Investigating Vascular Disease Treatment and Progression using Multiscale Hemodynamic Modeling and Machine Learning Algorithms

Abstract

Cardiovascular diseases (CVDs) are the leading cause of death worldwide and are continuing to increase in prevalence throughout the world. While significant strides have been made towards reducing the morbidity and mortality of CVD, many current interventional and treatment strategies are associated with high rates of patient complications. A critical component behind the high complication rate is the complexity of the underlying patient hemodynamics and its relationships with vascular disease, but understanding this interplay is essential in order to better predict the progression of vascular disease and improve interventional planning. To investigate disease progression, we studied hemodynamics in two aortic diseases: Stanford type B aortic dissection (TBAD) and coarctation of the aorta (CoA). To examine and identify the role of personalized simulation in interventional planning, we studied venoarterial extracorporeal membrane oxygenation (VA-ECMO), a treatment system for patients with cardiopulmonary failure. These hemodynamic relationships are difficult to assess in vivo, but, in this work, we developed computational models to simulate blood flow, offering faster, cheaper, and more adaptable alternatives. While simulations provide a highly practical method of exploring these disease applications, the immense computational demand of simulations in large arteries limits their applicability in clinical settings and their tractability for widespread research purposes. Thus, this work also focused on the development and application of hemodynamic modeling algorithms with reduced computational cost, leading to more accessible frameworks. For patients supported by VA-ECMO, we developed multiscale modeling frameworks to probe the influence of VA-ECMO on pathological cerebral hemodynamics and its relationship with neurological complications. Our multiscale modeling framework implemented high-fidelity massively parallel three-dimensional (3D) models in the cerebral vasculature with low-fidelity one-dimensional models in the lower body to reduce the computational demand. We found that increasing the VA-ECMO flow rate compared with cardiac output reduced the total cerebral blood flow to the brain. Additionally, a VA-ECMO flow fraction of over 75% was needed to fully perfuse the cerebral arteries with oxygenated blood. For the aortic disease models, we combined simulation with machine learning to improve the modeling pipeline and reduce the computational demand. In TBAD patients, we used 3D models to discover relationships between local hemodynamics and aneurysmal degeneration and found that flow rate to the false lumen could supplement current indicators for patient intervention. In patients with CoA, we developed a design of experiments framework with machine learning to predict the pressure drop across and local flow downstream of a stenosis subject to a variety of co-factors, reducing the computational demand of simulations by ~5.6x. This framework can be used to probe the effects of co-factors on clinical hemodynamic metrics in a variety of geometries and vascular diseases. To further reduce the computational cost of aortic diseases models by nearly 9.8x, we combined deep learning with simulation to develop a framework capable of predicting pulsatile hemodynamics from a steady state solution. Altogether, this work can be broadly grouped into two categories: 1) investigating treatment options with multiscale models that combine low and high-fidelity simulations to target specific regions of interest and 2) predicting CVD progression and patient risk by combining simulation with machine learning to reduce the number of simulations needed and the computational cost of each simulation. The disease applications that we studied and the frameworks that we developed can be used by researchers and clinicians to model hemodynamics and improve treatment strategies and clinical decisions for patients with CVD.