Magnetoencephalographic data analysis: Dynamical embedding followed by independent component analysis

With most nemophysiologic recordings, the recording of noise along with the desired signal is an accepted reality. Magnetoencephalographic (MEG) data recordings are no exception because of the very small signal levels compared to those of the background or contaminating noise. Independent Component Analysis (ICA) has been used in previous studies as a means of identifying and removing artifacmal waveforms from the recorded MEG. This includes waveforms such as ECG, eye-movements and 50/60 Hz mains frequency contamination. In general, ICA is performed on multiple-channel segments of MEG where the contaminants are visually selected from amongst the resulting Independent Components (ICs). This is not always practical however, especially when the number of recording channels is large (in excess of 150 channels in our case). The ICA algorithms currently available assume a square-mixing matrix; this implies that there must be as many ICs as there are channels of recorded data. For short segments of multichannel MEG, where the number of underlying generators can be assumed to be very small, much smaller than the number of measurement channels, this has the effect of producing ICs that are statistically independent in the strict sense, but are unrecognisable in the neurophysiological sense. Furthermore, dominant artifacts/signal characteristics in the measurement space tend to dominate in the component space and can result in localised artifacts not being modelled (and hence cannot be eliminated using this method). Using the technique known as Dynamical Embedding (DE) followed by traditional ICA it is possible to scrutinise single channels of MEG data individually. This allows localised artifacts to be removed (e.g. ECG contamination), as well as extracting/isolating localised, small amplitude, independent activity that would otherwise be lost in the multichannel case. DE is based on the presumption that the measured signal is due to some underlying small number of generators, which are non-linearly coupled. The analysis process is a two step one: (a) first a DE matrix is created out of a channel of MEG data through a series of overlapping delay vectors, then (b) ICA is applied to the matrix of delay vectors. The resulting ICs consist of statistically independent components that underlie the generation of the measured signal. DE coupled with ICA helps to extract temporal correlations or structure from de measured signal. The method is applied to various segments of multichannel MEG data recorded at the Wellcome Trust Laboratory for MEG studies at Aston University, on the 151 channel Omega MEG system (CTF Systems Inc.). In each case, the data is down-sampled to a suitable sampling rate (200 samples/set being typical) and mean corrected in the rows and columns. For the most part, single epochs of data are analysed, this is done primarily to investigate the potential of DE and ICA in extracting information from single epoch trials. Where possible, comparisons between results obtained from multichannel and single channel analysis are presented. We gratefully acknowledge funding from EPSRC grant #GR/L94673.