Could AI help prevent diabetes-related sight loss?

This is the second installment in a six-part series exploring how AI is transforming medical research and treatments.

Terry Quinn was a teenager when he was diagnosed with diabetes. Initially, he resisted frequent testing, not wanting to feel different.

While he feared potential foot amputation, vision loss wasn’t something he worried about. “I never thought I’d lose my sight,” says Quinn, who lives in West Yorkshire.

However, one day, he noticed bleeding in his eye. Doctors diagnosed him with diabetic retinopathy, a condition caused by diabetes-related damage to retinal blood vessels. Despite undergoing laser treatments and injections, his vision continued to decline. Quinn struggled with daily life—walking into lampposts, losing the ability to see his son’s face, and having to stop driving.

“I felt pathetic, like a shadow of myself,” he recalls.

Support from the Guide Dogs for the Blind Association helped him recover emotionally. Being paired with Spencer, a black Labrador, was transformative. “He saved my life,” says Quinn, who now fundraises for the organization.

In the UK, the NHS recommends diabetic eye screenings every 1-2 years. In the US, adults with type 2 diabetes should be screened at diagnosis and annually thereafter, but many miss these checkups.

“There’s strong evidence that screening prevents vision loss,” says Roomasa Channa, a retina specialist at the University of Wisconsin-Madison.

In the US, barriers like cost, communication, and accessibility hinder regular screenings. Dr. Channa suggests improving test accessibility would encourage more patients to participate.

Diabetic retinopathy screenings involve photographing the eye’s fundus (the rear interior surface). Currently, manually analyzing these images is labor-intensive and repetitive, but AI may streamline the process.

AI can be trained to detect diabetic retinopathy, which progresses in predictable stages. It could help determine if a referral to an eye specialist is needed or assist human image graders, making the process faster and more cost-effective.

Diabetes patients are advised to undergo eye scans every one to two years.

One such system, developed by Portuguese health technology company Retmarker, flags potentially problematic fundus images for review by human experts.

“We primarily use it as a support tool to help humans make decisions,” says João Diogo Ramos, Retmarker’s CEO. He believes that hesitation to embrace change is slowing the adoption of AI-driven diagnostic tools.

Independent studies show that systems like Retmarker Screening and Eyenuk’s EyeArt have acceptable sensitivity and specificity rates.

Sensitivity measures a test’s ability to detect disease, while specificity measures its ability to confirm the absence of disease.

However, higher sensitivity can increase false positives, leading to unnecessary specialist visits, added costs, and patient anxiety. Poor-quality images are a common cause of false positives in AI systems.

AI can be trained to analyze images of the fundus, the eye’s rear wall.

Google Health researchers have identified weaknesses in their AI system designed to detect diabetic retinopathy, particularly during trials in Thailand, where real-world conditions differed from ideal scenarios.

The algorithm struggled with less-than-perfect fundus images, which were affected by dirty lenses, inconsistent lighting, and varying levels of operator expertise. Researchers emphasized the importance of using better data and involving diverse stakeholders during development.

Despite these challenges, Google remains confident in its model. In October, the company licensed it to partners in Thailand and India and is collaborating with Thailand’s Ministry of Public Health to evaluate its cost-effectiveness, a critical factor in AI adoption.

João Diogo Ramos of Retmarker estimates their screening service could cost around €5 per test, depending on location and volume. In contrast, US medical billing codes set screening costs significantly higher.

In Singapore, Daniel S W Ting and his team analyzed the costs of three diabetic retinopathy screening approaches. Human assessment was the most expensive, while fully automated AI was not the cheapest due to its higher rate of false positives.

The most cost-effective solution was a hybrid model where AI performed initial screenings, followed by human review. This approach is now integrated into Singapore Health Service’s national IT system and is set to launch in 2025.

Prof. Ting credits Singapore’s existing strong screening infrastructure for enabling these cost savings.

Bilal Mateen argues that medical AI should be accessible in countries beyond wealthy nations.

The cost-effectiveness of AI tools in healthcare is likely to vary significantly across regions.

Bilal Mateen, chief AI officer at the health NGO PATH, points out that while cost-effectiveness data for AI tools to prevent vision loss is strong in wealthy nations like the UK and some middle-income countries like China, this is not true for much of the world.

“With rapid advancements in AI, the question should shift from whether it’s possible to whether we’re building solutions for everyone or just the privileged few,” Dr. Mateen emphasizes. He argues that we need more than just effectiveness data to make informed decisions.

Dr. Roomasa Channa highlights the health equity gap even within the US and hopes that AI can help bridge it. “We need to expand it to areas with limited access to eye care,” she says. However, she also stresses that older adults and those with vision issues should still see eye doctors regularly. While AI is effective at detecting diabetic eye disease, other eye conditions, such as myopia and glaucoma, remain challenging for AI to diagnose.

Despite these challenges, Dr. Channa is optimistic about the potential of AI. “I would love to see all our patients with diabetes screened in a timely manner. Given the burden of diabetes, this is a potentially great solution,” she says.

In Yorkshire, Terry Quinn shares the same hope. Reflecting on his own experience, he says, “If AI had been available to detect my diabetic retinopathy earlier, I’d have grabbed it with both hands.”