Disentangling physiological heterogeneity in retinal aging using a deep learning-based biological age framework

Biological age estimators quantify aging-related variation but provide limited insight into organ-specific aging processes. The retina enables non-invasive visualization of microvascular and neural structures and has emerged as a promising modality for biological age prediction. However, existing retinal aging models typically produce unidimensional age estimates with limited interpretability. Here we develop a deep learning framework based on a large-scale vision foundation model to estimate retinal biological age from fundus images and to characterize the physiological heterogeneity underlying retinal aging. Using a reference cohort of 56,019 relatively healthy individuals, the model achieved a Mean Absolute Error of 2.48 years in age prediction. Analysis of age deviations in a real-world clinical cohort (n = 46,369) revealed non-linear associations with cardiometabolic risk and population heterogeneity in aging patterns. Integrating multidimensional physiological profiling, feature attribution and unsupervised analysis, we identified distinct retinal aging signatures associated with systemic inflammation and hemodynamic variation. To further characterize age-related deviations, we introduced a residual learning framework that decomposes retinal aging signals into a normative age-related component and additional components associated with physiological variation, achieving a Mean Absolute Error of 1.80 years on the independent healthy test set. This approach provides an interpretable representation of retinal aging and a framework for studying organ-level aging processes and their relationship to systemic health using large-scale imaging data.