Predicting left ventricular contractile function via Gaussian process emulation in aortic-banded rats

Abstract

Cardiac contraction is the result of integrated cellular, tissue and organ function. Biophysical in silico cardiac models offer a systematic approach for studying these multi-scale interactions. The computational cost of such models is high, due to their multi-parametric and nonlinear nature. This has so far made it difficult to perform model fitting and prevented global sensitivity analysis (GSA) studies. We propose a machine learning approach based on Gaussian process emulation of model simulations using probabilistic surrogate models, which enables model parameter inference via a Bayesian history matching (HM) technique and GSA on whole-organ mechanics. This framework is applied to model healthy and aortic-banded hypertensive rats, a commonly used animal model of heart failure disease. The obtained probabilistic surrogate models accurately predicted the left ventricular pump function ( R 2 = 0.92 for ejection fraction). The HM technique allowed us to fit both the control and diseased virtual bi-ventricular rat heart models to magnetic resonance imaging and literature data, with model outputs from the constrained parameter space falling within 2 SD of the respective experimental values. The GSA identified Troponin C and cross-bridge kinetics as key parameters in determining both systolic and diastolic ventricular function. This article is part of the theme issue ‘Uncertainty quantification in cardiac and cardiovascular modelling and simulation’.

Linking statistical shape models and simulated function in the healthy adult human heart

“…And thirdly, our models are coupled to fixed Windkessel models as boundary conditions, as opposed to a closed loop cardiovascular system, where changes in cardiac stroke volume, in either ventricle, can feedback and result in alterations in the preload and afterload that can, in turn, change the cardiac cycle timings. However, previous studies [ 44 ] found that boundary conditions did not have a significant influence on cardiac output, albeit in a rat model. Our results from the local sensitivity analysis agree with [ 44 ], showing that the results are not overly sensitive to the stiffness of the boundary conditions.…”

“…We performed a GPE-based GSA as in [ 44 ]. Briefly, we trained one GPE per phenotype defined in Table 1 , with the exception of phenotypes whose range of variability was below a threshold, determined below, across the cohort.…”

Modelling the Effects of Conductive Polymers and Stem Cells Derived Myocytes on Scarred Heart Tissue

“…In order to perform a systematic investigation on the importance of each model parameter, we fitted a Gaussian process emulator [12,13] to our model, enabling us to perform a GSA. GSA's aims to quantify the relative importance of input variables in determining the value of an assigned output variable [14], which in our model is represented as activation times, collected on the top right corner of the tissue slab.…”

Surrogate models based on machine learning methods for parameter estimation of left ventricular myocardium

“…Di Achille et al [ 32 ] inferred the unloading LV geometry used Gaussian process and further statistically learned the infarct shape and size on LV performance in patients extracted from a public database [ 33 ]. More recently, Longobardi et al [ 34 ] predicted left ventricular contractile function via Gaussian process emulation in aortic-banded rats, the Bayesian history matching was applied to constrain the initial parameter sets in order to exclude those points which generate non-physiological biomechanical models. They further performed a Sobol sensitivity analysis using the trained emulator.…”