Biomedical Engineering Reference
In-Depth Information
In multichannel recordings, the dimension of the information is commonly
not as high as the number of channels. Hence, independent component analysis
(ICA) is usually introduced to identify underlying signal sources (Sect.
3.2.1
).
As a result, ICA can lead to artifacts extraction and data dimension reduction,
which are benefi cial for the subsequent calculation of activity features. Two
types of activity features are widely used to quantify event-related spatio-spec-
tro-temporal brain dynamics: time-frequency representations (TFRs, Sect.
3.2.2
)
and connectivity measures (Sect.
3.2.3
). The two features offer different infor-
mation outlines that complement each other. The last step in offl ine analysis is
the identifi cation of statistically signifi cant features underlying the target neural
processes (Sect.
3.2.4
).
3.2.1
Independent Component Analysis
ICA is a technique that is used to separate multichannel signals into their constituent
underlying sources, such that each source contributes as much distinct/independent
information to the data as possible. Adding this step to ECoG analysis may provide
lower-dimensional, nonartifactual, and statistically independent source signals.
Consequently, ICA can reduce the computational load and prevent colinearity in the
data, which are essential for calculating more complex activity features, such as
connectivity measures (Sect.
3.2.3
).
Preprocessing Before ICA
Data preprocessing is the most overlooked step and could diminish the effectiveness
of ICA or even lead to spurious results if not performed carefully. The most crucial
step in the preprocessing step before performing ICA is to reject artifacts in the raw
signals. For multitrial data, such as the data from the fear recognition experiment, a
thorough strategy for artifact rejection can be found in EEGLAB wiki (Delorme and
Makeig
2004
). The basic concept is to reject trials or channels that appear to contain
artifacts using visual inspection, statistical thresholding, or a combination of both.
To analyze the fear recognition data, we rejected channels and trials with abnormal
spectra, which has been suggested as the most effective method (Delorme et al.
2001
). An alternative strategy is to select nonartifactual independent components
(ICs) after performing ICA on the raw data.
Model Order and Component Selections
When the number of channels exceeds the number of real sources, the model order,
i.e., the number of ICs to estimate, can have a signifi cant impact on the quality and
accuracy of ICA. If the model order is greater than the actual number of sources,
