Deep Learning-Based Emulation of Human Cardiac Activation Sequences

Abstract

The vision of digital twins for precision cardiology is to combine expert knowledge and data of patients’ cardiac pathophysiology with advanced computational methods, in order to generate accurate, personalised treatment strategies. When studying cardiac electrophysiology, the twinning pipeline commonly requires a large amount of simulations, e.g. when exploring parameter spaces for personalisation or when scaling up to large cohorts of virtual patients in Big Data studies. In these cases, state-of-the-art methods are computationally expensive, even when applying relatively fast algorithms such as the Eikonal model. In this work, we investigate the performance of a U-Net-based model for electrical excitation throughout the human ventricles. The approach provides the advantage of reducing the input parameter space by representing anatomical and electrophysiological properties of the heart in a standardised three-dimensional space. Results demonstrate the ability of the model to emulate the Eikonal simulation scheme and predict cardiac activation time maps with average accuracy of 4.7 ms RMSE and an improved performance at point of prediction, yielding results up to 500 times faster. This new method provides promising results for personalised simulations of cardiac propagation in large cohorts of human heart models.