Biomedical Engineering Reference
In-Depth Information
BCI core for control command “decoding”
Signal processing
Pattern classification
“Hard link”
EEG
Acquisition unit
Device control unit
Voluntary intent
encoding
Brain
External devices
Feedback
“Soft link”
Figure 8.1
Components of BCI system.
trol command, such as moving a cursor up or down, is assigned a specific mental
state beforehand. The subject needs to perform the corresponding mental task to
“encode” the desired control command, either through attention shift or by volun-
tary regulation of his EEG [2]. Currently, several types of EEG signals exist—such
as sensorimotor rhythm (SMR; also known as
rhythm) [11-13], steady-state
visual evoked potential (SSVEP) [14, 15], slow cortical potential (SCP) [16, 17], and
P300 [18, 19]—that can be used as neural media in the qEEG-based BCI system.
Among these EEG signals, SMR and SCP can be modulated by the user's voluntary
intent after training, whereas the SSVEP and P300 can be modulated by the user's
attention shift. In fact, the design of the EEG-based BCI paradigm is largely about
how to train or instruct the BCI user to express (“encode”) his or her voluntary
intent efficiently [20]. The more efficient the user's brain encodes voluntary intent,
the stronger the target EEG signal we may have for further decoding.
μ
/
β
8.1.2.2 BCI Core: Control Command “Decoding” with a BCI Algorithm
Feeding the BCI system with a clear input is the function of a biological intelligent
system—the brain, whereas translating input EEG signals into output control com-
mands is the purpose of an artificial intelligent system—the BCI algorithm. Besides
a high-quality EEG recording, appropriate signal processing (SP) and robust pattern
classification are two major parts of a successful BCI system. Because scalp EEGs
are weak and noisy, and the target EEG components are even weaker in a BCI con-
text, various SP methods have been employed to improve the SNR and to extract
meaningful features for classification in BCI [10].
Basically, these methods can be categorized into three domains: time, fre-
quency, and space. In the time domain, for example, ensemble averaging is a widely
used temporal processing technique to enhance the SNR of target EEG components,
as in P300-based BCI. In the frequency domain, Fourier transform and wavelet
analyses are very effective to find target frequency components, as in SMR and
SSVEP-based BCI. In the space domain, spatial filter techniques such as common
Search WWH ::
Custom Search