Pivotal trial of an autonomous AI-based diagnostic system for detection of diabetic retinopathy in primary care offices

Michael D. Abramoff, Philip T. Lavin, Michele Birch, Nilay Shah & James C. Folk

People with diabetes fear visual loss and blindness more than any other complication.1 Diabetic retinopathy (DR) is the primary cause of blindness and visual loss among working age men and women in the United States and causes more than 24,000 people to lose vision each year.2,3 Adherence to regular eye examinations is necessary to diagnose DR at an early stage, when it can be treated with the best prognosis,4,5 and have resulted in substantial reductions in visual loss and blindness.6 Despite this, less than 50% of patients with diabetes adhere to the recommended schedule of eye exams,7 and adherence has not increased over the last 15 years despite large-scale efforts to increase it.8 To increase adherence, retinal imaging in or close to primary care offices followed by remote evaluation using telemedicine has also been widely studied.911

Artificial intelligence (AI)-based algorithms to detect DR from retinal images have been examined in laboratory settings.1215 Recent advances incorporate improved machine learning into these algorithms have led to higher diagnostic accuracy.16,17 However, in addition to high diagnostic accuracy, responsible and safe implementation in primary care requires autonomy (i.e., a use case that removes the requirement for review by human experts), instantaneous image quality feedback to the primary care based operator in order to reach a reliable disease level output in the vast majority of patients, a realtime clinical decision at the point of care, and consistent diagnostic accuracy across age, race and ethnicity.12,13,18,19 Studies comparing an AI system against an independent, high-quality gold standard that includes fundus imaging and Optical Coherence Tomography (OCT) imaging protocols have not previously been conducted; FDA has not previously authorized any such system.