AI-Based Noninvasive Retinal Imaging for CKD Risk Prediction in DM Patients without Established CKD

Introduction and Objective: 

Dr.Noon CKD is a deep-learning model analyzing retinal images to predict the likelihood of eGFR ≤ 60mL/min/1.73m² (reti-eGFR) and estimate 5-year CKD risk using age, sex, and reti-eGFR in a Cox regression model. This study evaluates its utility in DM patients without CKD by comparing CKD risk predictions between hypertensive and normotensive patients.

Methods: 

After excluding patients with CKD (eGFR ≤ 60 or UACR ≥ 30mg/g), UK Biobank data were used for 1:1 propensity score matching. Matching criteria included age, sex, UACR, BMI, LDL-C, HDL-C, BP, and smoking status for reti-eGFR and eGFR analyses. For 5-year CKD risk prediction, age and sex were excluded from the criteria. Dr.Noon CKD scores were averaged across both eyes. Statistical analyses included Wilcoxon rank-sum tests, Holm-adjusted p-values, and odds ratios (OR) with 95% confidence intervals (CI).

Results:

 While eGFR values were similar between groups, hypertensive patients had higher reti-eGFR and Dr.Noon 5-year CKD risk scores, as shown in Figure 1. The 5-year CKD risk score (OR: 4.19, 95% CI: 1.19-15.12) and reti-eGFR (OR: 3.70, 95% CI: 1.15-12.16) were strong predictors, with males showing higher risk than females (OR: 1.55, 95% CI: 1.08-2.23).

Conclusion: 

In DM patients without CKD, Dr.Noon CKD offers a non-invasive approach for CKD risk stratification, complementing eGFR in identifying risk differences.