Detection and Termination of Broken-Spiral-Waves in Mathematical Models for Cardiac Tissue: A Deep-Learning Approach

Defibrillation, the elimination of pathological waves of electrical activation in cardiac tissue, plays an important role in the elimination of life-threatening cardiac arrhythmias like ventricular tachycardia (VT) and ventricular fibrillation (VF). We develop a deep-learning method, which uses a convolution neural network (CNN), to develop a new defibrillation scheme applicable in 2D tisue. We begin by training our CNN with a huge dataset of spiral waves $\left( \mathcal{S} \right)$ and non-spiral waves $\left( {\mathcal{N}\mathcal{S}} \right)$ that we obtain from our direct numerical simulations (DNSs) of a variety of mathematical models for the propagation of electrical waves of activation in cardiac tissue. Our trained CNN can distinguish between $\mathcal{S}$ and $\mathcal{N}\mathcal{S}$ patterns; in particular, it also detects a broken spiral wave as $\mathcal{S}$. We demonstrate how to use our CNN to develop a heat map, from a broken-spiral-wave image, that yields the approximate locations of these spiral cores. We develop a defibrillation scheme that applies current, with two-dimensional (2D) Gaussian profiles of standard deviation (σ), centred at square lattice sites (N G × N G ) imposed on the simulation domain (N ×N); the amplitudes of these Gaussians are taken from the heatmap. We explore the dependence of our Gaussian defibrillation scheme on a noisy image, which closely mimics the noisy optical image data.