Physics-informed Machine Learning for Precision Cardiology

Cardiac modeling for precision medicine represents a transformative approach in cardiology. The development of patient-specific models is however hindered by significant challenges. Cardiac simulations are resource-intensive, with many patient-specific parameters; clinical data is scarce, sparse, and multi-modal. Consequently, relying solely on data-driven or model-based methods falls short in the creation of digital twins.

In this presentation, we will explore solutions through a physics-informed machine learning methodology. Our first challenge involves deducing the cardiac electrical properties—such as the conductivity tensor, sites of early activation, and the Purkinje network—from clinical electrical measurements. The second focus is on predicting the likelihood of atrial fibrillation in a detailed anatomical model of the human atria without the need of running simulations.