Are Cardiovascular Risk Scores from Genome and Retinal Image Complementary? A Deep Learning Investigation in a Diabetic Cohort

Mohammad Ghouse Syed, Alexander S. Doney, Gittu George, Ify Mordi, Emanuele Trucco

Abstract: Risk of cardiovascular diseases (CVD) is driven by both genetic and environmental factors. Deep learning (DL) has shown that retinal images contain latent information indicating CVD risk. At the same time, genome-wide polygenic risk scores have demonstrated CVD risk prediction accuracy similar to conventional clinical factor-based risk scores. We speculated that information conveying CVD risk in retinal images may predominantly indicate environment factors rather than genetic factors, i.e., provide complementary information. Hence, we developed a DL model applied to diabetes retinal screening photographs from patients with type 2 diabetes based on EfficientNetB2 for predicting clinical atherosclerotic cardiovascular disease (ASCVD) risk score and a genome-wide polygenic risk score (PRS) for CVD. Results from 6656 photographs suggest a correlation between the actual and predicted ASCVD risk score (R2 = 0.534, 95% CI [0.504, 0.563]; MAE = 0.109 [0.105, 0.112]), but not so for actual and predicted PRS (R2 = −0.005 [−0.02, 0.01]; MAE = 0.484 [0.467, 0.5]. This suggests that retinal and genetic information are potentially complementary within an individual’s cardiovascular risk, hence their combination may provide an efficient and powerful approach to screening for CVD risk. To our best knowledge, this is the first time that DL is used to investigate the complementarity of retinal and genetic information for CVD risk.