Biomedical Engineering Reference
In-Depth Information
the artifact component, one may also remove some part of EEG signal.
James and Gibson [James and Gibson, 2003] proposed temporally constrained
ICA which extracts signals that are statistically independent, yet which are con-
strained to be similar to some reference signal which can incorporate a priori in-
formation related to a particular kind of artifact. The authors argued that the relative
morphology of the reference is relatively unimportant as long as the temporal fea-
tures of interest are captured. However, the phase of the reference must be closely
matched to that of the desired underlying component. The method was applied to
remove artifacts from EEG and MEG signals. An example of removal of different
kinds of artifacts form MEG signals by means of the above described method is
shownin
Figure4.7.
The BSS approach for artifact rejection, called second order blind inference
(SOBI), was proposed by [Joyce et al., 2004]. ICA algorithm assumes that the com-
ponents are statistically independent at each time point, which is not a well-founded
assumption. SOBI considers the relationship between component values at differ-
ent time lags and insists that these values be decorrelated as much as possible. The
remaining correlated components can isolate highly temporally correlated sources,
which according to the authors, is a crucial feature for ocular artifact detection. An-
other advantage of SOBI put forth in [Joyce et al., 2004] is that it uses only second
order statistics that can be more reliably estimated than higher order statistics used
in the ICA approach. It seems that different methods devised in the framework of the
BSS approach require further tests.
In spite of the large number of works devoted to the methodology of artifact re-
moval very few of them were tested on sufficient experimental material and were
developed into fully working systems. The semi-automatic system for rejecting arti-
Makeig, 2004]. It includes several routines to identify artifacts based on detection
of: extreme values, abnormal trends, improbable data, abnormally distributed data,
abnormal spectra. It also contains routines for finding independent components.
Please note that if the aim of the further analysis is finding the causal relations
(Sect. 3.3.2) between the signals from a multichannel set, the subtraction of the ar-
tifacts cannot be used as a preprocessing step, since it disturbs the phase relations
between the channels of the process.
The elimination of artifacts in the overnight polysomnographic sleep recordings
is of especial importance because of the very large volume of data to be inspected—
whole night sleep recording as printed on standard EEG paper would be over half a
kilometer long. There was an attempt to design the system of artifact elimination in
sleep EEG in the framework of the EC SIESTA project [Schlogl et al., 1999, Anderer
et al., 1999], but this system was far from being robust and automatic and it was
tested only on selected 90 min. epochs taken from 15 recordings.
A fully automatic system of artifact detection in polysomnographic recordings de-
vised by [Klekowicz et al., 2009] was tested on over 100 polysomnographic record-
ings of overnight sleep. The parameters reflecting explicitly the likelihood that a
given epoch is contaminated by a certain type of artifact were calculated for the fol-
lowing types of artifacts: eye movements, eye blinks, muscle artifacts, ECG, low
Search WWH ::
Custom Search