Use of artificial intelligence on retinal images to accurately predict the risk of cardiovascular event (CVD-AI)

Purpose To create and evaluate the accuracy of an artificial intelligence platform capable of using only retinal fundus images to predict both an individual ’ s overall 10 year Cardiovascular Disease (CVD) risk and the relative contribution of the component risk factors that comprise this risk (CVD-AI). Methods The UK Biobank and the US-based AREDS 1 datasets were obtained and used for this study. The UK Biobank data was used for training, validation and testing, while the AREDS 1 dataset was used as an external testing dataset. Overall, we used 110,272 fundus images from 55,118 patient visits. A series of models were trained to predict the risk of CVD against available labels in the UK Biobank dataset. Results In both the UK Biobank testing dataset and the external validation dataset (AREDS 1), the 10-year CV risk scores generated by CVD-AI were significantly higher for patients who had suffered an actual CVD event when compared to patients who did not experience a CVD event. In the UK Biobank dataset the median 10-year CVD risk for those individuals who experienced a CVD was higher than those who did not (4.9% [ICR 2.9-8%] v 2.3% [IQR 4.3-1.3%] P<0.01.]. Similar results were observed in the AREDS 1 dataset The median 10-year CVD risk for those individuals who experienced a CVD event was higher than those who did not (6.2% [ICR 3.2%-12.9%] v 2.2% [IQR 3.9-1.3%] P<0.01 Conclusion Retinal photography is inexpensive and as fully automated, inexpensive camera systems are now widely available, minimal training is required to acquire them. As such, AI enabled retinal image-based CVD risk algorithms like CVD-AI will make cardiovascular health screening more affordable and accessible. CVD-AI ’ s unique ability to assess the relative contribution of the components that comprise an individual ’ s overall risk could inform personalized treatment decisions based on the specific needs of an individual, thereby increasing the likelihood of positive health outcomes.