Deep learning algorithm for detecting obstructive coronary artery disease using fundus photographs

Abstract

Abstract Previous studies validating fundus photographs to provide information about coronary artery disease (CAD) risk are limited. Deep learning further facilitates and enhances the use of fundus photography. Therefore, we aimed to design and prospectively validate a deep learning model for detecting obstructive CADin patients with suspected coronary artery disease.The algorithm was trained to predict obstructive CAD using fundus photographs of 4808 participants in validation group and 1385 patientsin external test group. The performance of the model was evaluated using area under the receiver operating characteristic curve (AUC) with the cardiologist's diagnosis as the reference standard and compared to pre-test probability models. The algorithm had an AUC of 0.833 and 0.751 for detecting obstructive CAD in the validation and external test groups, respectively, which was higher than the Updated Diamond Forrester Method and the Duke Clinical Score. The proposed deep learning model has a moderate performance in diagnosing obstructive CAD. The results from this multicenter study advance the development of clinically applicable and interpretable deep learning systems for detecting obstructive CAD from fundus photographs.