Retinal-image models detect diabetic kidney disease across multiple cohorts
Source check: Source checked. Not independent scientific confirmation.
Finding: A retinal-image model trained for diabetic kidney disease achieved AUCs of 0.791–0.826 across ten external datasets totaling 65,406 participants. A separate task distinguished isolated diabetic nephropathy from other kidney disease in small biopsy-referenced cohorts. These results support diagnostic investigation without establishing improved patient outcomes or replacement of confirmatory tests.
Technology: ResNet-50 with momentum-contrastive pretraining
Publication: Peer-reviewed paper · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Contrastively pretrained image features were supervised for kidney-disease categories; a frozen image encoder and linear classifier supported the separate nephropathy task.
- Test and result
- DeepDKD used 734,084 pretraining images and 486,312 development/internal-validation images from 121,578 participants. The image-only classifier averaged scores across each participant’s available images. Internal DKD AUC was .842; ten external cohorts had AUC .791–.826. The biopsy-related task used 267 development participants and 244 external participants, with external AUC .733–.844. A 325-person prospective primary-care proof-of-concept reported sensitivity 89.8% and specificity 47.6% against subsequent confirmatory findings.
- Limitation / still unknown
- The represented principal result is retrospective multicohort diagnostic validation. External datasets and a prospectively collected cohort do not establish improved clinical outcomes. Camera/domain differences and population selection constrain transfer. The small prospective study selected a previously defined high-sensitivity operating point; specificity was low, and the apparent specificity improvement versus metadata was not significant (P=.072). Model-development split details and diagnostic-reference definitions are in an appendix not independently read here. The retinal model is an ancillary screen, not a non-invasive tissue biopsy or proven substitute for laboratory/biopsy confirmation.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://discovery.ucl.ac.uk/id/eprint/10213219/1/Wu_1-s2.0-S2589750025000408-main.pdf · https://pubmed.ncbi.nlm.nih.gov/40312169/
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: Version-specific original computational/statistical Methods and Results plus primary bibliographic dates. Clinical benefit, causal effects and independent replication are not inferred..
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 1/5 · 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.