Structural abnormalities of the myocardium account for a large part of cardiac arrhythmia and sudden cardiac death. Accurate representation of these abnormalities in simulations of cardiac electrophysiology requires spatial resolutions in the order of a cell diameter. We tested if this would be feasible with a model of the whole human ventricles.
This study was prompted by the opportunity to use a new cluster of 2292 compute nodes with 128 cores each, and a total of 573TB memory. We integrated a monodomain reaction-diffusion equation with realistic membrane dynamics on a geometry of the human ventricles at 200, 100, 50, and 25 micrometer resolution. At the finest resolution the mesh had 11 billion nodes. Both explicit and implicit Euler integration was tested. Simulations were performed with the Propag-5 software, adapted to run its initialization phase with a limited number of processes, to avoid problems in the mesh partitioner and geometry input.
Scaling of the explicit method was over 80% efficient on 131072 cores for model resolutions of 100 micrometer and finer. The implicit method was generally slower. It matched the explicit method only at 25 micrometer resolution and on a relatively small number of cores.
These results show that whole-heart simulations with 25 micrometer resolution are technically feasible today, and that they would probably be performed best with an explicit integration method, despite the large number of time steps that such a method needs to take. This is because implicit models scale less well, and this disadvantage grows with the model size.