Researchers have developed artificial intelligence (AI) algorithms that automated the screening process of remote eye tests for diabetic retinopathy. The team from Queen’s University Belfast, the Philippine Eye Research Institute and Joslin Diabetes Center of Harvard Medical School used images taken in the Philippines.
It was said the technology could prevent blindness while making huge cost-savings, particularly in less developed countries, and matched existing commercial algorithms. Researchers noted the introduction of diabetic retinopathy screening programmes in countries like the Philippines faced challenges, such as low resources, high population and lack of healthcare infrastructure.
They said the use of digital technology and portable, handheld retinal cameras operated by trained imagers was suited to addressing these issues. Images were captured on a retinal camera to check for any signs of diabetic eye disease and sent electronically to a centralised reading centre.
The images were analysed by trained graders and the results, including follow-up details and recommendations, were sent back to the patient and screening sites. However, the volume of images requiring analysis in a timely manner was still a huge challenge, researchers said, which is where AI could play a role.
Reducing the burden
The team of researchers created code-free automated machine learning models (AutoML) for diabetic retinopathy screening using images from handheld retinal cameras. They said the development of these AI algorithms could allow more images to be assessed, more accurately and at less cost.
Professor Tunde Peto, professor of clinical ophthalmology from the school of medicine, dentistry and biomedical sciences at Queen’s, said: ‘AutoML allows the development of code-free algorithms at minimal cost by individuals without extensive background in computer programming language.
‘As coding skills are not common among healthcare workers, the use of AutoML models for diabetic retinopathy screening can potentially address disparities in the delivery of eye care to patients with diabetic eye disease by ensuring a more rapid assessment of retinal images at point of care with minimal cost, thereby ensuring prompt referrals and timely intervention for those patients who require more specialised eye care.
‘Our research in the Philippines has shown that this is particularly useful in low-resource settings where there is a huge clinical burden of diabetes, a high cost of services and a lack of eye care human resources.
‘This work can help clinicians and healthcare managers in the planning and scaling-up of diabetic retinopathy screening programmes operations, with the ultimate goal of saving sight in people with
diabetes.’
The study included 17,829 de-identified retinal images from 3,566 eyes with diabetes, which were acquired using handheld retinal cameras in a community-based screening programme.
Alleviating demand and identifying pathology
AI has the potential to allay the escalating demand for eye care in hospitals, according to Professor Pearse Keane, professor of artificial medical intelligence at University College London (UCL) and medical retinal consultant at Moorfields Eye Hospital.
Keane highlighted a 33% increase in ophthalmology appointments over the past five years with nearly 10m appointments held each year while speaking on the Optical Suppliers Association stand at 100% Optical at London ExCeL on February 24-26.
‘We have the largest ophthalmic imaging resource in the world at Moorfields – larger than the combination of the top five US providers combined – plus we have gold standard governance for engagement and privacy,’ he said.
With reference to RETFound, the collaboration between UCL and Moorfields, which utilises 1.6m retinal images to assist early research into progressing to a functioning algorithm, Keane said he was proud that AI was at the forefront of AI in medicine, but cautioned that we were still at the early stages.
‘The most important application in the short term is, perhaps not in direct patient care, but in clinical trial planning. We can help to overcome the challenges of recruitment by sending an algorithm to clinics to identify segment pathology to find patients suitable to be approached to participate,’ he said.
More than 500 medical AI systems are in the US Food and Drug Administration approval system and using retinal scan data to track patients for early signs of systemic disease is progressing, said Keane, but insurance, infrastructure, payment systems and formatting governance were still ‘in the melting pot’.