Learned image groups reveal a low-sensitivity slice in cardiomegaly detection
Source check: Source checked. Not independent scientific confirmation.
Finding: A fitted subgroup model assigned chest X-rays in a 22,281-image CheXpert-Plus test set to image groups. Across discovered groups, cardiomegaly sensitivity ranged from 0.10 to 0.98. One group associated with higher BMI and male sex had sensitivity of 0.10 and specificity of 0.95. The group was defined by learned image features and classifier scores, and its demographic profile was assessed afterward.
Technology: DOMINO Gaussian mixture; CLIP; BiomedCLIP; ResNet-50
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- A modified DOMINO Gaussian mixture learned image subgroups using pretrained CLIP or BiomedCLIP features and the predictions of a trained ResNet-50 cardiomegaly classifier. The mixture was fitted on the validation set and assigned memberships in the test set.
- Test and result
- The authors fitted subgroup memberships on 22,263 validation images and then assigned an existing separate test partition. Real-world subgroup analyses were distinct from synthetic-artifact checks.
- Limitation / still unknown
- The subgroup is defined by the fitted model, with demographic enrichment assessed afterward. The low sensitivity is imprecise (0.103 ± 0.097, reported 95% Wilson uncertainty). Methods describes an image split and does not establish patient-disjointness. The study does not establish causal effects of BMI or sex or externally reproduced clinical benefit.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.32362v1 · https://arxiv.org/html/2610.32362v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: arXiv v1 front matter, Results, Discussion example SG A0 and Methods including dataset, cluster fitting and classifier.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.