Artifact correction in continuous recordings of the electro- and magnetoencephalogram by spatial filtering
Nicole Ille · MADOC (University of Mannheim) · 2001
In this thesis, two novel spatial filter approaches for artifact correction in continuous recordings of the electro- (EEG) and magnetoencephalogram (MEG) are presented. The spatial filters differ from earlier approaches such as multiple source eye correction (MSEC) or independent component analysis (ICA) in the way artifact and brain signal topographies are estimated. Comparable to MSEC, artifact topographies are derived from single or averaged artifact prototypes. In order to model brain signal topographies, two novel concepts are introduced: the preselection approach and spatially constrained ICA (SCICA). In the preselection approach, a relevant number of eigenvectors is extracted from an artifact-free subset of the data segment. The subset is obtained by excluding sample vectors from the original data segment that exceed a certain amplitude or correlation with the predefined artifact subspace. In SCICA, brain signal topographies are estimated from the whole data segment. SCICA uses the prior knowledge about artifact topographies and combines this spatial information with the temporal-statistical strategy of ICA to estimate brain signal topographies. Starting from n artifact topographies, the artifact-contaminated data segment of m channels is decomposed iteratively into m-n further components until all waveforms are maximally independent under the spatial constraint. An algorithm to perform the SCICA decomposition is introduced. Spatial filtering of EEG/MEG segments applying preselection and SCICA shows that both approaches can remove artifacts completely without distortion of relevant brain signals. However, the preselection approach depends crucially on the subjective choice of parameters such as the correlation threshold. Advantages of SCICA over ICA are demonstrated.