Paper Overview
Field: Computer Vision / Computational Pathology Authors: Shengjia Chen, Ruchika Verma, Kevin Clare, Paula Restrepo, Jannes Jegminat Published: 2026-11-06 arXiv: 2611.09851
Summary
How much demographic-associated information is detectable across organs and disease states in routine clinical histopathology has remained unclear. This paper analyzes 1.3 million hematoxylin and eosin (H&E) whole-slide images from 257,322 patients spanning 25 organs, using pathology foundation-model representations combined with multiple-instance learning to systematically map information associated with sex, age group, and self-reported race.
Key Findings
- Across 69 organ-demographic combinations, demographic attributes were broadly recoverable: pan-organ mean AUROCs of 0.828 for sex, 0.733 for age group, and 0.662 for self-reported race, with substantial variation across organs.
- Demographic-class balancing produced relatively small changes in performance.
- Evaluation within diagnosis-text clusters showed only mild attenuation of demographic signal.
- With diagnosis-grouped training, demographic-associated signals remained detectable in multiple organs:
- Sex in stomach tissue: AUROC 0.865
- Age group in breast: AUROC 0.816
- Self-reported race in skin: AUROC 0.728
- A literature-based analysis generated candidate tissue-morphological and contextual explanations for these cases.
Significance
The work provides a systematic, health-system-wide atlas of demographic-associated information in routine histopathology. It enables investigation of the sources of these signals, assessment of potential shortcut learning in deep learning models, and development of more interpretable computational pathology models.
---
*Auto-collected on 2026-11-10*