Abstract:
Cardiovascular disease remains the leading cause of death globally. To aid in guiding
patient therapy, predicting prognosis, and improving diagnosis related to cardiovascular diseases, patient-specific computational models of the heart have been developed.
Personalisation of these models requires the estimation of input parameters, such as
the mechanical properties of the myocardium, from clinical measurements. These mechanical properties, such as the stiffness of the heart wall could be used as biomarkers,
for example, for early detection of heart failure where the structure of the myocardium
changes and thereby its mechanical properties change. Detecting these changes before irreversible damage would provide a novel avenue for diagnosing, treating, and
monitoring disease. Robust estimation of the constitutive parameters that describe the
mechanical behaviour of the tissues is difficult and confounded by the inherent uncertainties associated with real-world clinical measurements. This motivates the need
for quantifying the effect of uncertainty in clinical data on the constitutive parameter
estimates. Bayesian statistical methods and parameter estimation methods, specifically Bayesian hierarchical modelling (BHM) and Approximate Bayesian Computation (ABC), have been used in other scientific fields to perform uncertainty quantification (UQ), but these have not yet been widely applied in cardiac mechanics.
This thesis presents the use of Bayesian statistical methods to investigate how uncertainties in clinical measurements affect the estimation of constitutive parameters
of the left ventricle (LV) during the passive filling phase of the cardiac cycle. Two
approaches were developed and implemented to perform parameter estimation and
UQ. One used a BHM framework along with a statistical surrogate model of cardiac
mechanics, while the other used a simulation-based approach and full FEM cardiac
mechanics, specifically using ABC. The two approaches were applied to a cardiac mechanics model to quantify uncertainty in myocardial constitutive parameter estimates
due to i) noise in the displacement of the left ventricular wall (observed data); and ii)
heartbeat-to-heartbeat variability in LV measurements from catheters that were applied
as boundary conditions in the cardiac mechanics model. In particular, the practical
identifiability of c1 (intrinsic myocardial stiffness) and c2 (non-linearity in the direction of the myocardial fibres) of the Guccione constitutive relation, which describes
the passive response of the myocardium, were investigated. Numerical experiments
were performed to compare and contrast constitutive parameter estimates from the two
methods against a ground truth. Both Bayesian statistical methods were found to be
useful and effective in performing parameter estimation and UQ to account for multiple sources of error, especially when prior knowledge from a specific patient was
included in the analysis (e.g. the shape of their LV). The results show that noise in the
observed data and the beat-to-beat pressure variability contribute to the reduction in
the precision of estimates of the constitutive parameters. Overall, this work can help
further eliminate the barriers to the clinical translation of cardiac mechanics models for providing mechanical biomarkers for early diagnosis of diseases, such as heart failure