Risk Assessment for the onset of cardiovascular disease using artificial intelligence with ocular fundus images

Abstract

Cardiovascular disease (CVD) is a leading cause of death globally. Traditional risk assessment methods rely on invasive data like cholesterol and blood pressure, which are time-consuming to collect. This study proposes a deep-learning approach to assess CVD risk using fundus images. We utilized a dataset of 7,595 fundus images from the Japan Ocular Imaging Registry and developed a multi-task learning framework based on an improved Inception-ResNet-v2 architecture. Purpose: The model integrates a feature pyramid structure and attention mechanisms to predict 8 predicting factors pointed out in the Framingham study, including age, gender, cholesterol levels, blood pressure, smoking, and diabetes status, as well as the Brinkman Index. Results: The model showed strong performance in predicting gender (AUC 0.85), diabetes status (AUC 0.80), hypertension treatment (AUC 0.82), and age estimation (R² > 0.4). The prediction of the Brinkman Index achieved an R² of 0.4, while forecasts for cholesterol levels and blood pressure demonstrated promising but moderate performance.