Machine Learning-Integrated PDE Model for Blood Flow Simulation and Arterial Plaque Progression Detection in Cardiovascular Diagnosis

Abstract

The main issues in the modern cardiovascular diagnosis are the accurate prediction of the blood flow dynamics and the possibility to identify the arterial plaque progression at the initial stage. Partial differential equations (PDE)-based traditional computational fluid dynamics (CFD) models, especially the Navier-Stokes equations, provide high-fidelity hemodynamic models, but are expensive and demand considerable computational power and cannot be readily adapted to patient pathophysiological variations. Machine learning (ML) on the other hand, is good at real-time inference, but is not always interpretable and physically consistent. The paper provides a hybrid ML-based PDE model, which integrates physics-based modelling and learned surrogate modules to speed up the process of simulation, improve prediction of plaque-progression, and be physiological-valid. The ML model was trained on a dataset of coronary artery CT scans and Doppler ultrasound measurements and the PDE-based solver was validated. The hybrid method was better at prediction and 20 times less costly in computational cost compared to classical solvers. All the statistics and tables that are mentioned in the text are presented in the article.