Clinical validation of an artificial intelligence-based diabetic retinopathy screening tool for a national health system

Abstract

Objective: To evaluate the accuracy and validity of an automated diabetic retinopathy (DR) screening tool (DART, TeleDx, Santiago, Chile) that uses artificial intelligence to analyze ocular fundus photographs for potential implementation in the national Chilean DR screening programme. Method: This was an observational study of 1123 diabetic eye exams using a validation protocol designed by the commission of the Chilean Ministry of Health personnel and retina specialists. Results: Receiver operating characteristic (ROC) analysis indicated a sensitivity of 94.6% (95% CI: 90.9–96.9%), specificity of 74.3% (95% CI: 73.3–75%), and negative predictive value of 98.1% (95% CI: 96.8–98.9%) for the automated tool at the optimal operating point for DR screening. The area under the ROC curve was 0.915. Conclusions: The results of this study suggest that DART is a valid tool that could be implemented in a heterogeneous health network such as the Chilean system.