Department of Biostatistics Seminar/Workshop Series
Statistical Analysis of Brain Signals
Hernando Ombao, PhD
Associate Professor of Biostatistics, Co-Director, Biostatistics Graduate Program, Brown University
Wednesday, March 3, 1:30-2:30pm, MRBIII Conference Room 1220
Brain signals collected from designed experiments and observational studies are realizations of complex spatio-temporal processes. In this talk, we shall describe some current collaborative projects. In the first project, the main goal is to describe and estimate the brain network involved during a visual-motor experiment using electroencephalograms. We characterize cross-dependence between EEG channels using partial coherence (which is utilized to identify frequency bands that drive the linear cross-association). Jointly with Mark Fiecas, we developed a novel semi-parametric estimation method for brain network. We also briefly describe a project (with Sebastien Van Bellegem) where we develop a new mixture model for characterizing stimulus-induced changes in the cross-dependence structure. Our mixture consists of vector autoregressive models and the weights capture the effect of the stimulus. The parameters are estimated using the E-M algorithm. One crucial problem prevailing in the analysis of brain signals is the notion that the type I error is held fixed even when the size of data and amount of information increases. Jointly with Jeffrey Blume and Hakmook Kang, we develop the Likelihood approach which overcomes this limitation. In preliminary work, spatially-temporally correlated fMRI time series were generated and empirical FDR and Likelihood approaches were compared. The results suggest that the probability of false positives under the likelihood approach is typically half that of empirical FDR. In addition to identifying more true positive voxels, the likelihood paradigm yields dramatically fewer false positive voxels.