A research team at Severance Hospital has developed artificial intelligence (AI) software that can predict the risk of chronic kidney disease (CKD) based on retinal examination results.

A research team, led by Professors Joo Young-soo (left) at Yongin Severance Hospital and Park Jung-tak at Severance Hospital, developed an AI program to predict the risk of CKD based on retinal examination.

CKD is usually diagnosed when a person has persistent decreased kidney function or proteinuria for more than six months, with the main cause of the disease being diabetes and high blood pressure.

CKD is also often called a silent disease, as patients normally do not show symptoms until it has progressed severely. However, as there is no cure for this disease, doctors agree that prevention is crucial.

Until now, blood tests have been used to estimate glomerular filtration rate to assess CKD risks, but such a method had limitations as the glomerular filtration rate is easily affected by external factors such as age and exercise, making it difficult to measure risks in people with normal kidney function.

As a result, the team, led by Professors Joo Young-soo at Yongin Severance Hospital and Park Jung-tak at Severance Hospital, developed an AI program to predict the risk of CKD based on retinal examination and analyzed its effectiveness.

Mediwhale, a local company specializing in AI diagnostic solutions, also participated in the study.

The team came up with the idea that the retina is composed of microvessels like the kidneys and that it can be easily examined through a fundus examination and the fact that patients with diabetes, a major cause of CKD, also periodically undergo regular retinal examinations to check for complications.

The team first trained the AI with data on retinal examinations and glomerular filtration rates of 80,000 people who were examined at Severance Hospital. And then, they created an algorithm to analyze the risk of developing CKD.

The AI looked at the blood vessels in the retinal images to predict a decrease in glomerular filtration rate and assess the risk of developing CKD.

Afterward, the team validated the efficacy of the AI's predictions on a combined sample of 35,000 people with diabetes from the U.K. Biobank study and patients from Korea.

The subjects' retinal photographs were examined to determine whether they developed CKD over a period of up to 10.8 years.

As a result, the AI software showed higher accuracy than the glomerular filtration rate estimation method.

"It was difficult to predict the possible occurrence of CKD if the kidney function is normal," Professor Park said. "The AI developed by the team creates an opportunity to screen high-risk groups for CKD and prevent the disease."

The results of the research were published in the npj Digital Medicine.