AI Science Tracker
EN

Retinal-image models detect diabetic kidney disease across multiple cohorts

Shanghai Sixth People’s Hospital; Shanghai Jiao Tong University · 30 Apr 2025 · medicine

Discovery

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.

Open in the full catalog for filters, charts and related discoveries. Browse medicine discoveries.