NOV 25, 2019
This cross-sectional study compared the diagnostic accuracy of a smartphone-based artificial intelligence system versus ophthalmologist judgment for retinopathy screening in patients with diabetes.
This 2-month study was conducted among diabetic patients in India. A smartphone-based artificial intelligence system was used to grade images that were taken on a smartphone-based, nonmydriatic retinal camera by minimally trained health care workers. No internet connectivity was required as the AI system was housed within the smartphone itself.
After dilation with tropicamide 1%, an anterior segment photograph and 3 fundus photographs were obtained from 213 patients. The AI system prompted the photographer to retake the images if they were not of sufficient quality. A vitreoretinal surgeon and resident graded the images and the results were compared with the AI system.
The sensitivity of the AI system to diagnose referable diabetic retinopathy [moderate nonproliferative diabetic retinopathy (NPDR) or worse] was 100% and the specificity was 88.4%. The sensitivity to diagnose any diabetic retinopathy was approximately 85% with a specificity of 92%.
This was a comparatively small study that needs to be validated in a larger cohort. A large number of patients had no retinopathy, and from the manuscript, it is unclear how the system performed on eyes with more significant pathology and whether the system can detect other concomitant pathology.
This study has several important findings. Firstly, the system was both sensitive and specific in detecting referable diabetic retinopathy. This platform was portable and could be taken to patients are rather than having the patients come to a center for imaging. This may dramatically increase the ability to screen patients both in developing as well as in developed countries.
The photographers had minimal photographic training, which will hopefully reduce the cost of screening and also increase the availability of this technology to a larger group of patients. The imaging was done with an inexpensive smartphone camera that did not require internet connectivity. This is critical in areas without significant internet and wireless infrastructure and will further reduce screening burden and expense.