A new Australian study has found that an automated AI camera can accurately detect diabetic eye disease with more than 93% accuracy in non-eye care settings.

The study’s authors—Associate Professor Lisa Zhuoting Zhu and Sanil Joseph from the Center for Eye Research Australia and University of Melbourne, and Professor Mingguang He of the Hong Kong Polytechnic University—say their findings demonstrate the potential of AI eye screening to become part of routine clinical care for people with diabetes.

Globally, more than 529 million people are living with diabetes and at risk of vision loss and blindness from diabetic eye disease.

Early treatment can prevent blindness in 90% of cases, but ensuring that everyone with diabetes has access to the eye scans needed to detect the disease is a huge challenge for health systems worldwide.

Now the findings of a two-year Australian trial, published in the British Journal of Ophthalmology, show the potential of AI to increase access to sight-saving eye screenings.

More than 860 people with diabetes took part in the trial in the waiting rooms of GP and endocrinology clinics in Melbourne and an Aboriginal Health Service in Western Australia between August 2021 and June 2023.

The trial used an automated portable retinal camera—powered by an AI algorithm trained on more than 200,000 retinal images graded by 21 ophthalmologists—which allowed participants to take photos of their own eyes while they waited for their medical appointments.

Trial participants received a printout with a QR code linking to the results of their scan to take into their appointment, and those found to have signs of the disease were referred for follow-up with an eyecare specialist. To determine accuracy, all results were compared to the gold standard assessments of human grading. Participants and health professionals also took part in a satisfaction survey.

Although many studies have compared AI to human grading for diabetic eye disease, the Melbourne study is one of the first to occur in real world clinical settings.

The study found:

• A high accuracy rate of 93.3% compared to human grading
• 86% of participants were satisfied with the technology
• 85% of clinicians rated the technology highly.

The study identified areas for further improvement including:

• Need to improve image quality to reduce the number that could not be graded
• Ongoing development of the algorithm to reduce false negatives
• Improved follow-up to get more patients to act on referrals
• The need for targeted strategies in diverse communities.

“AI scans could be a great benefit in rural and remote areas where there is a shortage of trained eye care specialists,” says Dr. Zhu.

“It is also a cost saving for the health system, as it enables early screening to occur without the need for an eye care specialist for every patient.”

Sanil Joseph said AI-powered eye scans can also be more convenient for patients. “People with diabetes often have many medical appointments and prioritize appointments with other specialists over eye care. The AI scan enables them to combine their eye test with other medical visits.”

Provided by
Centre for Eye Research Australia