Deep-stratification of the cardiovascular risk by ultrasound carotid artery images

Cardiovascular risk estimation functions predict the risk of cardiovascular events with clinical data and survival models. These functions accurately stratify individuals into low, moderate, and high -risk categories. However, they tend to classify a considerable number of individuals into the middle -risk category, and often, a subsequent reclassification into high -risk groups is required. Atherosclerosis is the leading cause of cardiovascular events, and ultrasound images of the Carotid Artery (CA) can detect its burden by measuring the carotid intimamedia thickness and identifying atherosclerotic plaques. Current risk estimation functions do not consider ultrasound imaging. This paper proposes the use of deep ultrasound CA image features in survival models to improve risk stratification. In particular, we define new deep CA image features, extracting information from a convolutional neural network, and add them to an existing risk function. The experiments carried out show that using deep image features improves the AUC of the risk function to 0.842, and these features are enough to replace the information provided by blood biomarkers. Furthermore, the use of these features resulted in a 20% improvement in the reclassification of risk categories, specifically for individuals who suffered an event, as shown by the net reclassification improvement metric.