Page 188 - Artificial Intelligence for Computational Modeling of the Heart

160 Chapter 4 Data-driven reduction of cardiac models

high-frequency oscillations were observed, which lead to large er-
rors. Possible reasons for this effect could be insufficient training
data and network capacity. Another reasons could be that the in-
put features do not include explicitly information about the neigh-
boring points, but rather only account for neighbors indirectly
through the computation of internal forces. Explicitly adding fea-
tures describing the neighborhood of each point could simplify
model training in more complex scenarios. In addition, high er-
rors might arise from the accumulation of the prediction error
over a number of time steps. Although the magnitude of the input
features is scaled to be within the range of observed data dur-
ing training, error propagation could cause the input features to
progressively drift towards values that are poorly sampled in the
training database. A more detailed analysis will be the focus of fu-
ture work.

4.4 Summary

This chapter presented three data-driven approaches to esti-
mate surrogate models or accelerate numerical methods, with the
goal to perform real-time hemodynamics, electrophysiology and
biomechanics simulations. A deep learning model for FFR assess-
ment was introduced, which yielded high accuracy in identifying
functionally significant coronary artery lesions while by-passing
time-consuming computational fluid dynamics simulations. An
approach to estimate a meta-model of a complex cellular model,
while preserving its physiological fidelity, has been introduced
and demonstrated on a cardiac electrophysiology use case. Fi-
nally, a method to accelerate time integration in numerical solvers
of biomechanics has been described. In all three cases, the combi-
nation of existing models, to create the training data in particular,
and astute integration of neural networks in the simulation work-
flow allowed to achieve high computational efficiency while being
as accurate as the original models. These examples are only a few
of the many potential applications of the combination of artificial
intelligence and computational modeling. Among other research
areas that are opening in the same context, it is worth mentioning
the integration of generative neural networks with computational
models [375], and the estimation of new biophysical laws that can
take into account data noise and uncertainty directly, for more ac-
curate predictions [376].