Real world outcomes from artificial intelligence to detect diabetic retinopathy in the primary care setting: 12 month experience

Abstract

Purpose : Recently, artificial intelligence (AI)- based testing platforms for diabetic retinopathy (DR) have been put into clinical use but limited outcomes data exists. We performed a retrospective, observational study to report real world outcomes of follow up and accessibility to ophthalmologic care in patients with diabetic retinopathy utilizing a community, AI-based testing program.

Methods : The STATUS (Stanford Teleophthalmology Autonomous Testing and Universal Screening) program utilizes the FDA cleared autonomous AI technology, IDx-DR, to detect diabetic retinopathy in seven primary-care and endocrinology clinics in the San Francisco Bay area. Adult patients with type one or type two diabetes mellitus without a prior diagnosis of DR and dilated fundus exam in the prior 12 months were offered inclusion in the program. Two nonmydriatic fundus images of each eye – one 45-degree image centered on the macula and one centered on the optic disc – were obtained by a medical assistant in the clinics. Sufficient images were graded on the presence or absence of referral-warranted more-than-mild DR (MTMDR), defined as ETDRS level 35 or higher, and/or diabetic macular edema, in at least one eye. Patients with MTMDR or an ungradable image were referred for an ophthalmologic exam. Patient demographics and characteristics were collected from those with referrable disease. A random sample of patients were given a survey about their experience and follow up.

Results : 1222 patients were screened with IDx-DR over a 12-month period. 145 patients (12%) were given a diagnosis of MTMDR and referred for in-person ophthalmologic evaluation. 93 patients (64.1%) saw an ophthalmologist at a median 60 days after the screening visit. 30.3% of those referred after an AI-based exam followed up at the university eye institute compared with 11.5% of people referred for follow up after teleophthalmology screening. The percentage of patients that received an annual eye exam increased from 65.2% to 72.8% in the clinics utilizing IDx-DR, exceeding the 90^th percentile national benchmark (67.89%) for the Healthcare Effectiveness Data and Information Set (HEDIS) quality measure.

Conclusions : Diabetic eye exams increased in community clinics utilizing IDx-DR. Incorporating AI into a remote diabetic retinopathy detection workflow may increase in person follow up compared to teleophthalmology alone.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.