The development of low-amplitude defibrillation schemes, for the elimination pathological electrical waves of activation, in cardiac tissue, is a major challenge in the treatment of life-threatening cardiac arrhythmias. We com- bine the dataset generated from extensive direct numerical simulations (DNS) using mathematical models for cardiac tissue with a deep-learning method to construct a new defibrillation scheme to eliminate unbroken and broken spiral waves, which are the mathematical analogs of ventricular tachycardia and ventricular fibrillation. We use a convolution neural network (CNN), where we train the CNN to distinguish spiral-wave patterns (S) from those that do not have spiral waves (N S). For the training, the dataset we use is ' 25000 different pseudocolor plots of the transmembrane potential Vm obtained from DNS of both two-variable and ionically realistic models for cardiac tissue.
We have checked that our trained CNN can, distinguish between patterns: with spirals, and patterns that do not have spirals. Next, we show how of the broken spiral waves, which has high intensity in regions with spiral-wave cores; this heatmap plays a central role in our new defibrillation scheme: We first show that a control (defibrillation) current with a two-dimensional (2D) Gaussian profile (with width σ ' 75% of linear size of our simulation domain) eliminates the spiral wave when we apply it exactly at the spiral core. We then demonstrate that the defibrillation-current profile, comprising of 2D Gaussians on a square lattice, whose amplitudes are proportional to the heat-map intensity at a given point in our simulation domain, eliminates broken spiral waves.