Purpose
Evaluate an artificial intelligence (AI) system to screen people with diabetes at point-of-care for diabetic retinopathy (DR) including diabetic macular edema (DME).
Methods
We conducted a prospective multi-center study (NCT03112005) in which patients with diabetes were enrolled consecutively initially and later preferentially based on enrichment criteria. The study subjects underwent undilated 2-field fundus photography (macula centered and disk centered images) for the EyeArt AI eye screening system and dilated 4-wide field stereoscopic fundus photography. The EyeArt system provided eye-level results for referable DR (rDR), which is defined as moderate non-proliferative DR (NPDR) or higher (International Clinical DR (ICDR) severity scale) or clinically significant DME. The EyeArt system was evaluated against the clinical reference standard based on adjudicated grading of the 4-wide field photographs by expert graders at the Wisconsin Fundus Photograph Reading Center using the Early Treatment Diabetic Retinopathy Study (ETDRS) Severity Scale. Statistical analyses were performed with variance adjustment to account for the correlation between eyes of the same patient.
Results
1786 eyes from 893 subjects were included. Of these, 1718 eyes were gradable using the clinical reference standard and 326 of these were positive for rDR (290 moderate NPDR, 4 severe NPDR, 31 proliferative DR, and 85 clinically significant DME) and 1392 eyes were negative for rDR (1134 no apparent DR and 258 mild NPDR). Sensitivity of the EyeArt system using undilated images was 95.5% [95% CI: 92.4% – 98.5%], specificity was 86.0% [95% CI: 83.7% – 88.4%, and gradability rate was 87.5% [95% CI: 85.4% – 89.7%]. With a dilate-if-needed photography protocol (where dilated images were used for subjects with ungradable EyeArt results on undilated images), the gradability rate of the EyeArt system improved to 97.4% [95% CI: 96.4% – 98.5%], sensitivity was 95.5% [95% CI: 92.6% – 98.4%], and specificity was 86.5% [95% CI: 84.3% – 88.7%].
Conclusions
The EyeArt AI system compared favorably with the clinical reference standard from 4-wide field stereoscopic images and met the predetermined sensitivity and specificity endpoints for the detection of referable DR in people with diabetes.
Study Reference
Jennifer Lim, Malavika Bhaskaranand, Chaithanya Ramachandra, Sandeep Bhat, Kaushal Solanki, and Srinivas Sadda. “Artificial Intelligence Screening for Diabetic Retinopathy: Analysis from a Pivotal Multi-Center Prospective Clinical Trial.” In ARVO Imaging in the Eye Conference 2019. Vancouver, BC, Canada, 2019.