Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images.
Level 3 - non-randomized controlled study
Retrospective deep-learning model development and external cohort validation with prospective and longitudinal testing
PubMed 34131321 · doi:10.1038/s41551-021-00745-6
What was done
Deep-learning models were developed and validated to detect chronic kidney disease (CKD) and type 2 diabetes (T2D) from retinal fundus photographs alone or combined with clinical metadata (age, sex, height, weight, body mass index, and blood pressure). The development and validation dataset comprised 115,344 fundus images from 57,672 individuals. The authors also tested model capability to estimate continuous biomarkers (estimated glomerular filtration rate [eGFR] and blood glucose), stratify progression risk in a longitudinal cohort, validate across external population-based cohorts, and evaluate performance prospectively using smartphone-captured fundus images.
What was found
Models detected CKD and T2D with areas under the receiver operating characteristic curve (AUC) between 0.85 and 0.93. For continuous biomarker estimation, the models predicted eGFR with mean absolute errors (MAE) of 11.1 to 13.4 ml min⁻¹ per 1.73 m² and blood glucose with MAE of 0.65 to 1.1 mmol l⁻¹.
Why it matters
This approach shows that non-invasive retinal photography—including smartphone-based imaging—can serve as an accessible opportunistic screening tool for common chronic metabolic and renal conditions in remote or resource-constrained settings.
Limits
The abstract lacks breakdown of diagnostic performance metrics (e.g., sensitivity, specificity, positive predictive values) across separate cohorts. Long-term clinical utility, downstream diagnostic confirmation pathways, and whether early image-based detection improves hard clinical outcomes were not reported.
Cited by
- supports AI applied to high-resolution retinal imaging can detect early signs of Alzheimer's disease and kidney dysfunction.