Predictive modeling of cardiovascular risk factors using OCT and Fundus Images with deep learning techniques

Abstract

PURPOSE: The study aims to develop and test predictive models using fundus and Optical Coherence Tomography (OCT) images to detect retinal structural patterns linked to cardiovascular risk factor and events, specifically arterial hypertension (AHT), type II diabetes mellitus (T2D) and dyslipidemia. METHODS: The study included patients over 18 years old registered in the hospital information system, regardless of cardiovascular disease history. Imaging data comprised macula-centered and optic nerve-centered OCT images, as well as 45º or greater fundus images, collected between January 2016 and May 2024. A total of 30,773 OCT images were extracted, including 3,837 OCTs from health subjects, which were used as control group across three predictive models. Cohorts included 6,321 OCTs from patients with AHT, 3,479 from those with T2D and 6,824 from patients with dyslipidemia, with some images overlapping across cohorts due to comorbidities. Three predictive models were developed, each targeting one cardiovascular risk factor. For each cohort, two reference architectures, SwinTransformerV2 and RETFound, were trained and tested to compare their ability to capture structural retinal patterns associated with the studied risk factors. Predictive performance of the models was assessed using the area under the receiver operating characteristic curve (AUC).