Data-driven modelization for personalized medicine: an application to the thoracic aorta

Precision medicine (PM) represents a transformative approach in healthcare, promoting individualized medical interventions for specific patient subgroups. However, its application to cardiovascular disease presents intricate computational and modeling challenges attributed to cardiovascular systems’ multi-scale and multiphysics nature. This work undertakes a multidimensional exploration of data-driven methodologies, spotlighting their potential in advancing PM in cardiology. The work focuses on (but is not limited to) the thoracic aorta. First, a study of the geometrical variability of the thoracic aorta is presented by deploying a novel Statistical Shape Model (SSM) of the anatomical region. By creating a specifically designed non-rigid registration algorithm, we are able to create the first SSM of the aorta without any geometrical or topological simplification in the literature. Secondly, we create an automated machine learning workflow to (1) extract a 3d model of the aorta from a CT scan, (2) perform statistical morphological analysis, and (3) hemodynamic indices in almost real-time. This is done by training (1) a deep learning-based segmentation method on a collection of CT scans and their relative segmentations and (2) a Reduced Order Model (ROM) of hemodynamic bio-markers from a large-scale dataset of Computational Fluid Dynamic (CFD) simulations. The large-scale CFD dataset is created by performing numerical simulations of the Navier-Stokes equations on a synthetic cohort generated using the previously developed statistical shape model. Particular attention is devoted to analyzing the ROM accuracy when evaluating new aortic geometries. A method to estimate the ROM accuracy for a given geometry is proposed. In the last chapter, we present the first patient-specific FSI numerical simulations in the literature of the left part of the cardiovascular system using a data-assimilation (DA) method to integrate in-vivo measurements from a dynamic CT scan into the in-silico simulation. DA lets us reduce the uncertainties of both the FSI in-silico model and the in-vivo measurements of the soft tissue kinematics by optimally combining them in a single prediction.