Page 205 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 5 Machine learning methods for robust parameter estimation 177

Figure 5.8. EP results: personalization success rate (blue, dark gray in print
version) and average number of iterations (red, mid gray in print version). Left:
performance for increasing number of training data. Each dot represents results
from one experiment (cross-validated personalization of all 75 datasets), solid
lines are low-pass filtered means. Right: Performance of both reference methods.
Each shade represents 10% of the results, sorted by performance.

decreased when there was more training data available: 36.2 at
3
4
5
10 samples, 31.5 at 10 samples, or 31.8 at 10 samples.
These results suggested that the proposed method can achieve
similar performance as an advanced, manually engineered
method. Further, the number of required training samples was
not excessive, instead well manageable and computable in a rea-
sonable time-frame.

5.3.4 Application to whole-body circulation
The second considered task is the personalization of a lumped-
parameter whole-body circulation (WBC) model from pressure
catheterization and volume data. As some of the datasets from the
EP experiments lacked catheterization data, only a subset of 56
patients was used in the following experiments. The WBC model
is described in section 1.4.2 and [387,386]. In short, it consists in
time-varying elastance models for the four chambers of the heart,
the valves are modeled through a resistance and an inertance, and
Windkessel (WK) models are used for the systemic and pulmonary
circulation: 3-element WK for arterial and 2-element WK for ve-
nous circulation, respectively.
The goal of this experiment was to analyze personalization per-
formance of the RL-based method for the part of the model de-
scribing systemic circulation, in various configurations. In partic-
ular, we evaluated the method for increasing number of parame-
ters to tune and objectives to match, from two to six parameters
(setups 2p, 3p, 5p, 6p), see Table 5.5 for details. The reference val-
ues δ that define the action space A were set to 0.5% of the admis-
sible parameter range Ω. Personalization objectives are listed in
Table 5.6 and illustrated in Fig. 5.9.