Diabetes-related retinopathy (DR) is a leading cause of visual loss in working-age populations, and early detection of DR is critical for improved visual outcomes. Yet, adherence to screening guidelines is only approximately 66% in the United States, and even lower in Africa and Asia, 36% and 51%, respectively.

Experts who have utilized artificial intelligence (AI)-enabled DR screening programs testified that these tools are cost effective and can improve adherence to screening guidelines during the symposium, Maximizing Impact—Cost-Effectiveness and Real-World Outcomes of AI-Enabled Diabetic Retinopathy Screening Programs, presented at the 85^th Scientific Sessions in Chicago.

One might initially assume that adopting new AI tools for DR screenings would be prohibitively expensive. While factors such as employing autonomous versus assistive AI, prevalence of disease, varying health care settings, the size of the health care system, or the accuracy of an AI system can variably impact associated costs, Roomasa Channa, MD, Associate Professor of Ophthalmology at the University of Wisconsin School of Medicine and Public Health, illustrated that AI-enabled DR screening programs can be not only cost effective but also cost saving.

A 2020 study from Singapore performed an economic analysis modeling study using three types of DR screenings:

1. A fully manual screening with humans only
2. An autonomous AI screening system with no human oversight
3. A hybrid model where the AI tool performed the initial analyses of the screenings and referred all concerning or inconclusive results to human graders

The hybrid model was the most cost effective, carrying an average expense of $62 per patient, compared to $77 per patient for the fully manual screenings and $66 per patient for the autonomous AI screenings.

“By 2050, Singapore is projected to have 1 million people with diabetes, and they estimated they would be saving $15 million if they switched to this human plus AI combination,” Dr. Channa said.

This research was undertaken to enhance a national screening program in Singapore. Colin S. Tan, MD, Associate Professor at the Duke-NUS Medical School, Singapore, Associate Professor at the Lee Kong Chian School of Medicine, Singapore, and Head of the Fundus Image Reading Center National Healthcare Group Eye Institute, Tan Tock Seng Hospital in Singapore, explained the results of the country’s screening program as a first-hand participant.

The Singapore Integrated Diabetic Retinopathy Programme (SiDRP) screens 120,000–200,000 individuals with diabetes each year across 23 primary care clinics. The program trained its SELENA+ AI system by feeding it nearly 500,000 retinal images to detect DR, glaucoma, and age-related macular degeneration. The AI model has displayed impressive sensitivity and specificity scores, with an area under the curve value of 0.936 for referable DR cases and 0.958 for those with vision-threatening DR.

SiDRP is currently operating fully with human grading, but Dr. Tan explained the program plans to transition to the two-tier human and AI hybrid model. With this updated model, AI will identify all normal or non-referable cases and pass the referable or inconclusive cases to human graders for final evaluation. Echoing Dr. Channa, Dr. Tan emphasized that cost effectiveness, in tandem with the system’s encouraging performance, was a key factor for pivoting to the new hybrid model.

“This has the potential to reduce unnecessary referrals to ophthalmologists and tertiary care, which can result in cost savings,” Dr. Tan said. With the semi-automated AI model, less than a quarter of cases were flagged for human review. “Only 23% of cases were flagged as potentially abnormal and needed a human to review, meaning that 77% of cases did not need a human grader. So, this is a potential reduction in the workload for human graders.”

Some data suggests that AI-enabled DR screening programs could also close gaps in care. Risa Wolf, MD, Associate Professor of Pediatric Endocrinology, and Director of the Pediatric Diabetes Program at Johns Hopkins University, explained that compliance with traditional DR screening exams is low.

The SEE study, led by Dr. Wolf, sought to determine the diagnostic efficacy of autonomous AI for eye exams in youth with diabetes. Using an autonomous AI system, adherence to DR screening guidelines improved from 49% to 95%.

The ACCESS trial, also led by Dr. Wolf, hypothesized that AI could improve screening completion compared with the current standard care. Participants were randomized in the trial to receive the usual care or AI-enabled screenings. Only 22% of patients in the former arm of the study completed their diabetes-related eye exam. Conversely, 100% of the latter AI group completed their exams.

“From this study, we concluded that AI does actually close care gaps,” Dr. Wolf said.

Juan Ding, OD, PhD, Assistant Professor in the Department of Ophthalmology & Visual Sciences at the University of Massachusetts (UMass) Chan Medical School, and Director of Optometric Services at the UMass Eye Center, echoed her co-presenters’ assertions that AI-enabled DR screenings are cost effective and can close screening gaps in the primary care setting. She also highlighted some of the challenges associated with adopting AI technologies into practice, such as high upfront costs, negotiations with payers, and data safety.

However, the primary hurdle to overcome relates to image quality, Dr. Ding explained.

“The main challenge is that there are frequently artifacts and blurry images,” she said. “So, we recommend that we should improve training for staff and maybe re-educate staff periodically.”

On-demand access to recorded presentations from this session will be available to registered participants of the 85^th Scientific Sessions through August 25.