Cardiovascular Modeling With Adapted Parametric Inference

Abstract

Computational modeling of the cardiovascular system, promoted by the advance of fluid-structure interaction numerical methods, has made great progress towards the development of patient-specific numerical aids to diagnosis, risk prediction, intervention and clinical treatment. Nevertheless, the reliability of these models is inevitably impacted by rough modeling assumptions. A strong in-tegration of patient-specific data into numerical modeling is therefore needed in order to improve the accuracy of the predictions through the calibration of important physiological parameters. The Bayesian statistical framework to inverse problems is a powerful approach that relies on posterior sampling techniques, such as Markov chain Monte Carlo algorithms. The generation of samples re-quires many evaluations of the cardiovascular parameter-to-observable model. In practice, the use of a full cardiovascular numerical model is prohibitively expensive and a computational strategy based on approximations of the system response, or surrogate models, is needed to perform the data as-similation. As the support of the parameters distribution typically concentrates on a small fraction of the initial prior distribution, a worthy improvement consists in gradually adapting the surrogate model to minimize the approximation error for parameter values corresponding to high posterior den-sity. We introduce a novel numerical pathway to construct a series of polynomial surrogate models, by regression, using samples drawn from a sequence of distributions likely to converge to the posterior distribution. The approach yields substantial gains in efficiency and accuracy over direct prior-based surrogate models, as demonstrated via application to pulse wave velocities identification in a human lower limb arterial network.