94 | Chapter 5 DISCUSSION In this work, we successfully applied adaptive multiple importance sampling (AMIS) to estimate posterior distributions of model parameters describing local passive and active tissue behaviour based on echocardiographic deformation measurements. Estimated deformation closely resembled the clinically measured myocardial deformation with a realistic level of uncertainty originating from both the measurement and the model. Estimated RV tissue properties reflected progression of the disease substrate over time present in the clinical case studies. Model-based inference Personalization of cardiac computational models is becoming more popular and several approaches have been proposed. Schiavazzi et al.7 used MCMC to estimate model parameters in a simplified model of the single-ventricular heart in a close-looped circulation, based on clinically measured pressures and flows. Corrado et al.23 used a Reduced Order Unscented Kalman Filter to estimate model parameters to optimize body surface potential maps and myocardial displacement. Meiburg et al.8 used the Unscented Kalman Filter to predict post-intervention hemodynamics after trans-aortic valve implantation. Zenker24 used importance sampling to estimate model parameters in a cardiovascular model. Dhamala et al.25 used high-dimensional Bayesian optimization for parameter personalization of a cardiac electrophysiological model. Coveney and Clayton26 used history matching to calibrate the maximum conductance of ion channels and exchangers in two detailed models of the human atrial action potential against measurements of action potential biomarkers. Daly et al.27 used sequential Monte Carlo Approximate Bayesian Inference to quantify the uncertainty amplification resulting in a cellular action potential model. Camps et al.28 used the same technique to estimate key ventricular activation properties based on non-invasive electrocardiography and cardiac magnetic resonance imaging. These studies used computational models with different levels of model complexity in both anatomical and physiological detail. Complex models allow personalisation with a high number of details, however, they suffer from a high-dimensional unknown space increasing the difficulty of personalization due to unidentifiability of the model parameters. This problem can be solved by reducing the complexity of the optimization problem by assuming global model parameters29 or regional model parameters.30 However, this does not reduce the computational cost and increases model discrepancy. It is suggested to use a surrogate model to approximate the exact posterior probability density function31, but this creates a new source of uncertainty. Including model discrepancy in the estimation often fails due to the non-identifiability between model parameter estimations and model discrepancy.32 The pseudo-true parameter value found by ignoring model discrepancy can still be valuable for clinical interpretation. Another approach is to reduce the complexity of the model. Various lumped parameter models of the heart and circulation have been used for fast personalization.7,8,24 The cost of low complexity may lead to an increase in model discrepancy due to model assumptions and simplifications.32 It was, however, demonstrated before that the CircAdapt model is highly efficient in simulating regional mechanics and is able to simulate realistic hemodynamics.21,33 We previously showed that the CircAdapt model can simulate segmental mechanics with a similar spatial resolution as in clinical strain imaging measurements with low discrepancy.4,21 Therefore, we assume the CircAdapt model is a suitable model for modelling regional strain in AC patients.
RkJQdWJsaXNoZXIy MTk4NDMw