Development of a Customizable Framework for Data Driven Hemodynamic Analysis with Applications in Neurovascular and Cardiovascular Scenarios
Hillner, Rachel Mae
MetadataShow full item record
Cardiovascular disease, including but not limited to heart failure and aneurysms, remain a leading cause of death worldwide. Persons over the age of 40 experience upwards of a 20% chance of developing heart failure in their lifetime. With a decrease in viable donor hearts, the reliance on mechanical heart assist devices is on the rise. However, such devices are not without their complications. Specifically, for left heart failure, a common treat, the left ventricular assist device, is characterized by clotting, stroke, and low survival rates as a result. Many studies suggest that such unfavorable results are correlated to adverse hemodynamic stimuli due to nonoptimal positioning of the device’s inflow cannula at the apex of the left ventricle. Studies have further suggested that extreme angulation of the inflow cannula in the either direction may result in nonoptimal hemodynamic stimuli, i.e. increased stasis and recirculation. Our analysis suggests that even slight variations in inflow cannula angulation (varying from -14°, -7°, 0°, 7°, and 14° from axial alignment) produces significant changes to blood flow patterns within the left ventricle. Further, the analysis of cerebral aneurysms is also heavily reliant on high level hemodynamic analysis to best understand rupturing risk and to quantify pre and post treatment hemodynamics. Even with recent advances in imaging processes and with the development of high fidelity computational fluid dynamics, there is currently no feasible method for obtaining such in depth analysis on a patient specific level in a timely manner. Imaging techniques do not provide enough detail and CFD is too time consuming, thus, there has been a push for a more optimal method to record high fidelity hemodynamic data on a patient specific level. The use of artificial intelligence, particularly neural networks, is preliminary works is proving to be promising. The use of neural networks have provided ground breaking results in across various applications, however, neural networks require a significant amount of data for adequate training. Lack of high fidelity hemodynamic training data has plagued the progression of neural networks in this application. The goal of this work is to develop an automated and customizable framework for the generation of hemodynamic data across a range of cardiovascular and neurovascular geometries and to perform preliminary testing of a neural network with the ground truth data generated. Results indicate the successful scripting of ICEM CFD and Ansys Fluent and immense time savings as a result. Additionally, preliminary testing of a neural network highlight the network’s ability to pick up on general hemodynamic trends, i.e. identifying areas of increased and decreased velocities, recirculation, and stasis. However, further testing is required to sufficiently train a neural network for assessing clinical cases.