Scientific Machine Learning for Cardiac Electrophysiology Applications
DescriptionCardiac modeling for precision cardiology is an emerging technology in clinical practice. Thanks to the sophistication of state-of-the-art electrophysiology models, it is possible to tailor treatments to patient characteristics, thus improving the therapeutic outcome. Patient-specific modeling requires a deep integration of clinical data into existing models. This aspect is, however, not straightforward. Cardiac models are computationally expensive, with several patient-specific parameters of difficult identification. Clinical data is scarce, multi-modal, and sparse in space-time. Thus, neither purely data-driven nor model-driven approaches are optimal in the digital twinning process. In this talk, we will solve two clinically-relevant problems using a physics-informed strategy. The first problem consists in recovering the conductivity tensor in the heart, starting from sparse electric recordings collected by clinicians. We propose FiberNet, a physics-informed neural network method for solving the corresponding inverse problem. We weakly impose the physiological electric propagation with the anisotropic eikonal model. The second application concerns atrial fibrillation inducibility in a complex anatomical model of human atria. We propose a multi-fidelity classifier that learns the inducibility map on a manifold. Finally, we generalize the classifier so that it does not require any new simulation when the anatomy changes.
TimeTuesday, June 2717:00 - 17:30 CEST