Clinical Implementation of Autonomous Artificial Intelligence Systems for Diabetic Eye Exams: Considerations for Success
Risa M. Wolf, Roomasa Channa, Harold P. Lehmann, Michael D. Abràmoff, T. Y. Alvin Liu
Abstract
Diabetic eye disease (DED), including diabetic retinopathy (DR) and diabetic macular edema (DME), is a complication of diabetes and the leading cause of vision loss among working-age adults in the United States (1-3).Screening for DED can lead to its early identification and treatment, thereby preventing irreversible vision loss (4-7).However, rates of diabetic eye exams for DED screening remain suboptimal, with reported rates in the United States ranging from 11 to 70% (8,9).Autonomous artificial intelligence (AI)-based diabetic eye exams have the potential to increase access to these exams and facilitate the early identification of DED so that timely treatment can be administered to prevent blindness.Autonomous AI systems use a robotic nonmydriatic fundus camera with a built-in AI algorithm to provide feedback to the operator to acquire high-quality fundus photographs for determining the presence or absence of referable DED, with immediate results after image acquisition at the point of care (10-12).The autonomous AI system guides the operator to acquire two color fundus images determined to be of adequate quality using an image quality algorithm, with one each centered on the fovea and the optic nerve, and guides the operator to retake any images of insufficient quality.It is important to note that these systems have been rigorously validated against a prognostic standard to identify DR and DME and do not diagnose other eye conditions.The first autonomous AI diagnostic system (Luminetics-Core, formerly IDx-DR, Digital Diagnostics, Coralville, IA) was de novo-authorized by the U.S. Food and Drug Administration (FDA) in 2018 after completion of a pivotal trial demonstrating its safety, efficacy, and equity in diagnosing referable DED compared with a prognostic standard outcome (10).In the pivotal trial, the autonomous AI system demonstrated 87% sensitivity and 90% specificity in detecting referable DR and/or DME in adults (10,13).Since this first autonomous AI system was approved, two additional autonomous AI systems have been authorized under the FDA's 510(k) authorization process using it as a predicate, and there are several other autonomous AI systems in different stages of development (11,12).