Abstract
Purpose :
The Veterans Healthcare Administration (VHA) has led the nation’s most successful screening program for the detection of diabetic eye disease to date and the demand for such screening continues to increase. While Artificial intelligence (AI) has strived to revolutionize the field of Ophthalmology, actual real-world data has been less than promising, particularly within the VHA setting. We hypothesize that a novel AI-based algorithm created specifically to enhance, rather than replace, the existing standard image analysis by experienced optometrists may be utilized to improve efficiency and patient safety.
Methods :
A proof-of-concept AI algorithm based on machine-learning was developed to prescreen patients within the VHA without reducing their quality of care. The algorithm analyzes images for individual image quality, image set completeness, evidence of diabetic retinopathy, and the presence of other ocular pathology. To validate this algorithm, fundus photographs of 100 consecutive Veterans within VISN 8 within a single month (2/2017) were assessed by the algorithm and also reviewed by the optometrists experienced at teleretinal screening. The conclusions of both readings were compared.
Results :
1004 images from 100 patients (~10 images/set) at 17 imaging sites within VISN 8 were included in this study. Quality metrics succeeded in being at least as stringent as the reads of the experienced reviewers. 33% of all patients were found to have image sets of adequate quality (vs. 85% by reviewers). None of the 4 patient image sets that were deemed “abnormal” for retinopathy by the experienced optometrists was passable as “normal” by the algorithm. Of the 33 adequate quality patient image sets, the algorithm detected a total of 14 image sets as “normal.” There was a zero false negative rate, supporting the safety of the algorithm. The remaining 19 image sets were subsequently reviewed by a retina specialist. One set, which was previously deemed normal by optometry reviewers, was actually “abnormal” due to the presence of mild nonproliferative diabetic retinopathy.
Conclusions :
In this pilot study, the AI algorithm was able to provide adequate screening of fundus images obtained for diabetic screening with a zero false negative rate. Our algorithm serves as an example of a means of increasing efficiency and safety while not sacrificing the quality of care to patients.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.