Comparative Evaluation of Different Emulators for Cardiac Mechanics

Abstract

This paper outlines a comparison of different emulation based approaches to the task of parameter inference in a biomechanical model of the left ventricle of the heart, where the emulation models can account for variations in left ventricle geometry. Models considered include Gaussian processes, neural networks and random forests. We are able to achieve accurate parameter estimation for two of the model parameters, while the extension of statistical emulation to the multi geometry case allows us to observe identifiability issues in some of the model parameters. This was not observed in our previous single geometry emulation studies. Overall, this study shows the ability to generalize the single geometry emulation strategy to multiple geometries, pushing us closer towards in clinic decision support systems.

Graph Neural Network Emulation of Cardiac Mechanics

“…The issue here is that each LV geometry must be represented by a computational mesh comprised of hundreds or potentially thousands of nodes to obtain accurate numerical simulationsgenerating a training data set with a dense coverage over a space of this dimensionality is not feasible. For this reason, previous research in this area has used dimensionality reduction techniques to find a lower order representation of the LV geometry mesh, before then training an emulator in this reduced space [6,7]. The problem with this approach is that the loworder representation of the LV geometry mesh will not perfectly match the true geometry.…”