Abstract

Introduction

Diabetic eye disease (DED) is a complication of diabetes that is the primary cause of blindness in working-age adults in the U.S.^1,2. Early detection (‘screening’) and treatment can frequently prevent progression, but the majority of the 34 million people with diabetes in the US have a DED screening care gap, due to a lack of access, and education around the need for a diabetic eye exam^3,4. This care gap is a major source of health disparity, with racial and ethnic minorities, and under-resourced communities, having worse outcomes, and a disproportionately higher prevalence of DED^2,5–9.

While DED prevalence is lower in youth (defined as those aged <21 years) with diabetes, it affects approximately 4–9% of youth with type 1 diabetes (T1D) and 4-15% of youth with type 2 diabetes (T2D)^10–14. The risk for DED increases with the duration of diabetes in T1D, and recent data from the TODAY2 follow-up study demonstrated a diabetic retinopathy prevalence rate of 49% at a mean diabetes duration of 12 years in youth onset T2D¹⁵. The American Diabetes Association (ADA) and American Academy of Ophthalmology (AAO) recommend DED screening in T1D within 3–5 years of diagnosis and age greater than 11 years, and in T2D at the time of diagnosis in youth¹⁶. However, only 35-72% of diabetic youth undergo recommended screening exams, with even higher care gap rates in minority and lower socioeconomic background youth^10,17. Commonly reported barriers to screening include miscommunication regarding the need for a diabetic eye exam, time for an additional doctor’s visit, and transportation barriers^17,18.

While the introduction of telemedicine over the last two decades has improved screening and facilitated early detection of diabetic eye disease^19–23, the development of diagnostic autonomous artificial intelligence (AI) systems for diagnosing DED has ushered in the next chapter of DED screening^24–29. Autonomous AI systems require a camera operator to obtain point-of-care fundus images, which are then interpreted by an AI algorithm to provide a diagnosis without human oversight. The regulatory approval of the first diagnostic autonomous AI system for diabetic eye exams was based on a pivotal trial against a prognostic standard, i.e., patient outcome²⁴, showing its safety, efficacy, and lack of racial and ethnic bias for diagnosing DED in adults with diabetes with 87% sensitivity and 91% specificity²⁴. In prior studies of youth with diabetes, we demonstrated that the diagnosability of autonomous AI in youth was 97.5%, with 85.7% sensitivity and 79.3% specificity in detecting DED, with no difference in diagnosability across demographic groups³⁰. We have also shown that this system can be implemented in a multidisciplinary diabetes clinic and has the potential to increase DED screening rates in underserved youth, while also being cost-savings to patients and caregivers^30,31.

While diagnostic accuracy has been a focus of study of diagnostic AI systems^25,30,32, the effectiveness of autonomous AI to increase adherence and follow-up compared to traditional referral has not been evaluated in a rigorously designed randomized trial. We hypothesized that autonomous AI closes the diabetic eye exam care gap, and increases follow-up, compared to traditional eye care provider(ECP) referral in youth. To test this hypothesis, we designed ACCESS (AI for Childrens’ diabetiC Eye examS Study), a pre-registered, rigorously designed randomized control trial (RCT), to measure diabetic eye exam completion rates in a racially and ethnically diverse cohort of youth with T1D and T2D.