Comparison of Complex-Valued Independent Component Analysis Algorithms for EEG Data


Independent Component Analysis (ICA) has been successfully applied to a variety of problems, fromspeaker identification and image processing to functional magnetic resonance imaging (fMRI) of the brain. Inparticular, it has been applied to analyze EEG data in order to estimate the sources form the measurements.However, it soon became clear that for EEG signals the solutions found by ICA often depends on the particular ICAalgorithm, and that the solutions may not always have a physiologically plausible interpretation. Therefore, nowadaysmany researchers are using ICA largely for artifact detection and removal from EEG, but not for the actual analysis ofsignals from cortical sources. However, a recent modification of an ICA algorithm has been applied successfully toEEG signals from the resting state. The key idea was to perform a particular preprocessing and then apply a complexvaluedICA algorithm.In this paper, we consider multiple complex-valued ICA algorithms and compare their performance on real-worldresting state EEG data. Such a comparison is problematic because the way of mixing the original sources (the “groundtruth”) is not known. We address this by developing proper measures to compare the results from multiple algorithms.The comparisons consider the ability of an algorithm to find interesting independent sources, i.e. those related to brainactivity and not to artifact activity. The performance of locating a dipole for each separated independent component isconsidered in the comparison as well.Our results suggest that when using complex-valued ICA algorithms on preprocessed signals the resting state EEGactivity can be analyzed in terms of physiological properties. This reestablishes the suitability of ICA for EEG analysisbeyond the detection and removal of artifacts with real-valued ICA applied to the signals in the time-domain.