Exploratory and Data Visualization Methods for High-Dimensional Brain Signals

• Activity Number: 479 - Statistical Analysis of Complex Imaging Data
• Type: Invited
• Date/Time: Wednesday, August 2, 2017 : 10:30 AM to 12:20 PM
• Sponsor: IMS
• Author(s): Hernando Ombao* and Yuxiao Wang and Chee-Ming Ting
• Companies: KAUST and UC Irvine and University of California, Irvine and KAUST
• Keywords: Dimension reduction ; Multivariate time series ; Spectral analysis ; Principal components analysis ; Brain signals

Abstract

We present exploratory approaches to modeling dependence between components of high-dimensional brain signals. To address high dimensionality, we derive region-specific factors which are lower dimensional signal summaries of electrical activity in each brain region. Initial plots of both electroencephalograms (EEGs) in humans and local field potentials (LFPs) in rats exhibit a high degree of multicollinearity between channels or nodes that are projected from the same anatomical region. These suggest that it would be sensible to summarize the activity in each region using only a few number of factors. In this work, we derive summary factors that minimize the expected squared error between the reconstructed signals (obtained by expanding the summary factors) and the original signal. These factor signals are filtered versions of the original high dimensional signals where the coefficients are obtained from principal components analysis (PCA) of the spectral matrix. Dependence between brain regions are then explored through the regionally-derived summary factors. We conclude this talk by a demonstration of new data visualization and exploratory toolbox applied to EEGs and LFPs.