An ensemble Kalman filter approach to parameter estimation for patient-specific cardiovascular flow modeling

• Published in: Theoretical and computational fluid dynamics Vol. 34; no. 4; pp. 521 - 544
• Main Authors: Canuto, Daniel, Pantoja, Joe L., Han, Joyce, Dutson, Erik P., Eldredge, Jeff D.
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.08.2020; Springer; Springer Nature B.V
• Subjects: Boundary conditions; Classical and Continuum Physics; Computational Science and Engineering; Computer simulation; Covariance; Engineering; Engineering Fluid Dynamics; Flow simulation; Heart; Hypertension; Inflow; Kalman filters; Mathematical models; One dimensional models; Original Article; Outflow; Parameter estimation; Parameters; Pulmonary circulation; Three dimensional flow; Three dimensional models; Patient-specific; Cardiovascular; Multiscale; Data assimilation
• ISSN: 0935-4964; 1432-2250
• DOI: 10.1007/s00162-020-00530-2

Many previous studies have shown that the fidelity of three-dimensional cardiovascular flow simulations depends strongly on inflow and outflow boundary conditions that accurately describe the characteristics of the larger vascular network. These boundary conditions are generally based on lower-dimensional models that represent the upstream or downstream flow behavior in some aggregated fashion. However, the parameters of these models are patient-specific, and no clear technique exists for determining them. In this work, an ensemble Kalman filter (EnKF) is implemented for the purpose of estimating parameters in cardiovascular models through the assimilation of specific patients’ clinical measurements. Two types of models are studied: a fully zero-dimensional model of the right heart and pulmonary circulation, and a coupled 0D–1D model of the lower leg. Model parameters are estimated using measurements from both healthy and hypertensive patients, and demonstrate that the EnKF is able to generate distinct parameter sets whose model predictions produce features unique to each measurement set. Attention is also given toward the quality of model predictions made in the absence of direct clinical counterparts, as well as techniques to improve filter robustness against shrinking ensemble covariance.