Predicting the Risk of Cardiovascular Diseases Based on Retinal Fundus Images Using a Deep Learning Model

To develop a deep learning model to predict the risk of cardiovascular diseases (CVDs) events based on features found in fundus images.

We developed a predicting model for cardiovascular diseases based on retinal fundus images using the deep learning method. We trained our model using 2,091 retinal fundus images obtained from 211 patients. Our dataset included demographic information of each person, conventional CVD risk factors, CVD risk estimated number (calculated using the Framingham method), strokes and heart attack incidents during 5 years (patients who were referred to the ICU or CCU), and retinal fundus images for each person. We used receiver operating characteristic (ROC) analysis to assess the accuracy of our classification model.

Our proposed algorithm was able to identify high-risk individuals from no-risk individuals with 83% accuracy and a high confidence level (AUC = 0.91, P value < 0.0001). The results also showed that our model could predict cardiovascular events such as stroke with a probability of 72% (AUC = 0.83, P value< 0.0001). In comparing our model’s ability to predict CVD risk with the Framingham risk score, the Framingham model’s accuracy was 65 % in our dataset (with a best AUC of 0.78).

Our deep learning prediction model developed based on retinal fundus image findings to predict the risk of CVD, showed a relatively high accuracy. Its accuracy was higher than traditional prediction models like the Framingham model and comparable to other models based on fundus images for predicting CVD.