MEAM Ph.D. Thesis Defense: “Towards Digital Twins for Cardiovascular Flows: A Hybrid Machine Learning and Computational Fluid Dynamics Approach”

To design personalized treatment strategies, a number of measurable quantities (biomarkers) that relate a patient’s clinical representation to the existence, progress, and outcome of a disease need to be identified and measured. In cases where a biomarker is strongly correlated with the disease outcome, e.g. vascular pressure for hypertension, changes to the biomarker will perfectly describe changes of the disease outcome. However these ofter require invasive procedures to be measured. When the desired biomarkers correspond to physical properties, computational mechanics can be leveraged to obtain predictions in-silico. Unfortunately, computational models require a list of patient specific parameters, such as precise boundary conditions, which also cannot be easily measured in-vivo. Inaccurate calibration of these parameters is often the cause of poor predictions, therefore hindering the translational impact of computational methods. These challenges motivate the need flexible and computationally efficient frameworks that can operate under uncertain model assumptions and partial measurements.

The goal of this thesis is to introduce a novel approach to precision medicine by synthesizing artificial intelligence (AI) and computational modeling. We start by exploring how one can use available patient data to estimate parameters in computational fluid dynamics models of arterial blood flow, and show that this is prohibitively expensive. Then we accelerate the prediction of biomarkers by training surrogates to reconstruct available measurements by building physics-informed machine learning models to infer correlations between measurable (e.g., blood velocity) and unmeasurable quantities (e.g., vascular pressure) through underlying laws of fluid mechanics. We show that even though this is a successful approach it also faces challenges in generalizing to new clinical scenarios. Finally we propose a purely data-driven approach for making online biomarker predictions. In many biological scenarios the data acquisition process can be expensive and time consuming, limiting the amount of available training data. For this purpose, we propose creating a virtual patient database via computational fluid dynamics to train a neural operator model which we then use to make online predictions for new patients and clinical conditions. This computational efficiency that this brings has the potential to bridge the gap between modeling and clinical decision making.