Attention-based multi-fidelity machine learning model for fractional flow reserve assessment

Abstract

Coronary Artery Disease (CAD) is one of the most common forms of heart disease, caused by a buildup of atherosclerotic plaque in the coronary arteries. When this buildup is extensive, it can result in obstructions in the lumen of the blood vessels (known as stenosis) that lead to insufficient delivery of essential molecules like oxygen to the heart. Fractional Flow Reserve (FFR), defined as the ratio of pressures distal and proximal to the stenosis, is the physiologic gold standard for assessing the severity of CAD in the cardiac catheterization laboratory and relies upon the placement of an invasive coronary wire. Despite its strong diagnostic value, invasive FFR assessment is underutilized due to its cost, time-consuming nature, technique-dependent variability, and the small potential of increased risk to the patient. In this study, an attention-based multi-fidelity machine learning model (AttMulFid) is proposed for efficient and accurate virtual FFR (vFFR) assessment, including uncertainty quantification, without the use of an invasive coronary wire. Within AttMulFid, an autoencoder is used to select geometric features from the coronary arteries, with additional attention to the stenosis region. A convolutional neural network (feature fusion U-Net) combines multi-fidelity data, geometric features, and boundary conditions to produce accurate estimates of vFFR. We present results that demonstrate the good performance of AttMulFid against CFD FFR data, as well as in vivo, invasive FFR assessment from patients. Our results also show that the selected geometric features learned by the autoencoder can accurately represent the entire geometry, with greater attention on key features such as stenosis. AttMulFid thus presents itself as a feasible approach for non-invasive, rapid, and accurate vFFR assessment.