Chronic kidney disease: an AI that spots it in a photo of the eye

Why is kidney disease hard to spot?

Chronic kidney disease means the kidneys filter the blood less well, for months or years. At first it often causes no symptoms, so many people do not know they have it.

Finding it takes a blood test that measures how well the kidneys filter. In this study, the disease was defined as an estimated filtration rate below 60 mL per minute. Yet this kind of screening is hard to organize in the community and in family practice, even in wealthy countries.

The kidneys and the retina both contain very small blood vessels. The researchers' idea: could a photo of the eye signal a kidney problem?

How did the researchers go about it?

The team used three large studies of the general population in Singapore and China, covering people of different backgrounds. A deep learning AI, which learns by itself from examples, was trained on photos from 5,188 people, then checked on 1,297 others.

It was then tested on two independent groups: 3,735 people in Singapore and 1,538 people in Beijing.

The researchers compared three approaches: the AI reading the photo alone, a model based on simple risk factors (age, sex, ethnicity, diabetes and blood pressure), and a hybrid AI combining both.

What did they find?

Results are given as a score called the area under the curve, where 1 is a perfect result and 0.5 is no better than chance. In the validation group, the photo AI scored 0.911, the risk factor model 0.916 and the hybrid AI 0.938.

In the two test groups, the photo AI scored 0.733 in Singapore and 0.835 in Beijing. In those groups, the risk factor model did as well as it or better.

In people with diabetes or high blood pressure, who are most at risk of kidney disease, the scores stayed close to those of the validation group.

What does this mean for patients?

The authors see retinal photos as an extra, opportunistic tool: someone having an eye check could also get a signal about their kidneys, then be referred for a blood test.

This could help where blood tests are harder to reach. But it is not a test offered in routine care today: in practice, a blood test is still how kidney function is checked.

What are the limits?

The photo AI did less well in the test groups, especially in Singapore (0.733). And a simple model based on age, diabetes or blood pressure did as well or better, so the value of the photo on its own is still unproven.

The study is cross-sectional, meaning it looked at people at a single point in time. It covered Asian populations and did not test the tool in real care.