Efficient methods for large multivariate time series connectivity analysis.

摘要

Progressive advances in electroencephalography (EEG) and magnetoencephalography (MEG) acquisition hardware design allow for collection of signals at ever increasing temporal and spatial resolutions. At the same time, computationally complex analysis methods are also becoming increasingly popular. Therefore, such implementations of these methods that take advantage of the newest hardware and the latest developments in computational algorithms is of paramount importance to analyze EEG and MEG signals efficiently.;Moreover, we conduct exploratory analysis on EEG and MEG time series arising from various cognitive experiments aimed at creating representative profiles of brain activation for different populations. Using coherence as a measure of connectivity and investigating the power content of signals, we create brain activation profiles of various populations involved in the tasks, and identify those regions in the brain whose activation is significantly different under varying experimental protocols. Finally, we apply principal component analysis to the concentration of dopamine D1 and D2 receptors in the brain to study how localization of these neuroreceptors changes before and after working memory training and find that our data-driven approach corroborates with a study based on a priori regions of interest.;In this dissertation, we present computational methods and optimizations that enable computation of large multivariate models for EEG/MEG analysis. In particular, we focus on the implementation of the Granger causality algorithm on computer clusters with several nodes and multiple cores per node. We explore the applicability of both shared and distributed memory programming paradigms to enable the analysis of arbitrarily large data sets. Our solutions are validated using a variety of clinical data sets including EEG data from subjects suffering from traumatic brain injury and MEG data from normal subjects participating in a study of affect as well as MEG data from autistic subjects. Computation of brain activation profiles using Granger causality as a measure of connectivity among the various brain regions has resulted in reduction of computation time by up to 98.5%. In addition to Granger causality, we parallelize an iterative artifact removal algorithm based on independent component analysis that we have developed in-house. This has resulted in a speedup of an order of magnitude and it can now provide quasi real time performance with hundreds of recording channels and a sampling rate of more than 128 Hz.