Techniques of EEG Recording and Preprocessing - Quantitative EEG Analysis Methods and Clinical Applications - page 43

− 1

xWu
k

= (2.4)

k k

where u k is the set of the activation matrix rows for components in set k and W -1 k is
the scalp map matrix columns k . This process is called the back-projection of the
component to the data. Back-projected activity is at the original channel locations
and in the original recording units (e.g., V). Figure 2.7(b) schematically depicts the

projection of the first component onto the scalp channels. This is also easily com-
puted by setting the artifactual and/or irrelevant component activations to zero, as
interpolated and plotted in Figure 2.7(b) (lower scalp maps). In this case, columns
of the inverse unmixing matrix W -1 associated with these components become
nonfactors in the back-projection, whereas the column of the inverse unmixing
matrix associated with the first component determines the amplitude distribution of
the component across scalp channels. For each component, the distribution of cur-
rent across the scalp electrodes is fixed over time, but the actual potential values
(including their polarities) are modulated by the corresponding time course of com-
ponent activation, the relevant row of the output data matrix, in this case, u 1 ( t ),
depicted in the lower panels as the intensity fluctuations of the scalp maps over time.

μ

2.4.3 Artifact Removal Based on ICA
As mentioned earlier, one of the most pervasive problems in EEG analysis and inter-
pretation is the interference in the data produced by often large and distracting arti-
facts arising from eye movements, eye blinks, muscle noise, heart signals, and line
noise. Figure 2.8(a) shows a sample 5-second portion of continuous EEG time series
data collected from 20 scalp electrodes placed according to the International 10-20
system and from two EOG electrode placements, all referred to the left mastoid.
The sampling rate was 256 Hz. In this example, ICA was trained with 10 seconds of
spontaneous EEG data. Figure 2.8(b) shows component activations and scalp

Original EEG Component activations Corrected EEG

(a) (b) (c)
Figure 2.8 Demonstration of EEG artifact removal by ICA. (a) A 5-second portion of an EEG time series
containing a prominent eye movement. (b) Corresponding ICA component activations and scalp maps
of six components accounting for horizontal and vertical eye movements and temporal muscle activity.
(c) EEG signals corrected for artifacts by removing the six selected ICA components in (b).