FEB 06, 2023
According to this limited population study, artificial intelligence-based image analysis appears to be as effective as manual reading when screening for diabetic retinopathy, especially in younger patients, although the logistics of implementation have yet to be determined.
This single-institution retrospective review analyzed records of 1052 patients who received diabetic retinopathy photoscreening using artificial-intelligence (AI) software during an 18-month period at their primary-care facility. Reflex dilation with 1% tropicamide was performed for ungradable nonmydriatic images, and all images received manual overread. If diabetic retinopathy was greater than mild or if other retinal pathology was detected, then this was deemed a positive result.
Forty-five percent of the 1052 patients required reflex dilation to obtain gradable images. Once dilation was performed, 91.7% of patients achieved gradable images. Of the 965 patients with gradable images, 827 had mild nonproliferative diabetic retinopathy (NPDR). Sensitivity for detection of disease was found to be 100% compared to manual grading. Gradeability of both nonmydriatic and mydriatic imaging highly correlated with age, with a higher proportion of patients in the older age groups requiring dilation and continuing to have insufficient image quality after dilation.
The fairly homogenous study population with a high proportion of Caucasians is not representative of the global population of patients with diabetes. In addition, recruitment in a primary care setting may bias the study population toward those with greater adherence to follow-up. The study also did not address whether a positive test led to adherence to the referral to an ophthalmologist. The current paper states that only 60% of patients with diabetes receive screening exams, but does the use of this AI system improve adherence? The study did not address whether patients with worse-than-mild NPDR pursued treatment.
This study demonstrated that 55% of patients can be screened with a nonmydriatic camera to detect diabetic retinopathy. Will primary care offices be comfortable dilating patients according to the protocol in this study in order to obtain the higher yield of gradable images? Furthermore, will patients who have positive results follow through with ophthalmologist appointments so that we can decrease the rate of preventable vision loss from diabetic retinopathy? Is the primary care setting the best location for this device, or would there be a higher yield in community centers or churches to better reach the population of patients who are most affected by diabetic retinopathy? This study begins to address the benefits of AI software for diabetic retinopathy screening, but how best to implement it to better serve our communities needs to be further studied.
Financial Disclosures: Dr. Lisa Schocket discloses no financial relationships.