A cross population study of retinal aging biomarkers with longitudinal pre-training and label distribution learning

Abstract

Retinal age has emerged as a promising biomarker of aging, offering a non-invasive and accessible assessment tool. We developed a deep learning model to estimate retinal age with enhanced accuracy, leveraging retinal images from diverse populations. Our approach integrates self-supervised learning to capture chronological information from both snapshot and sequential images, alongside a progressive label distribution learning module to model biological aging variability. Trained and validated on healthy cohorts (34,433 participants from the UK Biobank and three Chinese cohorts), the model achieved a mean absolute error of 2.79 years, surpassing previous methods. When applied to broader populations, analysis of the retinal age gap—the difference between retina-predicted and chronological age—revealed associations with increased risks of all-cause mortality and multiple age-related diseases. These findings highlight the potential of retinal age as a reliable biomarker for predicting survival and aging outcomes, supporting targeted risk management and precision health interventions.