Transfer entropy calculations using GPUS for determining epilepsy focus -

Abstract

About one third of epilepsy patients do not respond well to drug treatment. For part of this population, surgical intervention is a promising solution whereby the brain tissue causing seizure initiation is removed (the seizure onset zone or SoZ). Here, several clinical tests are normally conducted to detect and highlight the SoZ, including scalp EEG, MRI, SPECT, PET. Pre-surgical intracranial EEG (IEEG) is collected from multi-electrode arrays placed on the cortical surface to improve SoZ detection. Among the various multivariate techniques used to study the collected IEEG, conditional transfer entropy is very effective in finding causal relationships between the signals in recorded channels due to its generality and exhaustiveness. It is, however computationally very expensive so that using traditional CPUs is impractical. In this thesis, we used GPUs where thousands of cores could run in parallel to implement multi-variate CTE studies. Moreover, we focused on code optimization where the same functions could be implemented in untraditional way so that faster execution is achieved. Part of the code optimization is memory management where different types of memories with different speeds are available on GPUs. We reduced time needed from few days to few hours, thereby rendering the ability of applying CTE to IEEG data more readily attainable for research in SoZ prediction as well as other studies of high-dimensional causal interaction.