AI Algorithm May Help Identify Eligibility for Diabetic Retinopathy Clinical Trials

Published on: 

New data from AAO 2022 show how technology may refine a currently inefficient enrollment practice for retina disease assessment.

Using an artificial intelligence (AI) tool to prescreen potentially eligible participants for diabetic retinopathy (DR) may improve the enrollment rate and efficiency of clinical trials for the disease space.

In new data presented at the American Academy of Ophthalmology (AAO) 2022 Annual Meeting in Chicago this week, investigators observed a high accuracy associated with an AI model implemented into a tiered strategy for DR clinical trial screening. The addition of AI to screening strategies may be beneficial to the cost and labor of physicians and patients alike going forward.

Investigators from the A-EYE Research Unit and Wisconsin Reading Center in the department of ophthalmology at University of Wisconsin, led by Amitha Domalpally, MD, PhD, sought to interpret whether a US Food and Drug Administration (FDA)-regulated AI assist retina specialists in screening for clinical trial inclusion criteria.

As they noted, current DR clinical trial screening criteria warrants patients grade from 47-53 ETDRS letters, as well as grade on a 7-field stereoscopic imaging protocol. However, current criteria are imperfect; Wisconsin Reading Center alone receives approximately 50% failed screenings for DR clinical trials.

Domalpally and colleagues used an AI model to screen out any patients with <47 ETDRS levels via the macular field. Their AI model was trained with 572 retina images, then validated with 132 images stratified across the spectrum of nonproliferation DR (NPDR) cases. Patients who passed through the AI model were then eligible for a human grader review confirmation prior to DR clinical trial enrollment. Eligible eyes had moderately severe or severe NPDR, per levels 47-53.

Among the 132 validation images, investigators observed a model accuracy of 86.4%, with a sensitivity of 0.77 and a specificity of 0.89. The 22.6% false negative rate was explained as due to an imbalance of central and peripheral field pathology with other peripheral fields.

While the team emphasized the need for improvements to the rate of false negatives—a true risk in DR detection—they noted the AI algorithm shows capability to detect eyes with DR levels <47 ETDRS letters.

They now hope to externally validate the AI algorithm through a prospective clinical trial comparing its utility as a prescreening tool to that of a lone, traditional human grader approach.

“AI prescreening for eligible participants before grader confirmation can reduce screen failure rate, create cost efficiency and reduced burden for participants and clinical site staff,” the team concluded. “Real time automated assessment of potential eligibility could improve enrollment in DR clinical trials.”

The study, “AI-Enabled Prescreening for DR Clinical Trials,” was presented at AAO 2022.