Fast BCI Calibration
Comparing Methods to Adapt BCI Systems for New Subjects
Jean Thorey
1
, Parvaneh Adibpour
1
, Yohei Tomita
1
, Antoine Gaume
1
,
Hovagim Bakardjian
1,2,3
, Gérard Dreyfus
1
and François-B. Vialatte
1,2
1
SIGMA Laboratory, ESPCI ParisTech, Paris, France
2
LABSP, Riken BSI, Wako-Shi, Japan
3
IHU-A-ICM, Pitié-Salpêtrière Hospital, Paris, France
Keywords: Brain Computer Interface, EEG, SSVEP, Calibration.
Abstract: A Brain Computer Interface (BCI) is a system where a direct connection is established between the brain
and a computer, providing a subject with a new communication channel. Unfortunately, BCI have many
drawbacks: signal recording is problematic, brain signatures are non reproducible from individual to
individual, etc. A dependent-BCI prototype, the BrainPC project, was developed in the SIGMA laboratory.
Electroencephalographic (EEG) signals collected by a BrainAmp amplifier in responses to flickering light
stimuli (Steady State Visual Evoked Potentials) are converted into machine-readable commands. This
system is coupled with a human-machine interface. We propose a solution for fast calibration of the
automatic detection of SSVEP between experimental subjects. We tested different calibration methods;
harmonic and electrode selections were shown to be the most efficient methods.
1 INTRODUCTION
Brain–Computer Interfaces (BCI) are
communication systems that enable users to send
commands to a computer by using only their brain
activity (Nicolelis, 2011). This activity is generally
being measured through EEG, which is a
noninvasive technique for recording brain electrical
activity at the surface of the scalp. In a BCI, the
brain signals are recorded and analyzed to extract
features that represent the messages buried inside the
EEGs. Then a translation algorithm is needed to
convert the features to a command which is
supposed to be sent to the computer or external
machine. It is through this procedure that disabled
people can control a computer screen or navigate a
wheelchair (Wolpaw et al., 2002). SSVEP-based
BCIs are those BCIs that allow the users to
communicate with a computer or machine, by
SSVEP responses that are generated in their brain by
looking at a repetitive visual stimulus. Steady State
Visual Evoked Potential (SSVEP) is an oscillatory
activity in human visual cortex that is phase locked
to repetitive visual stimulation (Vialatte et al., 2010).
Studies on developing SSVEP-based BCIs have
used several algorithms for detecting the SSVEPs.
Most of the studies in literature identify user’s
intended target by calculating the frequency
spectrum analysis of the signal and this is typically
implemented using the Fourier Transform,
particularly Fast Fourier Transform (FFT). Detection
is usually based on considering a threshold for the
power spectrum at stimuli frequencies (Vialatte et
al., 2010). There are also other recent studies,
introducing new methods for detection of SSVEPs.
(Friman et al., 2007, Lin et al., 2006, Bin et al.,
2009, Zhang et al., 2012)
For a BCI to have applicability in daily life, it is
very important to make it work in different
situations and for different subjects. This is not
always easy due to the subject variability in the
spatial patterns and spectrotemporal characteristics
of brain signals (Volosyak et al., 2010). This subject
variability makes the pattern recognition part quite
difficult. For solving this issue, a rather long
calibration phase is usually added to BCI
experiments in order to collect EEG data from the
subject to train the classifier or to adapt the stimuli
parameters for each subject. The main problem with
the calibration phase is the long times it takes for
663
Thorey J., Adibpour P., Tomita Y., Gaume A., Bakardjian H., Dreyfus G. and Vialatte F..
Fast BCI Calibration - Comparing Methods to Adapt BCI Systems for New Subjects.
DOI: 10.5220/0004180806630669
In Proceedings of the 4th International Joint Conference on Computational Intelligence (SSCN-2012), pages 663-669
ISBN: 978-989-8565-33-4
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
recoding the EEGs. For several applications (e.g.
video games or neurorehabilitation), a fast
calibration is necessary. Some previous studies have
proposed methods for reducing the calibration time
but most of them require a database of recordings
from different subjects or several past recordings
from the same BCI user. Krauledat et al. (2008)
showed in their study on motor imagery BCI users,
how predefined spatial filters and classifiers on the
recorded data of previous training sessions of the
same user would eliminate the need for a whole
calibration phase at the beginning of each online
experiment. To do this, they adjusted the bias of the
classifier at the beginning of the online experiment.
