Can it help prevent blindness in diabetics?

This is the second feature in a six-part series examining how AI will transform medical research and treatment.

Terry Quinn was just a teenager when he was diagnosed with diabetes. In a way, he rebelled against labels and frequent testing, not wanting to feel different.

His biggest fear was that his leg would one day have to be amputated. Vision loss, another complication of diabetes, was never on his radar. “I never thought I would lose my sight,” says Quinn, who lives in West Yorkshire.

But one day I noticed that my eyes were bleeding. Doctors told him he had diabetic retinopathy, damage to blood vessels in the retina associated with diabetes. This required laser treatment followed by injections.

In the end, treatment was not enough to prevent vision loss. He hurt his shoulder when he hit a lamppost. He could not recognize his son’s face. And he had to give up driving.

“I felt pathetic. I felt like a shadow of a man who couldn’t do anything,” he recalls.

It was the support of Guide Dogs for the Blind that helped him rise from despair. They connected him with a black Labrador named Spencer. “He saved my life,” says Quinn, who now raises money for guide dogs.

In the UK, the NHS recommends that patients have a diabetic eye exam every 1-2 years.

U.S. guidelines say all adults with type 2 diabetes should be tested at the time of diabetes diagnosis and annually if there are no problems. But for many people, that doesn’t really happen.

“There is clear evidence that screening prevents vision loss,” said Rumasa Channa, a retina specialist at the University of Wisconsin-Madison.

In the United States, barriers include cost, communication, and convenience. Dr. Channa believes that making testing easier to access will help patients.

To screen for diabetic retinopathy, your health care professional will take a photo of the back inner wall of your eye, known as the fundus.

Currently, manually interpreting fundus images is “a lot of repetitive work,” says Dr. Channa.

But some believe that artificial intelligence (AI) could speed up the process and make it cheaper.

This means that AI can be trained to recognize diabetic retinopathy, as it develops in fairly distinct stages.

In some cases, AI could decide whether a referral to an eye specialist is needed or work in conjunction with a human image grader.

One such system was developed by Retmark, a health technology company based in Portugal.

Its system identifies potentially problematic fundus images and sends them to human experts for further investigation.

“Typically, we use it as a support tool to give humans information to make decisions,” says João Diogo Ramos, CEO of Retmark.

He believes fear of change is limiting the adoption of such AI-powered diagnostic tools.

Independent studies suggest that systems such as Retmarker Screening and Eyenuk’s EyeArt have acceptable sensitivity and specificity.

Sensitivity tells you how good a test is at detecting the disease, and specificity shows how good it is at detecting the absence of the disease.

In general, very high sensitivity can increase false positives. False positives can lead to unnecessary specialist visits, causing both anxiety and expense. In general, poor image quality can lead to false positives in AI systems.

Google Health researchers have been investigating weaknesses in the AI ​​system they developed to detect diabetic retinopathy.

Performance during the tie trials was significantly different compared to the hypothetical scenario.

One problem is that the algorithm requires an initial fundus image. This was a far cry from the reality of occasionally dirty lenses, unpredictable lighting, and photographers with varying levels of training.

Researchers say they have learned lessons about the importance of working with better data and consulting a wider range of people.

Google is so confident in its model that it announced in October that it would license it to partners in Thailand and India. Google also said it is working with Thailand’s Ministry of Public Health to assess the cost-effectiveness of the tool.

Cost is a very important aspect of any new technology.

Ramos said Retmarker’s service fee could be around 5 euros per screening, but would vary depending on volume and location. In the United States, medical billing codes are quite high.

Daniel SW Ting and colleagues compared the costs of three models of diabetic retinopathy screening in Singapore.

Human evaluation was the most costly. However, full automation was not the cheapest because it had more false positives.

The most affordable was a hybrid model where the initial filtering of results was performed by AI before a human took over.

The model is currently integrated into the Singapore Health Service’s national IT platform and is expected to go live in 2025.

However, Professor Ting believes that Singapore was able to achieve cost savings because it already had a strong infrastructure for diabetic retinopathy testing.

Therefore, cost-effectiveness can vary widely.

Bilal Mateen, chief AI officer at health NGO Path, said cost-effectiveness data on AI tools to preserve vision are very limited in rich countries like the UK and a few middle-income countries like China. It is said to be strong. However, this is not the case in other parts of the world.

“With rapid advances in what AI can do, we are not asking whether it is possible, but rather whether we are building it for everyone or for a privileged few. Increasingly, we need to ask: Effective decision-making requires more than just efficacy data,” Dr. Mateen urges.

Dr. Channa points out that there are health disparities within the United States, and hopes this technology can help close those gaps. “We need to extend it to places where access to eye care is even more limited.”

He also said that older people and people with vision problems should see an eye doctor, and that the convenience of AI to regularly detect diabetic eye disease will help ensure that all other eye diseases are taken care of. It emphasizes that it should not be hindered. Other eye diseases, such as myopia and glaucoma, have proven difficult to detect with AI algorithms.

But even with these caveats, “this technology is very exciting,” Dr. Channa says.

“I hope all people with diabetes can get tested in a timely manner. Given the burden of diabetes, I think this could be a really great solution.”

Back in Yorkshire, Mr Quinn is certainly hopeful that the new technology will catch on.

If AI had existed to detect his diabetic retinopathy early, “I would have grabbed it with both hands.”