Spatio-temporal Covariance Model for Human Brain MEG Analysis

December 4, 2003

Date: Thursday, December 4th
Time: 11am-12:15pm
Location: Woodward 149
Sergey Plis, (email) Department of Computer Science, UNM Abstract: There are many problems in data analysis or signal processing in which one wants to discern what is different from a "signal" state and a "background" state. This is also true in the brain mapping modality of Magnetoencephalography (MEG) in which one measures the magnetic field outside a person's head generated from the electrical currents from active sets of brain neurons. Here the signal state is the brain's response to a given stimulus and the background state is ongoing neural activity not time-locked to that stimulus. The ability to characterize the statistical nature of the background activity is important for correctly inferring the brain's response. This background is sufficiently complex and correlated that it is not practical to construct its full covariance using conventional means. We present a model for the spatiotemporal noise covariance that has a small enough number of free parameters to be estimated correctly from reasonable amounts of data. In addition this model has the property that the inverse of the covariance is easy to calculate, which is important for using this covariance in an analysis. Our approach models data with spatiotemporal dependencies better than an other approach (also to be presented) and is only linearly more complex in the number of parameters.