A noninvasive model for chronic kidney disease screening and common pathological type identification from retinal images.

Abstract

Chronic kidney disease (CKD) is a global health challenge, but invasive renal biopsies, the gold standard for diagnosis and prognosis, are often clinically constrained. To address this, we developed the kidney intelligent diagnosis system (KIDS), a noninvasive model for renal biopsy prediction using 13,144 retinal images from 6773 participants. The KIDS achieves an area under the receiver operating characteristic curve (AUC) of 0.839-0.993 for CKD screening and accurately identifies the five most common pathological types (AUC: 0.790-0.932) in a multicenter and multi-ethnic validation, outperforming nephrologists by 26.98% in accuracy. Additionally, the KIDS further predicts disease progression based on pathological classification. Given its flexible strategy, the KIDS can be adapted to local conditions to provide a tailored tool for patients. This noninvasive model has the potential to improve CKD clinical management, particularly for those who are ineligible for biopsies.