Biomedical Engineering Reference
In-Depth Information
user's occipital region on the scalp, typically on two sites around Oz in the 10-20
EEG electrode system. A tennis headband was modified to harness the electrodes on
the head surface.
The EEG signal was amplified by a customized amplifier and digitized at a sam-
pling rate of 256 Hz. After a 50-Hz notch filtering to remove the power line interfer-
ence, the digital EEG data were streamed to PC memory buffer through a USB port.
For the precision of frequency control, the periodical flickering of each LED was
controlled by a separate lighting module, which downloads the frequency setting
from the PC through the USB port. In one of the demonstrations, our BCI system
was used for dialing a phone number. In that case, a local telephone line was con-
nected to the RJ11 port of an internal modem on the PC.
8.2.2.2 BCI Software and Algorithm
The main software running on the PC consists of key parts of the EEG translation
algorithm, including signal enhancing, feature extraction, and pattern classification.
The following algorithms were implemented in C/C
and compiled into a
stand-alone program. The real-time EEG data streaming was achieved by using a
customized dynamic link library.
In the paradigm of SSVEP, the target LED evokes a peak in the amplitude spec-
trum at its flickering frequency. After a band filtering of 4 to 35 Hz, the FFT was
applied on the ongoing EEG data segments to obtain the running power spectrum. If
a peak value was detected over the frequency band of 4 to 35 Hz, the frequency cor-
responding to the peak was selected as the candidate of target frequency. To avoid a
high false-positive rate, a crucial step was taken to ensure that the amplitude of a
given candidate's frequency was higher than the mean power of the whole band.
Herein, the ratio between the peak power and the mean power was defined as
++
QP P
peak
=
(8.1)
mean
Basically, if the power ratio
Q
was higher than the predefined threshold
T,
then
the peak power was considered to be significant. For each individual, the threshold
T
was estimated beforehand in the parameter customization phase. The optimal
selection of the threshold balanced the speed and accuracy of the BCI system.
Detailed explanation of this power spectrum threshold method can be found in pre-
vious studies [30, 31].
Due to the nonlinearity that occurs during information transfer in the visual sys-
tem, strong harmonics are often found in the SSVEPs. Muller-Putz et al. investigated
the impact of using SSVEP harmonics on the classification result of a four-class
SSVEP-based BCI [32]. In their study, the accuracy obtained with combined har-
monics (up to the third harmonic) was significantly higher than that obtained with
only the first harmonic. In our experience, for some subjects, the intensity of the sec-
ond harmonic may sometimes be even stronger than that of the fundamental compo-
nent. Thus, analysis of the frequency band should cover at least the second
harmonic, and the frequency feature has to be taken as the weighted sum of their
powers, namely,
()
() (
) ()
Pi
=
α
P
i
+
1
−
α
P
i
i
=
1,
,
N
(8.2)
f
1
f
2
Search WWH ::
Custom Search