Artificial intelligence applied to non-contrast-enhanced cardiac computed tomography for the prediction of cardiovascular events

Abstract

Background Non-contrast-enhanced cardiac computed tomography (CT) may provide two measures that are emerging as independent predictors of cardiovascular events: coronary calcium score (CCS) and the volume of epicardial fat, a metabolically and immunologically active tissue surrounding the coronary arteries. The quantification of epicardial fat volume (EFV) is not routinely performed in clinical practice for the long time required for image reconstruction and the intra- and inter-observer variability. Purpose We evaluated if artificial intelligence (AI) might prove a valuable tool to interpret the CT data-set, and to better understand the relative prognostic value of CCS and EFV compared to “traditional” cardiovascular risk factors. Methods The Montignoso HEart and Lung Project is a community-based study carried out in a small town of Northern Tuscany (Italy). Starting from 2009, asymptomatic individuals from the general population underwent a baseline screening including a non-contrast cardiac CT, and were followed-up. For the present study, CCS and EFV were automatically measured from CT scans through a deep learning (DL) strategy based on convolutional neural networks. Because of the low incidence of the primary endpoint (myocardial infarction [MI]), the observed cardiac events were predicted with a random forest model built using a subsampling approach. Results Study participants (n=1528; 48% males, age 40 to 77 years) experienced 47 MI events (3%) over 5.5±1.5 years. CCS and EFV independently predicted this endpoint (p values <0.001 and 0.005, respectively) in a model including other predictors, namely weight, age, male gender, and hypertension. The model displayed a good prognostic performance, with an out-of-bag accuracy of 80.43% (accuracy on non-event prediction: 81.17%; performance on event prediction: 57,45%). The CCS emerged as the most important predictor, followed by EFV, weight and age. Interestingly, the incidence of cardiovascular events linked with CCS levels was associated with elevated EFV and the subjects with elevated CCS values but low EFV had no events (figure 1). Conclusions The tools of AI allow to perform an automated analysis of non-contrast-enhanced CT scans, with rapid and accurate measurement of CCS and EFV through a DL approach. In asymptomatic individuals from the general population, these features are more predictive of non-fatal MI than other variables related to the cardiovascular risk, as we can be demonstrated through an application of AI. Figure 1 Funding Acknowledgement Type of funding source: None