Can AI Aid in Preventing Diabetes-Related Vision Loss?

This article is the second installment of a six-part series exploring the impact of artificial intelligence (AI) on medical research and therapeutic approaches. Terry Quinn received a diabetes diagnosis during his teenage years. He resisted the diagnosis and the regular medical examinations, disliking the feeling of being distinct. His primary concern was the potential necessity of a foot amputation in the future. Vision impairment, another potential consequence of diabetes, was not a significant worry for him. Quinn, a resident of West Yorkshire, states, “I never thought I’d lose my sight.” However, he later observed bleeding in his eye. Medical professionals informed him he had diabetic retinopathy, which is damage to the blood vessels in the retinas caused by diabetes. This condition necessitated laser procedures followed by injections. Ultimately, these treatments proved insufficient to halt the decline of his eyesight. He sustained shoulder injuries by colliding with lampposts, struggled to recognize his son’s face, and was forced to cease driving. He recalls, “I felt pathetic. I felt like this shadow of a man that couldn’t do anything.” A key factor in overcoming his despondency was the assistance from the Guide Dogs for the Blind Association, which paired him with a black Labrador named Spencer. Quinn, who currently works as a fundraiser for Guide Dogs, asserts, “He saved my life.” In the UK, the NHS offers diabetic eye screening to patients every one or two years. US recommendations stipulate that all adults diagnosed with type 2 diabetes should undergo screening at the time of diagnosis, and subsequently on an annual basis if no problems are identified. Nevertheless, for numerous individuals, this practice is not consistently followed. Roomasa Channa, a retina specialist at the University of Wisconsin-Madison in the US, states, “There’s very clear evidence that screening prevents vision loss.” In the US, obstacles include expenses, communication challenges, and accessibility. Dr. Channa is of the opinion that improving access to these tests would benefit patients. For diabetic retinopathy screening, healthcare professionals capture images of the eye’s posterior internal surface, referred to as the fundus. Presently, the manual interpretation of fundus images constitutes “a lot of repetitive work,” according to Dr. Channa. However, some experts suggest that artificial intelligence (AI) has the potential to accelerate this process and reduce its cost. Diabetic retinopathy progresses through distinct stages, making it suitable for AI training to detect. In certain situations, AI might determine the necessity of a referral to an eye specialist or collaborate with human image evaluators. An example of such a system was created by Retmarker, a health technology firm located in Portugal. Their system pinpoints potentially problematic fundus images and forwards them to a human expert for additional examination. João Diogo Ramos, Retmarker’s chief executive, explains, “Normally we use it more as a support tool to give information to the human to make a decision.” He contends that resistance to change is hindering the adoption of AI-driven diagnostic instruments of this nature. Independent research indicates that systems such as Retmarker Screening and Eyenuk’s EyeArt demonstrate satisfactory levels of sensitivity and specificity. Sensitivity refers to a test’s effectiveness in identifying a disease, whereas specificity measures its accuracy in confirming the absence of a disease. Typically, exceptionally high sensitivity might correlate with an increased number of false positives. False positives generate both apprehension and financial costs, as they result in unneeded appointments with specialists. Generally, substandard image quality can contribute to false positives in AI systems. Researchers at Google Health have been investigating the limitations of an AI system they designed for detecting diabetic retinopathy. Its performance during trials in Thailand varied significantly from its behavior in theoretical situations. A particular issue was the algorithm’s need for immaculate fundus images, which contrasted sharply with real-world conditions involving sometimes unclean lenses, inconsistent illumination, and camera operators with diverse skill levels. The researchers report having gained insights into the significance of utilizing superior data and engaging with a diverse group of individuals. Google expresses sufficient confidence in its model that in October, the corporation declared its intention to license it to collaborators in Thailand and India. Google additionally stated its collaboration with the Thai Ministry of Public Health to evaluate the tool’s economic efficiency. The financial outlay represents a crucial element of this novel technology. Mr. Ramos indicates that Retmarker’s service might be priced at approximately €5 per screening, although this figure can fluctuate based on volume and geographical area. In the US, medical billing codes are established at significantly greater amounts. In Singapore, Daniel S W Ting and his associates conducted a cost comparison of three distinct diabetic retinopathy screening approaches. Human assessment proved to be the most costly. Nevertheless, complete automation did not emerge as the least expensive option due to a higher incidence of false positives. The most economical solution was a hybrid approach, where AI conducted the preliminary filtering of results before human intervention. This particular model has now been incorporated into the Singapore Health Service’s national IT infrastructure and is scheduled for implementation in 2025. Nonetheless, Professor Ting suggests that Singapore’s ability to realize cost efficiencies stems from its pre-existing robust infrastructure for diabetic retinopathy screening. Consequently, the economic viability is anticipated to show considerable variation. Bilal Mateen, chief AI officer at the health NGO PATH, observes that cost-effectiveness data concerning AI instruments for vision preservation has shown promising results in affluent nations such as the UK, and in a limited number of middle-income countries like China. However, this trend does not extend to the global remainder. Dr. Mateen emphasizes, “With the rapid advances in what AI is capable of doing, we need to ask less if it’s possible, but more and more whether we’re building for everyone or just the privileged few. We need more than just effectiveness data for effective decision-making.” Dr. Channa highlights the disparity in health equity, even within the US, a gap she hopes this technology can help close. She states, “We do need to expand it to places that have even more limited access to eye care.” She further underscores the importance of older individuals and those with visual impairments consulting eye doctors, and that the ease of AI in routine detection of diabetic eye disease should not divert focus from other ocular conditions. Other eye ailments, such as myopia and glaucoma, have presented greater challenges for AI algorithms to identify. Despite these qualifications, Dr. Channa remarks, “the technology is very exciting.” She adds, “I would love to see all our patients with diabetes screened in a timely fashion. And I think given the burden of diabetes, this is a really potentially great solution.” In Yorkshire, Mr. Quinn undoubtedly wishes for the widespread adoption of this new technology. He states that if AI had been available for earlier detection of his diabetic retinopathy, “I’d have grabbed it with both hands.”