Purchase this article with an account.
E Simon Barriga, Eva Rosita Dewi, Olivia Baldivieso, Jimmy Borda, Christian Diaz, Ehsan Rahimy, Jeremy Benson, Jeff Wigdahl, Gilberto Zamora, Rajat N Agrawal, Peter Soliz; Using a Handheld Retinal Camera and Artificial Intelligence for Diabetic Retinopathy Screening in Bolivia. Invest. Ophthalmol. Vis. Sci. 2020;61(7):1645.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To assess the clinical feasibility of diabetic retinopathy (DR) screening using artificial intelligence (AI) paired with a handheld retinal camera as compared to multiple trained readers in Bolivian patients.
Adult diabetic patients admitted to four clinics in the City of Santa Cruz, Bolivia were enrolled in a non-comparative, non-randomized, one-time observation-only study. Subjects were dilated with Tropicamide and images were collected using Volk Pictor Plus, a handheld non-mydriatic fundus camera (Volk Optical, Mentor, OH). The images were graded by two retina specialists according to the International DR grading scale. Human grading results were mapped into “refer” (severe NPDR, PDR, and/or CSME) and “non-refer” (no DR, mild NPDR, or moderate NPDR) categories. Disagreements in referral level between the two readers were sent to a third retina specialist for adjudication. The images were then processed using the EyeStar DR screening software (VisionQuest Biomedical, Albuquerque, NM) which provided a “refer” or “non-refer” recommendation. The EyeStar recommendations were compared to the adjudicated DR grades to assess sensitivity and specificity of the AI software.
Dilated fundus images were collected for 1,005 patients of 1,035 patients were enrolled in the study. The two readers had referral level disagreements in 79 (7.9%) of the cases. Cohen’s kappa agreement for the two readers was 0.58 (95% CI = [0.49,0.67]). After adjudication by the third retina specialist, 911 cases were classified as “non-refer” (90.8%) and 92 cases were classified as “refer” (9.2%).Due to bad image quality, the software was not able to process 78 cases (7.76%). Compared to the adjudicated grades, the EyeStar AI software’s sensitivity and specificity for referable DR and CSME were 91% and 81%, respectively, with an area under the ROC curve (AUC) of 91.4%. When removing 120 cases with worse than low average image quality as deemed by the readers, EyeStar achieved 93% sensitivity and 85% specificity with an AUC of 94%.
This study is the first to evaluate the use of AI for DR screening utilizing a portable retinal camera in urban and rural populations in Bolivia. The results demonstrate the feasibility of using the combination of portable camera and AI for DR screening, which increases the ability to screen diabetic patients for retinopathy in low-resource and rural settings.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only