However, their good classification results were
showing the power of their method for session to
session transfer for the same subject but not for
inter-subject variability. Lotte (2011) proposed a
method based on generation of artificial EEG trials
from the few previous collected trials in order to
increase the training data set of classifier.
Generating artificial EEG trials was based on
segmentation of the data from different trials and
then concatenating the segments from different trials
to make new trials. Shishkin et al. (2011) proposed
to use single stimulus for the calibration phase in a
P300-based BCI in order to avoid the conflicts of
non-target stimuli. The performances of their BCI
system did not deteriorate significantly even when
trained using a single-stimulus protocol. Wang et al.
(2006) showed that user variability could be reduced
by adapting channel, stimulus frequency and speed
of command detection for each subject. Volosyak et
al. (2010) compared two calibration methods of
single LED and multi-target group LED stimuli for
exploring the best stimulation frequencies. They
found a strong correlation between the selected
stimulation frequencies through both methods. They
concluded they could shorten the calibration time
significantly by using the multi-target group LED
stimuli for detecting the best stimuli frequencies.
We investigate here methods for fast calibration
of an SSVEP BCI based on selection of the channels
and dominant frequency between the first and
second harmonics of stimulation frequency
independently for each subject. Such a method
would allow us to use the system directly on new
subjects, without long calibration times, but
nevertheless exploiting previously collected data to
design an optimal classifier.
2 METHODS
2.1 Experimental Paradigm
A virtual phone keypad was used as the interface,
with 9 digits displayed, each of them were flickering
with a pre-decied frequency (5.45 , 20, 8.75, 4.62,
6.67, 7.5, 12, 5 and 4 Hz). The background colour
was black and the screen refresh rate was 60 Hz.
This display was was realised using Cogent
Graphics developed by John Romaya at the LON at
the Wellcome Department of Imaging Neuroscience.
Figure 1: SSVEP stimulation interface.
2.2 Data Acquisition
EEGs were recorded with a BrainAmp amplifier;
with a sampling rate of the 500 Hz. 16 active
electrodes were placed over the head, according to
the 10-20 international system for electrode
placement. The electrodes covered the frontal,
temporal and occipital sites (Fp
1
, Fp
2
, F
3
, F
4
, F
7
, F
8
,
C
3
, C
4
, T
3
, T
4
, T
5
, T
6
, Po
1
, Po
2
, O
1
, O
2
), with the
reference and grounds placed in the central positions
(F
z
and P
z
).The subjects were told to relax and focus
during 30 seconds on each command consecutively.
30 seconds of resting state eyes open were also
recorded in front of a black screen at the beginning
of each recording session.
We recorded 7 subjects. All subjects were young
adults without any known history or actual brain
disorder or anomaly.
2.3 Feature Extraction
The overall workflow of signal processing is as
follows: (1) supervised feature extraction, for all
subjects; (2) calibration, for each subject
independently; (3) command classification and
performance evaluation. After having selected a set
of seven relevant features for the BCI system, we
use the database collected to evaluate the
performance of the BCI on new unknown subjects
(classification is detailed in section 4). We compared
the results obtained on the raw data, with results
obtained after calibration of the data (calibration is
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
664
explained in section 3). When applying calibration,
it is applied for each subject: the “unknown” test
subject as well as those used as a training reference
for the classifier. All the selected features are pre-
processed, for each subject, using these calibration
techniques. We will now explain the supervised
feature selection approach used.
Most BCIs are designed around a pattern
recognition approach. In an EEG-based BCI, the
first step is to extract features describing the relevant
information buried in the EEG signals. They are then
fed into a classifier which identifies the class which
these features belong to. For detection of SSVEPs,
we extracted the following features from the signals:
Fourier Peak: For detecting the SSVEPs in the
signal, one first need to transform the time domain
EEG into the frequency domain using Fourier
transform. Once it is transformed into frequency
domain, peaks at stimulation frequency and its
harmonics are detectable. For detecting these peaks,
we took the maximum amplitude in the Fourier
spectrum of the signal at a small margin around each
stimulation frequency.
