Independent Component Analysis and Signal Separation

Abstract

ICA is often employed for the analysis of MEG stimulus experiments. However, the assumption of independence for evoked source signals may not be valid. We present a synthetic model for stimulus evoked MEG data which can be used for the assessment and the development of BSS methods in this context. Specifically, the signal shapes as well as the degree of signal dependency are gradually adjustable. We illustrate the use of the synthetic model by applying ICA and independent subspace analysis (ISA) to data generated by this model. For evoked MEG data, we show that ICA may fail and that even results that appear physiologically meaningful, can turn out to be wrong. Our results further suggest that ISA via grouping ICA results is a promising approach to identify subspaces of dependent MEG source signals.