Investigating the fundamental characteristics of retinal age models

Retinal age is a biological ageing marker estimated from retinal images. Its deviation from chronological age, termed the retinal age gap (RAG), has been associated with adverse health outcomes. However, fundamental characteristics of retinal age models that are critical for interpreting and applying RAG remain underexplored, including the consistency of these associations across age subgroups as well as model generalisability across different cohorts. In this study, we used 327,764 retinal images to develop a retinal age model and investigate associations between RAG and systemic diseases. We examined association consistency across age subgroups and evaluated model performance on three external datasets with distinct imaging devices and populations. RAG showed positive associations with systemic diseases in young and middle-aged subgroups, but negative associations in the aged subgroup, revealing substantial age-dependent inconsistency. We further showed that this inconsistency was explained by regression-to-the-mean effects which varied by health status. In external evaluations, the generalisability of RAG varied considerably by cohort and intended application. Large age estimation errors did not necessarily indicate limited clinical utility, instead suggesting that model generalisability should be defined and evaluated in an application-specific manner rather than based on age estimation accuracy alone. Overall, this study investigates key characteristics of retinal age models that affect the interpretation of RAG and its disease associations. Our findings highlight the importance of reporting age-dependent associations, and emphasise the need for multi-dimensional, application-specific external evaluation for reliable use of retinal age and RAG in future clinical applications. More broadly, our findings and practical recommendations may extend to wider biological age estimation models.