Predicting renal function using fundus photography: role of confounders.

BACKGROUND/AIMS: The kidneys and retina are highly vascularized organs that frequently exhibit shared pathologies, with nephropathy often associated with retinopathy. Previous studies have successfully predicted estimated glomerular filtration rates (eGFRs) using fundus photographs. We evaluated the performance of the Modification of Diet in Renal Disease (MDRD) and Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formulas in eGFR prediction.

METHODS

We enrolled patients with fundus photographs and corresponding creatinine measurements taken on the same date. One photograph per eye was randomly selected, resulting in a final dataset of 45,108 patients (88,260 photographs). Data including sex, age, and blood creatinine levels were collected for eGFR calculation using the MDRD and CKD-EPI formulas. EfficientNet B3 models were used to predict each parameter.

RESULTS

Deep neural network models accurately predicted age and sex using fundus photographs. Sex was identified as a confounding variable in creatinine prediction. The MDRD formula was more susceptible to this confounding effect than the CKD-EPI formula. Notably, the CKD-EPI formula demonstrated superior performance compared to the MDRD formula (area under the curve 0.864 vs. 0.802).

CONCLUSION

Fundus photographs are a valuable tool for screening renal function using deep neural network models, demonstrating the role of noninvasive imaging in medical diagnostics. However, these models are susceptible to the influence of sex, a potential confounding factor. The CKD-EPI formula, less susceptible to sex bias, is recommended to obtain more reliable results.