Applying Independent Component Analysis to the Artifact Detection Problem in Magnetoencephalogram Background Recordings

IntroductIon The analysis of the electromagnetic brain activity can provide important information to help in the diagnosis of several mental diseases. Both electroencephalogram (EEG) and magnetoencephalogram (MEG) record the neural activity with high temporal resolution (Hämäläinen, Hari, Ilmoniemi, Knuutila, & Lounasmaa, 1993). Nevertheless, MEG offers some advantages over EEG. For example, in contrast to EEG, MEG does not depend on any reference point. Moreover, the magnetic fields are less distorted than the electric ones by the skull and the scalp (Hämäläinen et al., 1993). Despite these advantages, the use of MEG data involves some problems. One of the most important difficulties is that MEG recordings may be severely contaminated by additive external noise due to the intrinsic weakness of the brain magnetic fields. Hence, MEG must be recorded in magnetically shielded rooms with low-noise SQUID (Superconducting QUantum Interference Devices) gradiometers (Hämäläinen et al., 1993). Unfortunately, the external noise is not the only undesired signal in MEG data. In these recordings, noncerebral sources (i.e., artifacts) appear mixed with the useful brain signals. The artifacts could bias the brain activity analyses, since both kinds of signals may have similar power and share the same frequency band. In MEG data, the main artifact is the cardiac one, whose amplitude is usually high enough to be visible in raw can disguise the actual brain activity in long recordings. Finally, power line noise may also be present in MEG signals (Hämäläinen et al., 1993). Diverse approaches have been used to detect and reject artifacts from EEG and MEG data, such as epoch rejection, regression methods (Croft & Barry, 2000), or principal component analysis (PCA) (Sadasivan & Dutt, 1996). In the early 1990s, a new method to obtain a blind source separation (BSS) became available: the independent component analysis One of the main advantages of ICA over other approaches is that artifacts must not be orthogonal to brain signals, and reference channels are not needed, although they can help to detect the artifacts (Barbati,