0.822 (Mammo-FM) versus 0.765 for age-only.
- Both models assign substantially higher predicted risk (2.5-2.6-fold) to radiologist-documented BAC-positive true negatives despite BAC never being used as a training label, with a stronger association than clinical-variable or age models.
- Grad-CAM attributions concentrate on retroareolar tissue and vessel-like structures, suggesting the models leverage vascular features directly from mammographic pixels.
- Screening mammography could provide an opportunistic source of cardiovascular risk information to complement conventional clinical assessment without additional imaging.
Strengths
Careful leakage controls
Prior-event exclusion, a 180-day post-mammogram exclusion window, patient-level splits, a single canonical index scan at test time, and scan-level multi-instance labels confined to training with explicit time-alignment rules (Section 2.1).
Honest baseline hierarchy
Reports the tabular clinical model (AUROC 0.859) above the image models, with paired bootstrap significance versus age-only 0.765 (p=0.0016 CLIP, p=0.0040 FM) rather than claiming superiority over clinical assessment (Section 3.1, Table 3).
Triangulated biological plausibility without cardiovascular labels
Clinically framed evaluation beyond AUROC
Weaknesses
Complementarity claimed but never tested
Underpowered equivalence framing
ICD-coded outcomes without adjudication
Underspecified baseline and label provenance
Protocol and editorial inconsistencies
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv:2609.19385, formatted for PMLR, marked Under Review) tests a specific transfer hypothesis: mammography foundation models pretrained only for breast-imaging tasks — Mammo-CLIP (Ghosh et al., 2024) and Mammo-FM (Ghosh et al., 2025) — can predict 5-year major adverse cardiovascular events (MACE) from screening mammograms alone, without cardiovascular labels, BAC annotation, or handcrafted vascular features. The stated contributions (Section 1) are cohort construction, transfer evaluation, benchmarking against clinical and age-only baselines, and interpretability. Headline claims: AUROC 0.823 (Mammo-CLIP) and 0.822 (Mammo-FM) versus 0.765 age-only; substantially higher predicted risk for radiologist-documented BAC among true negatives; Grad-CAM attribution on vessel-like structures; and the abstract's framing that mammography can 'complement conventional clinical assessment'.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Can a mammogram predict heart attack or stroke risk?
In this study, deep learning models built on mammography foundation models predicted the 5-year risk of heart attack, ischemic stroke, or acute heart failure with AUROC about 0.82 using mammograms alone — clearly better than age alone (0.765) but below a model using full clinical variables such as cholesterol and blood pressure (0.859). This is a research result from a single institution, not an approved clinical tool.
What does 'opportunistic cardiovascular risk prediction' mean?
It means estimating cardiovascular risk from an imaging exam that was already acquired for another purpose — here, routine breast cancer screening — without extra imaging, extra radiation, or added radiologist workload.
Did the models actually detect breast arterial calcification (BAC)?
BAC was never used as a training label. However, models assigned BAC-positive patients about 2.5 to 2.6 times higher predicted risk, and Grad-CAM heatmaps concentrated on vessel-like structures, suggesting the models pick up vascular signals indirectly from the image.
Is this technology ready for clinical use?
No. The study is retrospective and single-center, uses billing-code-defined outcomes without chart adjudication, lacks external validation, and does not demonstrate that the mammogram adds value on top of existing clinical risk factors.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are available only as captions in the extracted text (ROC curves, enrichment/precision-recall, decision curves, BAC distributions, Grad-CAM overlays); visual claims could not be verified.
- The stated code repository (github.com/PauFeld/MammoCVD) was not inspected; code completeness and correctness are unverified.
- The paper contains minor cross-reference errors (Sections 3.5 and the Conclusion cite Section 3.3 for the BAC analysis, which is Section 3.4), reflecting hurried editorial assembly.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.