Using MATLAB and GPU Acceleration to Perform High-performance EEG analysis on Supercomputers

In this poster, we demonstrate how supercomputers can help EEG researchers to perform sophisticated MATLAB-based data processing and analysis tasks, and report on our parallel GPU algorithm development efforts that can further reduce the execution time of these operations. Traditional EEG analysis is based on trial averaging. Stimulus evoked potential distribution is measured on the scalp for each trial, and then trials are averaged to reduce random noise and amplify stimulus-locked response. The analysis of data gathered in this manner is hampered by several problems. (1) Noise and artefacts are always present in EEG registrations; careful data cleaning and artefact removal is crucial if one is to produce correct average potentials. (2) Due to volume conduction (Nunez & Srinivasan, 2005) potential measured on individual electrodes is a mixture of different underlying cortical sources that can provide a misleading view of the location of sources. (3) Oscillatory mechanisms at different frequencies represent neural activation/inhibition and communication, necessitating detailed time-frequency analysis. (4) Various source localisation techniques (Grech et al., 2008) should be used to find the true location of cortical sources. These steps require the execution of a series of sophisticated methods, e.g. Wavelet Transform, Independent Component Analysis (ICA), and dipole localisation that each require considerable time. HPC resources can help researchers to speed up data processing but the migration of local scripts to supercomputers might be challenging. The aim of our work reported here is twofold, to speed up traditional EEGLAB script execution with parallel job execution and to develop high-performance GPU versions of the most time-critical operations to shorten individual pipeline runtimes. Acknowledgements We acknowledge the support of the DKF (Digital Government Development and Project Management) Ltd. for providing access to the Hungarian supercomputer Komondor used for EEG processing.