Page 208 - Artificial Intelligence for Computational Modeling of the Heart

180 Chapter 5 Machine learning methods for robust parameter estimation

Convergence analysis For each WBC setup, the proposed method
was evaluated based on varying amount of training data (n samples =
5
0
10 ...10 ) and using leave-one-patient-out cross-validation. We
used the same dynamic limits for number of iterations as for
the reference method. The results are presented in the graphs in
Fig. 5.10. Performance increased steadily with the amount of data
until reaching a plateau. The more complex the problem, i.e. the
more parameters and objectives involved in the personalization,
the more training data was needed to reach the plateau with best
performance. As an example, 80% success rate was reached with
less than n samples = 50 training samples per dataset in the 2p setup,
while for 6p, almost 90× as many samples were needed.
Compared to the benchmark, the RL-based method reached
equivalent or better success rates (up to 11% higher), while si-
multaneously being more efficient in terms of run-time. In the
3
2p setup, if n samples ≥ 10 , the RL-based converged after 3 itera-
tions on average, as compared to 22.6 iterations for the reference
method. This translates to a seven-fold speed-up. For all of the
other setups (3p, 5p, and 6p), significant improvements can also
be observed, with speed-up of at least a factor of 1.8.

5.4 Summary
This chapter presented two methods based on machine learn-
ing to estimate tissue properties. We first presented a method
based on polynomial regression to estimate the electrical dif-
fusivity of the myocardium from 12-lead ECG features, namely
QRS duration and electrical axis. The approach is as accurate
as traditional inverse problem methods, while it provides a first
order approximation of parameter uncertainty and is extremely
computationally efficient. The method, while illustrated on a
mono-domain model of cardiac electrophysiology, is generic and
could be applied to any type of cardiac electrophysiology models
(mono-domain, bi-domain, Eikonal, etc.). Although the focus of
this discussion was on cardiac depolarization and electrical dif-
fusivity, a similar strategy could be used to estimate the tissue
parameters related to cardiac repolarization, like action poten-
tial duration and restitution curve. Furthermore, the method can
be enhanced by providing more features as input without loss of
generality and hence reduce the uncertainty in estimated cardiac
parameters.
One important learning from this approach is the importance
of geometry in the data-driven approach. The relative position
of the heart in the patient’s torso has a noticeable impact on the
ECG traces. It is therefore crucial to take the geometry out of the