Brain–Computer Interfaces - Index of

Digital Signal Processing and Machine Learning 317
Initially, many features are extracted from various methods and from channels covering
all potentially useful brain areas. Then, feature selection is employed to find
the most suitable channel locations and feature extraction methods. This section
describes two aspects of feature selection for BCIs: channel selection and frequency
band selection.
4.1 Channel Selection
As mentioned before, not all electrodes distributed over the whole scalp are useful
in BCI systems in general. Hence, channel selection is performed to select the useful
channels according to the features used in a BCI system. For SSVEP-based BCIs,
the EEG signals from the visual cortex are selected. For mu/beta-based BCIs, the
channels in the sensorimotor area are selected. For P300-based BCIs, the channels
exhibiting an obvious P300 are selected. For P300 or SSVEP-based BCIs, channel
selection is often manually performed before other processing steps. Apart from
physiological considerations, channel selection can be performed by applying a
feature selection method to a training dataset. Examples of such a feature selection
method include optimizing statistical measures like Student’s t-statistics, Fisher
criterion [33] and bi-serial correlation coefficient [34] and genetic algorithms [35].
In the following, channel selection based on the bi-serial correlation coefficient
is described [34]. Let (x1, y1), ··· ,(xK, yK) be a sequence of one-dimensional observations
(i.e. a single feature) with labels yk ∈ {1, −1}. Define X + as the set of
observations xk with label of 1, and X − the set of observations xk with label of
−1. Then bi-serial correlation coefficient r is calculated as
rX =
√ N + N −
N + + N −
mean(X− ) − mean(X + )
std(X + � X− , (11)
)
where N + and N− are the numbers of observations of X + and X− respectively. r2- coefficient r2 X , which reflects how much of the variance in the distribution of all
samples is explained by the class affiliation, is defined as a score for this feature.
For each channel of a BCI system, a score is calculated as above using the observations
from this channel. If the objective is to choose the N most informative
channels, one would choose the channels with the top N scores.
4.2 Frequency Band Selection
As mentioned in Sect. 2.2, temporal filtering plays an important role in improving
the signal-to-noise ratio in BCI systems [7], and temporal filtering is also important
to extract band power features. Before temporal filtering, one or several frequency
bands need to be determined that affect the performance of BCIs. The optimal
frequency bands are generally subject-specific in BCIs. Frequency band selection
is often performed using a feature selection method. As an example, we use the
bi-serial correlation coefficient presented in Sect. 4.1 to carry out frequency band