Page 209 - Artificial Intelligence for Computational Modeling of the Heart

Chapter 5 Machine learning methods for robust parameter estimation 181

input features. In this work, we used evaluations of the forward
model with specific values of the parameters to estimate the max-
imal range of model outputs for each patient, thus being able
to normalize the ECG features and removing the expected effect
of patient-specific geometry. Alternatively, geometric information
could be used explicitly as input to the regression. More advanced
neural networks models could also be employed to learn encod-
ings of the geometric relationship between heart and torso, with
the goal to further improve electrical diffusivity estimates.
We then described a personalization approach based on re-
inforcement learning. Inspired by how humans approach per-
sonalization, this method first attempts to understand the model
characteristics by means of data-driven exploration. The obtained
knowledge is then utilized to learn a personalization strategy us-
ing RL. We describe a generic solution that requires only minimal
user input (parameter ranges, authorized actions, number of rep-
resentative states) to start learning in an autonomous way how to
personalize a model.
The approach was tested on two challenging personalization
tasks in cardiac computational modeling. Only minimal tun-
ing of the algorithm hyper-parameters was required, suggesting
good generalization properties. At the same time, we observed
increased robustness compared to state-of-the-art optimization
approaches. Training data generation can be computationally de-
manding. However, training is performed only once in an off-line
phase. On-line personalization, on the contrary, is not paralleliz-
able in its current form, because, just like for most state-of-the-
art methods, the parameters for each forward model run depend
on the outcome of the previous iteration. However, we observed
faster convergence with less evaluations of the forward model
(which can be computationally demanding), even when com-
pared to manually engineered algorithms. The on-line overhead
introduced by the RL-based method (convert data into an MDP
state, then query an action policy) is negligible.
Based on these promising results, we conclude that apply-
ing modern AI technology to model personalization could enable
a unified framework for personalization of any computational
physiology model. Eventually, such a framework may speed up
progress in the computational modeling community as it could
remove the need for tailored design and engineering of complex
optimization techniques and promote the adoption of machine
trained algorithms encoding the required domain-specific exper-
tise through learning from data.