A machine learning-based reduced order model for the investigation of the blood flow patterns in presence of a stenosis of the left main coronary artery

Coronary artery diseases represent a significant cause of death worldwide. In fact, their occlusion can cause lack of oxygen to the heart tissue, leading to heart failure or heart attack. Coronary artery bypass grafting is a surgical procedure to restore a proper blood supply. In this context, Computational Fluid Dynamics (CFD) can be a significant tool to improve coronary artery bypass grafts and to avoid unfavorable flow conditions in the region of the anastomosis, which can be associated with the failure of the surgical procedure. In this work the development of a Reduced Order Model (ROM) for the investigation of hemodynamics in a patient-specific configuration of coronary artery bypass graft is proposed. The method deployed extracts a reduced basis space from a collection of highfidelity solutions via a Proper Orthogonal Decomposition (POD) algorithm and employs Artificial Neural Networks (ANN) for the computation of the modal coefficients. The main goal is to characterize blood flow for different settings that are relevant in the clinical practice, such as several stenosis factors, in a rapid and reliable way. The Full Order Model (FOM) is represented by the incompressible Navier-Stokes equations discretized by using a Finite Volume (FV) technique. The computational domain is referred to coronary arteries, in particular left branches when a stenosis of the Left Main Coronary Artery (LMCA) occurs. In the first stage, only a stenosis that reduces the width of the vessel by 50% is taken into account, so the time is the only parameter considered in the model. Then, stenosis from mild to severe is added too. Several numerical results are analysed, underlining the computational performance of the proposed approach, such as the error between computed FOM and ROM solutions as well as the substantial speed-up achieved at the online stage.