Deep Learning Diagnostic Models for Coronary Heart Disease Based on Retinal Fundus Photographs and Optical Coherence Tomography

Background: Coronary heart disease (CHD) is one of the leading causes of death worldwide. This study aimed to investigate the ability of a multimodal deep learning model for predicting of CHD from ocular information. Methods: In this retrospective study, we developed a multimodal deep-learning model based on retinal fundus photographs and optical coherence tomography (OCT) images via a convolutional neural network and a convolution block attention module to predict the risk of CHD. In addition, a nomogram model combine the AI prediction results and clinical features were constructed. The predictive values of these models were evaluated by receiver operating characteristic curves and calibration curves. Results: Among a cohort of 507 patients, 284 of them were in CHD group. The AUC reached 0.992 by AI diagnostic model for predicting CHD in training cohort, while the predictive value was also validated in test and validation cohort with fairly high AUC and good consistency determined by the calibration curves. Moreover, nomogram based on risk factors of CHD and AI prediction scores also illustrated a good predictive value and consistency. Conclusions: We developed a deep learning artificial intelligence diagnostic models based on retinal fundus photographs and OCT images can help diagnosing CHD.