Page 26 - Artificial Intelligence for Computational Modeling of the Heart

Preface xxvii

are poised to become more and more automatic, while new de-
cision support tools will appear. For these reasons, healthcare is
believed to be one industry to be significantly disrupted by AI.

Computational modeling meets statistical

learning
Interestingly, both fields of computational modeling and AI
started almost at the same time, and both benefited from the re-
cent access to data and computational power to unlock their po-
tential in many applications. A recent scientific direction consists
in exploring ways to augment computational models with statis-
tical methods, and conversely. A first objective is model reduction
for fast but accurate simulations. The idea is to use a data-driven
approach, typically manifold learning, to reduce complex mod-
els into a reduced set of equations that is easier to tract while
capturing the overall trend of the physics under considerations.
In [37] for instance, the authors applied dimensionality reduction
to complex ionic models of sarcomeres to simplify and speed up
their computational solution. The expressive power of deep learn-
ing however could also allow to predict the dynamics of complex
physical systems directly from observable input features. For in-
stance in [38], the authors show that a neural network can simu-
late the universe evolution under gravity, a complex N-body sys-
tem, as accurately as the most sophisticated simulator while be-
ing several orders of magnitude faster. Not only, the deep neural
network could also predict universe evolution under conditions
not observed during training. Similar approaches are being in-
vestigated for computational fluid dynamics [39] and biomechan-
ics [40], mostly in the field of computer graphics, to develop fast
and accurate solvers.
Another fascinating area is the use of data and computational
models to train neural network systems that predict the occur-
rence of scenarios of interest. In [41], the authors train a model
to predict the next occurrence of earthquakes in a region, given
sparse and noisy seismographic data. Interestingly, the authors
found that the output of the neural network could be explained in
a physical sense. This highlights an interesting trend, where data-
driven models informed by training samples produced by com-
putational models can help generate new hypotheses, that could
then be further analyzed through computational modeling, and
even be used to augment these models.
In the same spirit, this book describes how computational
modeling and AI can be integrated for the study of human phys-