Deep learning on routine pathology slides prescreens FGFR-altered bladder cancers in a global study
A deep-learning device analyzed routine H&E whole-slide images to estimate whether advanced urothelial tumors carried selected FGFR2/FGFR3 alterations used for erdafitinib eligibility. In a separate 350-slide retrospective validation set, the device achieved 88.7% sensitivity and 31.8% specificity; it was also prospectively integrated into a non-interventional study spanning 89 sites in nine countries. The paper projected a 28.7% reduction in molecular tests assuming 15% biomarker prevalence; that estimate was simulated, not a measured population-level reduction.
Technology: FGFR Device, a multi-instance-learning attention model with a convolutional neural-network backbone
Publication: Peer-reviewed paper · Test context: prospective or independent check. Catalog categories, not an evidence grade.
Result, AI contribution and limits
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
- Publication status: Peer-reviewed paper
- Test type: prospective or independent check
- Tested: In a separate 350-slide retrospective validation set, the device achieved 88.7% sensitivity and 31.8% specificity; it was also prospectively integrated into a non-interventional study spanning 89 sites in nine countries.
- Still unknown: The 350-slide retrospective validation set was deliberately enriched (150 FGFR-positive and 200 negative; 43% prevalence), and the device fell slightly short of the prespecified 90% sensitivity target. Prospective real-time workflow validation covered only 17 samples in one month (15 predictions; two quality-control failures); production deployment reported 24 samples, with only 20 receiving molecular testing. The 28.7% test reduction is a simulation assuming 15% prevalence, not an observed clinical outcome. This is prescreening support, not a stand-alone diagnosis or evidence of patient benefit.
- AI role: Trained on more than 3,000 urothelial-cancer slides with molecular test labels, the deep-learning model inferred selected FGFR alteration status from H&E images and returned a likelihood for prescreening before confirmatory molecular testing. The device was prospectively integrated into a non-interventional study workflow; molecular tests remained the reference standard.
Evidence sources: https://www.nature.com/articles/s41467-024-49153-9.pdf
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 scope and date are not recorded; source checked does not establish full-text review.
Citations: in the top 10% of works of the same type, year and subfield (OpenAlex work metrics, 30 citations; as of 15 Nov 2026). Attention, not a measure of correctness; not part of the score.
Editorial assessment: Impact 3/5 · Validation 4/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.
Follow-up and relationships
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.