A prospective study (NCT04302012) evaluated the reliability of the EyeArt AI DR detection system (Eyenuk, Inc.) across repeated examinations, operators, and cameras. Using retinal images from off-the-shelf color fundus retinal cameras, this AI system is intended to detect patient eyes with referable DR (rDR) [moderate non-proliferative DR (NPDR) or higher on the International Clinical DR (ICDR) severity scale or clinically significant diabetic macular edema (CSDME)] and vision-threatening DR (vtDR) [severe NPDR or higher on the ICDR scale or CSDME].A total of 62 subjects with diabetes were enrolled at 2 U.S. primary diabetes care centers and completed digital DR examination using the AI system (including retinal imaging) multiple times. Of these, 31 subjects underwent 12 AI exams for each eye performed by 3 operators on 4 cameras, and another 31 subjects underwent 6 AI exams for each eye performed by 1 operator on 2 cameras. Subjects had a median age of 50 (22-70) years, 42% were female, 79% were White, 6.5% were Black or African-American, and 6.5% were Asian. Mean (± std. dev.) diabetes duration was 12.5 ± 9 years and 66% had type 2 diabetes. Of the 1116 repeated exams (= 31 subjects x 12 exams x 2 eyes/subject + 31 subjects x 6 exams x 2 eyes/subject), 16 eye exams had missing data (due to minor protocol deviations) and 22 eye exams (all from the same subject) did not have sufficient exam quality to rule out presence of disease. Of the remaining 1078 eye exams, 97.5% were in agreement for the referable DR detection results with a majority of the disagreements limited to only 5 eyes. For the vision-threatening DR detection results, 99.5% were in agreement. This study demonstrates that the EyeArt AI system is robust to variations and can be used to reliably detect diabetic retinopathy in non-ophthalmology settings across different operators and cameras.
E. Ipp: Research Support; Self; Eyenuk, Inc. D.R. Liljenquist: None.
