New AI model predicts future risk of diabetes-related blindness

Overview

A new AI model can predict up to three years ahead which people are at the highest risk of developing sight-threatening diabetic eye disease. It could save the NHS money by avoiding unnecessary screening for people at low risk, while preventing sight loss for those at high risk. The AI model was tested using HDR UK’s unique, diverse INSIGHT eye health dataset, to ensure that it works for everyone.

The challenge

In the UK, around 3.2 million people with diabetes attend the national diabetic eye screening programme every year. This is to check for signs of diabetic retinopathy (DR), a diabetes-related eye disease caused by raised blood sugar levels. DR can cause sight loss by damaging the small blood vessels in the back of the eye. Screening costs over £85 million per year in England alone. It involves taking pictures of people’s eyes, generating 13 million images each year, which then need to be checked for by 1,500 graders. Global diabetes rates are expected to rise by 51% between 2019 and 2045, which will put even more strain on DR screening programmes in future.

Previous studies, using blood tests to work out the risk of diabetes progression, have suggested that the vast majority of people who attend screening have a low risk of developing sight-threatening DR in the next two years. That means many people with diabetes could safely wait up to two years before their next screening appointment if it was possible to accurately distinguish between those at high and low risk.

The solution

King’s College London ophthalmologist Mr Paul Nderitu has led research to develop and test an AI system to predict people’s future risk of developing sight-threatening DR, using routinely collected images of the back of the eye. It’s the first ever non-invasive AI system in the UK for predicting sight-threatening DR up to one, two and three years ahead.

Mr Nderitu and colleagues developed three independent AI systems, one for each of the three prediction intervals. They trained the AI with images from 162,339 eyes from the South East London Diabetes Eye Screening programme. Next, they tested the AI systems on images from around 35,000 different eyes from the South East London Diabetes Eye Screening programme and the Birmingham Diabetes Eye Screening programme, which contributes data to HDR UK’s INSIGHT Hub. The AI systems accurately predicted the future development of sight-threatening DR at all three time intervals.

The impact

If it were to be used in national screening programmes, the AI systems – the first of their kind in the UK –-could save considerable NHS resources and money. Mr Nderitu explains:

“It could mean savings of about 20% for the screening costs, so around £15 million per year, and a saving of about 40% of appointments. So if you take the current 3.2 million screening appointments, you could potentially be saving over a million appointments every two years.”

Previous predictive models needed additional and invasive blood tests, whereas the new AI model only needs images that are collected during screening anyway, so it would be much more convenient for patients.

“Up to now there was no AI system developed in the UK that used non-invasive data, such as just images, to predict if someone’s going to develop sight threatening diabetic eye disease. AI systems developed elsewhere may not be applicable to the UK because we grade DR differently. Therefore, it was important that we developed a homegrown AI system that works for the UK,” explains Mr Nderitu.

HDR UK’s INSIGHT dataset was fundamental to the development of the AI system, and importantly it has contributed to making the AI system equitable because of the diversity of the data.

“There’s no other data source in the UK that had sequential pictures of people attending screening which was readily accessible – so HDR UK’s role was unique and very helpful in that respect. Another unique aspect about HDR UK was the diversity of the dataset, with a significant proportion of patients from South Asian communities of Birmingham. This was really important as it helped to show that our AI system worked as well as it did in the high Afro-Caribbean population we have here in London,” explains Mr Nderitu.

This research was a fruitful partnership between charities, hospitals and the NHS, with funding from Diabetes UK and King’s College Hospital charity and close links with NHS screening services in South London and Birmingham.