Page 190 - Artificial Intelligence for Computational Modeling of the Heart

162 Chapter 5 Machine learning methods for robust parameter estimation

Figure 5.1. A computational model f is a dynamic system that maps model input
parameters x to model state (output) variables y. The goal of personalization is to
tune x such that the objectives c, defined as the misfit between y and the
corresponding measured data z of a given patient, are optimized (the misfit is
minimized).

sign of such methods is trial-and-error-driven and involves te-
dious manual tuning, which results in methods that are model-
specific rather than generic, and still do not necessarily ensure
generalization to unseen data.
Several machine learning techniques have been explored in the
community to cope with the known limitations of optimization
techniques. For instance, Jiang et al. [381] demonstrated a method
to learn a mapping between body surface potentials maps and
epicardial potentials, without requiring complex optimization
procedures. Bayesian inference was also explored in [82]toesti-
mate tissue conduction properties from invasive local activation
time measurements. Finally, model reduction approaches as dis-
cussed in the previous chapter were applied to estimate a signifi-
cantly faster surrogate model, which was then applied to estimate
model parameters and their uncertainty [201]. These techniques
demonstrated superior computational efficiency compared to tra-
ditional optimization approaches, and also the capability of pro-
viding more information about the estimated parameters (such as
uncertainty).
On the other hand, when experienced humans are tasked with
manually individualizing a model from patient data, they almost
always succeed. There are several reasons why human experts are
often superior to automatic methods. First, human experts often
have the intuition of how a model behaves based on prior knowl-
edge of the modeled physiology. Second, an expert usually knows
the model design, assumptions, limitations and implementation
details. Finally, experience built from past similar attempts on
other datasets gives advantages. A combination of these and po-
tentially other factors can enable a human expert to individualize
models efficiently, even based on unseen data.
Motivated by the promising results of machine learning ap-
plied to parameter estimation, this chapter describes two ap-
proaches applied to concrete cardiac modeling use cases: a su-
pervised learning method based on polynomial regression to di-
rectly estimate cardiac parameters from 12-lead ECG features, and