Association between deep learning-derived retinal aging speed and rapid decline of renal function in the Japanese general population

Abstract Background The eye and the kidney share microvascular features and retinal changes have been postulated to reflect systemic cardiometabolic dysfunction. Recently, retinal age, estimated by deep learning (DL) from a single fundus photograph, has been shown to predict cardiovascular events. The present study aimed to determine whether retinal aging speed, defined as the individual-specific annual rate of change in DL-estimated retinal age, is associated with rapid estimated glomerular filtration rate (eGFR) decline. Methods We retrospectively analyzed the longitudinal data obtained from 7807 Japanese individuals who underwent repeated health check-ups with serial measurements of eGFR and fundus photographs (median 9.0 visits over median 8.0 years). We evaluated the relationship between retinal aging speed during the baseline period (from baseline to year 4) and rapid eGFR decline (≤−3 mL/min/1.73 m²/year) during the follow-up period (after year 4). Results The incidence of rapid eGFR decline was significantly higher in the highest quartile of retinal aging speed than in the lowest quartile as the reference group in the fully adjusted model adjusting age, sex, body mass index, blood pressure, diabetes, smoking, eGFR, and proteinuria (odds ratio (OR) 1.34; 95% confidence interval (CI) 1.06–1.68; P = 0.015). In the multivariable logistic regression models, retinal aging speed was associated with rapid eGFR decline after adjusting multiple key confounders (OR 1.14; 95% CI 1.02–1.29; P = 0.027). Conclusions Taken together, DL-derived retinal aging speed is independently associated with rapid eGFR decline in the Japanese general population.