Signal to Noise Ratio Peak: Signal to Noise
Ratio (SNR) is a measure that depends on the
frequency f and is computed as the ratio of Fourier
Power at frequency f and average Fourier power at
its adjacent frequencies. This is actually a way to
enhance SSVEP peaks (Wang et al. 2006) and is
computed according to the following formula:
X
f
∑
∆
/
∑
∆
/
,
(1)
where X(f) is the value for Fourier power of a signal
at the frequency f and X’(f) is the value of the SNR
at frequency f, and Δf is the frequency step. The
maximum SNR value at a small margin around each
stimulation frequency is then defined as the SNR
Peak.
We computed Fourier Peak and the SNR Peak
for occipital, parieto-occipital and Frontal channels.
(O
1
, O
2
, Po
3
, Po
4
, F
3
, F
4
)
Magnitude Squared Coherence: Magnitude
Squared Coherence (MSC) is a measure for
quantifying the synchronization between two
signals. This feature is computed between pairs of
EEG channels to see how similar their power
spectrums in terms of magnitude are. The magnitude
squared coherence is a function of the power
spectral densities (Pxx(f) and Pyy(f))and the cross
power spectral density (Pxy(f)) of x and y.
,
(2)
We computed this value for the following
channel pairs: O
1
- O
2
, F
3
-F
4
, Fp
1
-O
1
, Fp
2
-O
2
, Fp
1
-O
2
,
Fp
2
-O
1
.
Fourier and SNR Peak for concatenated Signals:
The FFT epochs of SSVEP signal require sufficient
data length to achieve a satisfactory frequency
resolution. However, increased epoch length comes
at the cost of time taken to collect EEG. Since for
the purpose of online BCI applications, time for
processing the data and estimation of the command
is a crucial element to be kept short, detection
should be done using short epochs of signals. Tomita
et al. (2011) proposed concatenation method to
improve the frequency resolution of the SSVEPs
using short time window epochs. In their proposed
method, they concatenated signals from different
channels in the time domain and showed that the
concatenated signal produces clearer SSVEP peaks
in the Fourier Spectrum due to the increased
frequency resolution.
For this study, two groups of concatenated
signals were built, one including two frontal
channels (F
3
and F
4
) and the other one including the
parieto-occipital and occipital channels (O
1
, O
2
, Po
3
and Po
4
). Then the Fourier and SNR Peak were
computed for both concatenated signals.
2.4 Feature Ranking
Since the number of candidate features was too large
(Nf = 22) given the number of examples in the
database (Ne = 100) for each stimulation frequency,
feature selection was performed by the orthogonal
forward regression (OFR) algorithm (Guyon and
Elisseef, 2003) to select the most relevant features
for discriminating the 9 Stimuli. For this purpose we
used OFR algorithm 36 times, each time finding the
most relevant features for discrimination of two
different Stimuli (36 different combinations for 9
different stimuli frequencies). OFR algorithm
calculates the angle between each candidate feature
u
i
and the quantity to be modelled y and defines the
most correlated feature as the feature that has the
smallest angle (i) with y.
Θ
,
(3)
Then y and all the remaining candidate features
are projected onto the null space of the selected
feature and the same procedure is iterated until all
candidate features are ranked. We performed our
feature ranking on half of our database in order to
avoid a bias. Finally 8 features were selected (see
Table).
FastBCICalibration-ComparingMethodstoAdaptBCISystemsforNewSubjects
665
Table 1: Top-ranked features from the feature selection
step.
Feature Channels
Frequency peak Average (O
1
,O
2
)
Frequency peak Average (F
3
,F
4
)
SNR peak Average (Po
3
,Po
4
)
Magnitude squared coherence O
1
, O
2
Concatenation: frequency peak F
3
, F
4
Concatenation: SNR peak F
3
, F
4
Concatenation: frequency peak O
1
, O
2
,Po
3
, Po
4
Concatenation: SNR peak O
1
, O
2
,Po
3
, Po
4
3 CALIBRATION
3.1 Distribution Calibration
We hope to reduce the inter-subject variability by
calibrating the data, so that the SSVEP responses
would be more homogeneous. ‘First level’
calibration is based on a mathematical projection of
the data into a reference space, which is defined
based on a short period of time. We investigated two
different first level calibration approaches:
- Resting state eyes open data. In this case, we
remove for each feature the mean value of 30
sec of resting state.
