Biomedical Engineering Reference
In-Depth Information
average of the uncontaminated trials before artifact removal [Figure 2.9(d)]. This
implies that the corrected recordings contained only event-related neural activity
and were free of artifacts arising from blinks or eye movements.
2.4.3.1 Cautions Concerning ICA-Based Artifact Removal
ICA-based artifact removal also has some shortcomings. First, it is important to dis-
tinguish among artifacts produced by processes associated with stereotyped scalp
maps, for example, eye movements, single muscle activity, and single-channel noise.
These may be well accounted for by a single independent component if sufficient
data is used in the decomposition. At the other end of the scale, nonstereotyped arti-
facts that produce a long series of noise with varying spatial distributions into the
data—for example, artifacts produced by the subject vigorously scratching her
scalp—defy the standard ICA model. Here, at each time point, artifacts may be asso-
ciated with a unique, novel scalp map, posing a severe problem for ICA decomposi-
tion. It is by far preferable to eliminate episodes containing nonstereotyped artifacts
from the data before decomposition because such artifacts can negatively affect the
ICA decomposition even at small amplitudes.
In addition, caution needs to be taken that ICA cannot keep track of sources
when processing several time windows of the EEG because the order of resultant
independent components is, in general, arbitrary. Therefore, artifact removal
requires visual inspection of the components and determination of which compo-
nents to remove. However, the distributions of spectral power and/or scalp topogra-
phies of artifactual components are quite distinct, which suggests that it is feasible to
automate
procedures
for
removing
these
artifacts
from
contaminated
EEG
recordings.
2.4.4 Decomposition of Event-Related EEG Dynamics Based on ICA
It is noteworthy that ICA is not only effective for removing artifacts from EEG data,
but also for direct analysis of distinct EEG components, which arguably represent,
in many cases, functionally independent cortical source activities [45]. During the
last decade, our laboratory and many others have applied ICA to decompose sets of
averaged ERPs, continuous EEG records, and/or sets of event-related EEG data tri-
als and have demonstrated that much valuable information about human brain
dynamics contained in event-related EEG data may be revealed using this method.
In our experience, ICA decomposition is most usefully applied to a large set of con-
catenated single-trial data epochs. Simultaneous analysis of a set of hundreds of sin-
gle-trial EEG epochs gives the concurrently active EEG source processes that
contribute to the response and/or the response baseline a far better chance of
expressing their temporal independence and thus being separately identified by ICA.
ICA algorithms thus can separate the most salient concurrent EEG processes active
within the trial time windows. Many studies (including but not limited to [3, 43,
45-47]) have shown that relatively small numbers of independent components
exhibited robust event-related activities near stimulus presentation and/or the sub-
ject behavioral response. These components tend to have near-dipolar scalp maps,
compatible with a compact cortical source area and suggesting that the brain areas
Search WWH ::
Custom Search