Using AI to Screen for Diabetic Retinopathy in Africa
By Lynda Seminara
Selected By: Deepak P. Edward, MD
Journal Highlights
Lancet Digital Health
2019;1:e35-e44
Download PDF
Bellemo et al. tested an artificial intelligence (AI) model of deep learning as part of a screening program for diabetic retinopathy (DR) in Zambia. Their AI system was able to detect referable DR, vision-threatening DR, and diabetic macular edema (DME). Moreover, AI-generated grading was faster than human grading and was equally accurate in identifying systemic risk factors for DR.
For this clinical validation study, the authors adopted an ensemble AI model involving two convolutional neural networks. The model was first trained on 76,370 retinal fundus photographs from 13,099 patients with diabetes who participated in the Singapore Integrated Diabetic Retinopathy Program. The model was then used to evaluate patients with diabetes who attended a mobile screening event at five urban centers in Zambia’s Copperbelt Province. Referable DR was defined as nonproliferative DR of moderate or worse severity, DME, or ungradable images.
The authors calculated area under the curve (AUC), sensitivity, and specificity for referable DR, and they compared sensitivities of vision-threatening DR and DME achieved by the AI system and by the grading performed by retina specialists. They also conducted multivariate analysis to compare the two methods of grading.
Among the 4,504 study images from 3,093 eyes (1,574 people with diabetes), the AI system identified referable DR in 697 eyes (22.5%). It detected vision-threatening DR in 171 eyes (5.5%) and DME in 249 eyes (8.1%). The AUC of the AI system for referable DR was 0.973. The corresponding sensitivity and specificity were 92.25% and 89.04%, respectively. The system’s sensitivity for detecting vision-threatening DR and DME was 99.42% and 97.19%, respectively. Outcomes for the AI model resembled those of human graders; both models identified major risk factors for referable DR, including longer duration of diabetes, higher systolic blood pressure, and higher levels of glycated hemoglobin.
Additional study is needed to assess the cost-effectiveness of AI models, said the authors. Moreover, it will be crucial to ensure that under-resourced areas have enough specialists to treat referable DR.
The original article can be found here.