RetiME uses retinal photographs to estimate how mortality risk changes with age

Researchers have trained an artificial-intelligence model to turn an ordinary photograph of the back of the eye into a sequence of estimated mortality risks across a person's future years. Their model, RetiME, is reported in Nature Biomedical Engineering on November 2. Its important innovation is not the familiar claim that eyes reflect general health. Instead of offering only a biological-age score, it estimates how risk might unfold over time, then examines whether different predicted patterns correspond to different disease histories. Those curves are population-informed probabilities, not a forecast of the date on which an individual will die.[1]

The team began with UK Biobank retinal images linked to medical histories and mortality records. A retinal foundation model was adapted to estimate probabilities of death for successive ages. The researchers then used survival-modelling methods to reconstruct personal risk trajectories. That differs from training a classifier to say simply whether someone dies during a fixed follow-up period: it creates a shape that can be compared with the timing of observed events and grouped with similar trajectories.[1]

Why a photograph might contain information beyond the eye

Blood vessels and other retinal structures are affected by processes that also influence systemic health. In the study, RetiME associations extended beyond eye disorders to cardiovascular, neurological and other conditions. The authors compared the signal with other ageing indicators such as telomere length and hand-grip strength in their analysis. The clinical question is whether a retinal score adds useful risk information beyond existing measurements across different settings.[1]

The paper reports external testing in three independent cohorts spanning Southeast Asia, Eurasia and North America, in addition to the European training context. Those datasets matter because cameras, populations, care systems and background risks differ. External validation makes a finding stronger than an impressive internal test alone, but it does not automatically establish performance in every ethnicity, age range, hospital or consumer camera. Patients are not randomly assigned risk levels, so retrospective association cannot show that using the score improves their health.[1]

The model's 're-timing' step deserves particular care. A conventional risk score might rank two people as relatively high and low risk. A full trajectory aims to represent how risk changes with age and whether its probability distribution peaks earlier or later. The authors report that predicted trajectories showed concordance with observed mortality outcomes, and clusters corresponded to broad disease-risk patterns including cancer, cardiometabolic illness and neurodegeneration. These clusters describe statistical groupings, not diagnoses of cancer or dementia for a particular person.[1]

What the five trajectory clusters represent

The trajectory analysis groups modelled curves into five clusters with different disease associations. These are statistical groupings produced by the analysis, and people can have overlapping risks. Their interpretation depends on the model, the recorded events and the assumptions used to estimate survival over time.[1]

There are unusually practical reproducibility details. The authors identify UK Biobank data-access permissions and state that several external datasets require institutional or ethical approval rather than unrestricted download. They have released code through a public repository. The article also discloses that several authors are inventors on a provisional patent application covering the RetiME method, while one senior investigator reports commercial links to eye-AI businesses. Such disclosures do not invalidate results, but they matter when assessing future product claims and independent deployment studies.[1][2]

A tempting application would be mass screening for people unlikely to receive sophisticated blood tests. That remains a future evaluation question. To establish benefit, a prospective study would need to show that the information changes an appropriate clinical action and improves outcomes beyond existing risk factors without provoking disproportionate unnecessary investigations. The access, cost and follow-up pathway matter as much as image acquisition.[1]

Report an error or suggest a correction