Page 70 - Artificial Intelligence for Computational Modeling of the Heart

40 Chapter 1 Multi-scale models of the heart for patient-specific simulations

the real heart. A multi-step optimization procedure was proposed
in [190] to estimate left ventricular passive myocardial proper-
ties. More recently, a combination of one-dimensional parameter
sweeps with gradient-free optimization was presented in [191]
to estimate myocardial passive material parameters on a single-
ventricle model. A similar multi-step idea was presented in [192],
where parameter sweeps (exhaustive search) are performed iter-
atively on pre-defined one- and two-dimensional sample grids.
Unfortunately, methods based on parameter sweeps quickly be-
come intractable as the number of parameters to optimize grows.
They can therefore enable the exploration of the full landscape of
the objective function only at a pre-defined, usually coarse resolu-
tion. Hierarchical, multi-level approaches have also been explored
to cope with the curse of dimensionality [18,193], as well as multi-
model techniques [194]. Finally, the authors in [195]presented an
approach for iterative regional personalization of cardiac electro-
physiology models, tailored for patients with left-bundle-branch
block.

1.5.2 Data assimilation
Data assimilation approaches frame the personalization prob-
lem in a way that aims at identifying unknown variables using
observations of a dynamical system. At each iteration of the per-
sonalization loop, a forecast of the computational model and the
assimilated observations are compared to estimate the current
goodness of fit between model and data, considering uncertainty
in the state and in the observations. Two categories of data as-
similation methods can be distinguished: variational [196]and
filtering approaches [121,197]. The latter includes the Kalman fil-
tering approach and its extensions (e.g. unscented Kalman filters).
Filtering approaches have been used for global [198]aswellas
regional [121] active contractility estimation, or passive material
parameter estimation [192].

1.5.3 Machine learning
The past few years saw the development of data-driven ma-
chine learning (ML) methods, such as the one-shot global EP
personalization approach described in Section 5.2.1.Inanother
work [199], the authors learn the non-linear relationship between
heart motion, derived from temporal images, and parameters of
electrical propagation. These examples demonstrated that ma-
chine learning approaches can learn the non-linear mappings
from observations to cardiac model parameters. Furthermore,