Changes in the eye’s blood vessels can reveal something about age. Researchers analysing flicker-light responses from 140 healthy volunteers trained a machine-learning model that achieved a mean absolute age error of 9.3 years on 28 held-out participants. Individual predictions varied widely: the largest reported miss was roughly 37 years.[1]
The Scientific Reports study examines a dynamic signal. Retinal vessels widen and narrow as the eye responds to stimulation, providing information about vascular reactivity that a still photograph cannot capture. The question is how much that changing response adds to age estimation—and how consistently it works across people.
A living signal instead of a still photograph
Much work on age estimation from eyes has analysed static photographs or scans. This study instead looked at how arteries and veins changed when volunteers were exposed to flickering light. Each recording included an initial baseline followed by three 20-second stimulation periods separated by longer rests. The full protocol covered approximately 350 seconds. Researchers measured vessel diameters in retinal regions and extracted changing patterns across the sequence.[1]
The data came from a clinical study carried out in 2005, not a newly recruited population in 2026. The sample included 79 men and 61 women between about 21 and 70 years of age. It was then divided by participant into training, validation and test sets: 98, 14 and 28 people, respectively. Keeping individuals out of both training and testing is important because repeated readings from the same eye are not independent examples.
Models included conventional tree ensembles as well as recurrent and convolutional deep-learning approaches. On the test data, the strongest Extra Trees model had a mean absolute age error of 9.3 years and a root mean squared error of 12.4 years. Among the evaluated deep-learning models, the best ResNet result had a larger mean absolute error of 11.09 years. 'More sophisticated' did not automatically mean 'more accurate' on this small dataset.
What an average hides
A nine-year average error may be tolerable when exploring whether a physiological signal contains information, but it is substantial if interpreted as a prediction for a person. One subject aged 59.5 received a prediction of 59.4 years. Another, actually aged 45.9, was predicted to be 32.9. In the reported worst example, a 68.9-year-old was assigned an age of 31.8: an error of roughly 37 years.[1]
The investigators used LIME, a method for inspecting which features influenced a model's individual predictions, to explore the changes over time that appeared important. Such explanations can help locate influential inputs; they do not prove which molecular process produced them or guarantee that the reasons generalise to other people.
The model predicts chronological age: years since birth. Whether a person’s prediction error tells researchers anything useful about vascular health is a separate question requiring health-outcome data. A large error in either direction cannot currently be interpreted as accelerated or slower biological ageing.
The reason this modest result matters
The retina provides a rare window onto tiny living blood vessels without cutting into the body. If patterns of reactivity vary with age, they may eventually complement other measures used in vascular research. But this cohort was small, healthy and collected at one research setting. Conditions such as diabetes, hypertension or vascular disease could change the response for reasons unrelated to chronological age. External validation and repeated measurements are necessary before treating the result as clinically robust.[1]
It would also be valuable to compare this dynamic approach fairly with a simple statistical age baseline and conventional retinal photography under the same conditions. That would show whether the added several-minute flicker protocol contributes enough information to justify its extra complexity.
The next useful comparison is whether vessel motion adds reliable information beyond simpler measurements. That requires testing on new participants and reporting the spread of individual errors alongside the average. The present results identify a signal; the variability shows how far it is from a dependable personal assessment.
Sources & further reading
- Estimating chronological age from flicker-induced retinal vascular reactivity: a machine learning approach Scientific Reports · 22 Oct 2026 · Accessed 24 Oct 2026