We show that through careful and model-specific optimizations of their GPU implementations, simulations of realistic, detailed cardiac-cell models now can be performed in 2D and 3D in times that are close to real time using a desktop computer. Previously, large-scale simulations of detailed mathematical models of cardiac cells were possible only using supercomputers. In our study, we consider five different models of cardiac electrophysiology that span a broad range of computational complexity: the two-variable Karma model, the four-variable Bueno-Orovio-Cherry-Fenton (BCF) model, the eight-variable Beeler-Reuter (BR) model, the 19- variable Ten Tusscher-Panfilov (TP) model, and the 67-variable Iyer-Mazhari-Winslow(IMW) model. For each of these models, we treat both their single- and double-precision versions and demonstrate linear or even sub-linear growth in simulation times with an increase in the size of the grid used to model cardiac tissue. We also show that our GPU implementations of these models can increase simulation speeds to near real-time for simulations of complex spatial patterns indicative of cardiac arrhythmic disorders, including spiral waves and spiral wave breakup. The achievement of real-time applications without the need for supercomputers may facilitate the adoption of modeling-based clinical diagnostics and treatment planning, including patient-specific electrophysiological studies, in the near future.
In Proc. of CMSB'11, the 9th International Conference on Computational Methods in Systems Biology, Paris, France, September, 2011, ACM.
*This work was supported by the NSF Expeditions Award CNS-09-26190, the
NSF CSR-AES05-09230 Award and the AFOSR FA-0550-09-1-0481 Award.