Development and validation of a deep-learning model to predict 10-year atherosclerotic cardiovascular disease risk from retinal images using the UK Biobank and EyePACS 10K datasets

Abstract

BackgroundAtherosclerotic Cardiovascular Disease (ASCVD) is a leading cause of death globally, and early detection of high-risk individuals is essential for initiating timely interventions. The authors aimed to develop and validate a deep learning (DL) model to predict an individual's elevated 10-year ASCVD risk score based on retinal images and limited demographic data. MethodsThe study used 89,894 retinal fundus images from 44,176 UK Biobank participants (96% Non-Hispanic White, 5% diabetic) to train and test the DL model. The DL model was developed using retinal images plus age, race/ethnicity, and sex at birth to predict an individual's 10-year ASCVD risk score using the Pooled Cohort Equation (PCE) as the ground truth. This model was then tested on the US EyePACS 10K dataset (5.8% Non-Hispanic White 99.9% diabetic), composed of 18,900 images from 8,969 diabetic individuals. Elevated ASCVD risk was defined as a PCE score of ≥7.5%. ResultsIn the UK Biobank internal validation dataset, the DL model achieved area under the receiver operating characteristic curve (AUROC) of 0.89, sensitivity 84%, and specificity 90%, for detecting individuals with elevated ASCVD risk scores. In the EyePACS 10K and with the addition of a regression-derived diabetes modifier, it achieved sensitivity 94%, specificity 72%, mean error -0.2%, and mean absolute error 3.1%. ConclusionThis study demonstrates that DL models using retinal images can provide an additional approach to estimating ASCVD risk, and the value of applying DL models to different external datasets and opportunities about ASCVD risk assessment in patients living with diabetes.