by King's College London

Deep learning system training schema for the 2 year prediction interval. Credit: Communications Medicine (2024). DOI: 10.1038/s43856-024-00590-z

Researchers at King's College London have used anonymized NHS eye data from more than 100,000 people with diabetes to build an AI model that can accurately predict who is at a high risk of developing sight threatening diabetic retinopathy (DR) up to three years in advance.

The study is published in the journal Communications Medicine.

DR is an eye condition that affects about 1 in 3 people with diabetes, and a leading cause of vision loss in working-age adults.

Anyone aged 12 years or older diagnosed with diabetes under the NHS is asked to attend an annual eye check for DR with the NHS Diabetic Eye Screening Program (DESP), which screens around 3.2 million people at a cost of over £85 million pounds per year in England alone.

The AI model could serve as a tool to enable individualized DESP screening by predicting if an individual is at low or high risk of developing sight threatening DR in one year, two years or three years using images from the back of the eye—a capability that the NHS does not currently have.

In developing the AI model, the researchers used more than 1.2 million retinal images from people with diabetes from South East London DESP. To ensure that the model was robust enough to work accurately on a diverse range of individuals, it was validated on a dataset of around 70,000 images from the INSIGHT Health Data Research Hub.

"This project shows the enormous value of collating, curating and sharing routinely collected clinical data. It also shows the way ahead, in terms of how AI might move from hype to tangible patient benefit, subject to further clinical studies," says Professor Timothy Jackson PhD.

"It has been a pleasure to work in this clinically driven project that leveraged the cross-cutting capabilities of the Faculty of Life Sciences & Medicine at King's. Using Artificial Intelligence to predict the incidence of retinal disease due to diabetes has tremendous social and economic prospects.

"AI can modernize diabetic screening programs without sacrificing their current ability to prevent sight loss," says Professor Christos Bergeles.

If implemented, individualized screening would reduce the screening burden for people at low risk of vision loss, while ensuring individuals at a high-risk of vision loss are seen urgently. This approach could save the NHS millions of pounds, and thousands of appointments every year.

More information: Paul Nderitu et al, Predicting 1, 2 and 3 year emergent referable diabetic retinopathy and maculopathy using deep learning, Communications Medicine (2024). DOI: 10.1038/s43856-024-00590-z

Journal information: Communications Medicine 

Provided by King's College London