Re-timing mortality risk through individualized modelling of lifespan trajectories.

Published in Nature biomedical engineering. Abstract: There is an urgent global need for affordable, large-scale health screening to enhance preventive care in the ageing population. Using low-cost, non-invasive retinal photographs and leveraging large foundation models, we developed a deep learning-based biomarker, Retinal Mortality Equivalent (RetiME), that directly mapped retinal features to individual mortality risk. Trained on a European cohort and externally validated across three independent external cohorts spanning Southeast Asia, Eurasia and North America, RetiME showed robust associations with all-cause and cause-specific mortality and predicted the incidence of a wide range of systemic and ocular diseases. Crucially, beyond static risk estimation, RetiME re-timed mortality risk by reconstructing personalized lifespan trajectories that were associated with observed mortality outcomes (for example, age of death). Clustering of these trajectories revealed distinct mortality patterns, spanning cancer-predominant, cardiometabolic and neurodegenerative risks, reflecting diverse ageing pathways captured by RetiME. Interpretability analyses showed that RetiME captures diverse biological and retinal features underlying mortality risk. Genome-wide analyses further identified genetic loci associated with RetiME. Collectively, these findings support further evaluation of retinal imaging and trajectory modelling as one component of a comprehensive assessment of mortality risk.