Slow Start for AI-Assisted Screening for Diabetic Retinopathy Despite Promise

While artificial intelligence (AI)-assisted detection of diabetic retinopathy has the potential to expand access to screening, few patients with diabetes have been evaluated with the technology since it received FDA approval in 2021, a retrospective cohort study suggested.

AI imaging was used in 2.2% since 2021 based on reimbursement codes, indicating that this new modality for diabetic retinopathy detection was used in only 0.09% of all patients with diabetes, reported Ravi Parikh, MD, MPH, of Manhattan Retina and Eye Consultants in New York City, and colleagues in JAMA Ophthalmology.

AI imaging was used 58.0 times per 100,000 patients in 2021, which slightly increased by 1% to 58.6 times per 100,000 patients by 2023.

Meanwhile, optical coherence tomography (OCT) imaging was used in 80.3% of patients, fundus photography in 35%, and traditional remote imaging (tracked with two different codes) in 1% and 2.5%, respectively.

Although use of traditional remote imaging increased by 185.4% from 2021 to 2023, AI imaging had a higher referral rate (7.74%) to OCT imaging compared with traditional remote imaging (5.53%). Use of all remote imaging modalities increased by 90.16% from 2021 to 2023 (P<0.001).

“In some ways, it’s a little disappointing that AI use is flat,” Parikh told MedPage Today. “But at the same time, use of remote imaging is increasing a lot. There’s momentum building toward this idea of re-engineering how we identify patients with diabetic retinopathy.”

According to Parikh, diabetic retinopathy is the leading cause of blindness and severe vision loss among working-age people in the U.S. Treatments such as laser therapy, surgery, and anti-VEGF drugs can prevent about 90% of severe vision loss, he said.

Annual vision screenings are recommended for patients with diabetes, but only 66% of patients were estimated to have been screened within the past 12 months in 2023. Other research has suggested that true screening rates are much lower, said Rithwick Rajagopal, MD, PhD, of Washington University School of Medicine in St. Louis, who was not involved in the study.

“Screening can be done via traditional dilated fundus exam, aka ophthalmoscopy, or via retinal photographs, interpreted by a human or one of several approved computer algorithms,” Rajagopal told MedPage Today.

“AI-based point-of-care screening offers a powerful solution to the longstanding problem of low adherence to recommended retinal screening guidelines,” he said. “There are multiple causes of such poor adherence, but point-of-care screening solves several of them: no need to take time off for an additional medical visit, no additional co-pay for eye doctor visits, and no need for dilation in many cases. Numerous studies have validated their screening efficacy, and some studies, such as our own from a few years ago, have shown that they markedly improve diabetic retinopathy screening rates and rates of follow-up care.”

Parikh said that AI-assisted imaging has been shown to be “effective and practical,” and it can be used in primary care clinics for earlier detection.

While the study shows that AI isn’t widely used nationally, it does appear to be helping to reduce disparities in screening. More than 80% of those who received AI imaging were from the South, a region that comprised only 40% of other imaging modalities, and almost half of the patients who received AI imaging were Black, compared with about a quarter in other imaging modalities.

“It’s a perfect example of how individuals who otherwise might not be in the tent at all could potentially come in,” Parikh said.

Moving forward, he noted, “AI imaging needs more investment from the government and higher reimbursement to remain economically feasible for clinics.”

Rajagopal said uptake of the technology has been slow for several reasons. For one, cameras used in AI-assisted screening are expensive, and “costs may not be recovered in a reasonable timeframe due to low reimbursement rates.”

Also, there’s “unfamiliarity with eye anatomy and physiology among non-ophthalmologists, which is associated with low level of comfort in capturing the photographs and interpreting the results, even though the cameras are increasingly easy to use and the AI software generates the diagnosis,” he added.

For this study, Parikh and colleagues used the TriNetX database, which includes more than 107 million patients across 62 healthcare organizations in the U.S. Using the records of 4.95 million patients with diabetes from January 2019 to December 2023, they found that 4.2% of patients received at least one of the targeted reimbursement codes. Mean patient age was 64 years, and 48% were women.

The researchers did not include dilated fundus exams, which may lead to referrals to more advanced screening.

Please enable JavaScript to view the

comments powered by Disqus.