Abstract
Background: Cardiovascular diseases (CVDs) are responsible for 32.4% of all deaths across the European Union (EU), and several CVD risk scores have been developed, with variable results. Retinal microvascular changes have been proposed as potential biomarkers for cardiovascular risk, especially in coronary heart diseases (CHDs). This study aims to identify the retinal microvascular features associated with CHDs and evaluate their potential use in a CHD screening algorithm in conjunction with traditional risk factors. Methods: We performed a two-center cross-sectional study on 120 adult participants-36 patients previously diagnosed with severe CHDs and scheduled for coronary artery bypass graft surgery (CHD group) and 84 healthy controls. A brief medical history and a clinical profile were available for all cases. All patients benefited from optical coherence tomography angiography (OCTA), the use of which allowed several parameters to be quantified for the foveal avascular zone and superficial and deep capillary plexuses. We evaluated the precision of several classification models in identifying patients with CHDs based on traditional risk factors and OCTA characteristics: a conventional logistic regression model and four machine learning algorithms: k-Nearest Neighbors (k-NN), Naive Bayes, Support Vector Machine (SVM) and supervised logistic regression. Results: Conventional multiple logistic regression had a classification accuracy of 78.7% based on traditional risk factors and retinal microvascular features, while machine learning algorithms had higher accuracies: 81% for K-NN and supervised logistic regression, 85.71% for Naive Bayes and 86% for SVM. Conclusions: Novel risk scores developed using machine learning algorithms and based on traditional risk factors and retinal microvascular characteristics could improve the identification of patients with CHDs.