Extraction of event-related potentials from electroencephalography data
Andriy Ivannikov · Jyväskylä University Digital Archive (University of Jyväskylä) · 2009
The research work reported in this thesis addresses the issues related to denoising of event-related potentials (ERP) in multichannel electroencephalography (EEG) data. The main idea behind the ERP denoising methods presented in this thesis lies in separating ERP and noise subspaces according to the linear instantaneous mixing model. When subspaces are extracted, the denoising of the channels of measurements is reached by the inverse transformation of the previously obtained ERP components ignoring components related to the noise subspace. The emphasis of the thesis is on finding appropriate problem-specific criteria, which allow ERP and noise components in multidimensional EEG data space to be reliably distinguished and, thus, for the separation of ERP and noise sub-spaces by finding the basis vectors that span them. The criteria that have been studied are based on exposing the data to some modification that influences signal and noise subspaces or, more precisely, signal and noise constituents of the data, differently. We explore those subspace-specific changes that are seen on the level of second-order statistical properties of the data. Namely, the two covariance matrices of data before and after the modification are compared. Moreover, we concentrate our attention on those modifications that exploit the data which have three dimensions of variability: channels, time samples, and trials. Yet the scope of this thesis goes beyond a sole proposition of the subspace separation criteria and touches also on such topics as (1) practical aspects of application of the denoising methods, (2) validation of the data and results, 3) developing a comparison framework, (4) analysis and interpretation of the results, (5) elaboration of suggestions and recommendations for improving the performance of the denoising methods.