Predicting renal function using fundus photography: role of confounders.

Area of Science:

• Ophthalmology
• Nephrology
• Artificial Intelligence

Background:

• Kidneys and retina share pathologies; kidney disease (nephropathy) often links to retinopathy.
• Previous research shows fundus photos can predict estimated glomerular filtration rates (eGFRs).

Purpose of the Study:

• Evaluate the performance of MDRD and CKD-EPI formulas for eGFR prediction using fundus photographs.
• Assess the utility of deep neural networks in noninvasive renal function screening.

Main Methods:

• Utilized a dataset of 45,108 patients with paired fundus photographs and creatinine measurements.
• Employed EfficientNet B3 models to predict age, sex, and creatinine levels from fundus images.
• Calculated eGFR using both MDRD and CKD-EPI formulas.

Main Results:

• Deep neural networks accurately predicted patient age and sex from fundus images.
• Sex was identified as a confounding variable affecting creatinine prediction, with MDRD being more susceptible.
• The CKD-EPI formula outperformed the MDRD formula in eGFR prediction (AUC 0.864 vs. 0.802).

Conclusions:

• Fundus photography, analyzed by deep neural networks, offers a valuable noninvasive method for renal function screening.
• The CKD-EPI formula is recommended over MDRD for more reliable eGFR estimation due to lower susceptibility to sex-related confounding.
• Acknowledges the influence of sex as a confounding factor in AI-driven diagnostic models.