Physics-Informed Fully Connected and Recurrent Neural Networks for Cardiac Electrophysiology Modelling

Abstract

Cardiovascular diseases, the leading cause of death and disability, are often underlined by cardiac arrhythmias. Cardiac electrophysiology models play an increasingly important role in dissecting arrhythmogenic mechanisms and improving treatments, but high computational costs hinder their application. We develop and compare four novel deep learning (DL) models to solve the Fitzhugh-Nagumo (FN) electrophysiology equations efficiently in OD, 1D and 2D. The training datasets were created using numerical solutions of FN equations. Physics-informed neural network (PINN) was based on the incorporation of FN equations into the Loss function, allowing the optimal combination of training data with physical constraints. Another recurrent NN (RNN) with a mean squared error (MSE) loss function was developed as a baseline. The DL models were evaluated using the MSE score. In OD and 1D, similar performances were achieved for all DL models, with a typical MSE of 10-2. In 2D, PINN and RNN succeeded in simulating plane and spiral waves with similar MSE of 10-2. Hence, PINNs can provide an efficient tool for cardiac electrophysiology simulations.