- Resting state eyes open data and active state
data (see Figure 2): active state data is a
collection of 30 sec of SSVEP response at a
given frequency. In this case, we remove the
separating threshold between active and
passive data, so that active data and non-active
data will be discriminated on their sign. The
threshold is determined using linear
classification (active vs. rest data). Active data
values should then be positive and non-active
data values should be negative.
Figure 2: Frequency #2 is defined as active state frequency
while the other pads are defined as non-active.
3.2 Feature Calibration
3.2.1 Harmonic Selection
Depending on the subject and the observed
frequency of stimulation f, features can have higher
value depending on if they are calculated at the
fundamental f or at the harmonic 2f. While the mean
value is commonly used, a calibration can be
performed to detect the frequency and feature that
emphasize this specificity for each subject. Based on
a 9 30-second recordings for each stimulation
frequency, a selection between the value of the
feature at f, 2f and the mean value is processed via
the Mann-Whitney test. This test determines whether
the medians are significantly different. A case where
the Mann-Whitney z-score is above 2 or -2 indicates
that either the fundamental or the harmonic
dominates. Otherwise the medians are not different
enough, and then the mean value is kept.
3.2.2 Channel Topography Selection
Eight classic features were isolated as explained
above, but these 8 features are not optimal for all the
subjects. We optimized the channel selection, by
subdividing the selected features into 7 groups of
features. The 7 groups are organized in order to
access a subject-specific topography (see Table).
Table 2: 7 Groups of expanded features These features are
more detailed topographic mappings of the selected
features, so that the topography can be further adapted to
each subject.
Group1 Concatenation: frequency peak
O
1
O
2
Po
3
O
1
O
2
Po
4
O
1
Po
1
Po
4
O
2
Po
1
Po
4
O
1
O
2
Po
3
Po
4
Group2 Concatenation: SNR peak
O
1
O
2
Po
3
O
1
O
2
Po
4
O
1
Po
3
Po
4
O
2
Po
3
Po
4
O
1
O
2
Po
3
Po
4
Group3 Magnitude squared coherence 1
O
1
O
2
Po
3
Po
4
O
2
Po
3
O
1
Po
4
Group4 SNR peak
F
3
F
4
F
3
F
4
Group5 SNR peak
O
1
O
2
Po
3
Po
4
Po
3
O
2
Po
4
O
1
O
1
O
2
Po
4
Po
3
O
1
O
2
Po
3
Po
4
Group6 Magnitude squared coherence 2
F
3
O
2
F
3
Po
4
F
4
O
1
F
4
Po
3
Group7 Frequency peak
O
1
O
2
O
1
O
2
Based on one 30-second recording at only one
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
666
stimulation frequency, considered as a reference,
one feature of each group is selected via OFR for
each subject (therefore personalizing the channel
topography for each subject). Each subject has then
his personalized7-feature set.
The selection processed is OFR based:
- find the best feature in the 33 features.
- remove all features corresponding to the same
group
- project the remaining candidate features (from
the other groups) onto the null space of the selected
feature.
The above two steps can be iterated in subspaces
of decreasing dimensions until one candidate of each
group has been selected.
4 CLASSIFICATION
The classes correspond to the frequencies of
stimulation, they indicate which of the buttons on
the dial pad the subject wishes to activate. The 9-
class classifier is in fact composed of 36 2-class
linear classifiers (LDA). Each 2-class classifier is
based on twice the features account, as the class
parameters are the features calculated at both
frequencies of analysis.
For a given 3s signal, the features are extracted
for the 9 stimulation frequencies. Then for each
couple (f
a,
f
b
) of frequencies, the corresponding
features are compared to determine whether the
signal corresponds to class #a or class #b, in other
words if the observed command flickered at
frequency f
a
or at f
b
. The command estimation is
based on the best mean score after the 36
comparisons.
5 RESULTS
5.1 Distribution Calibration
We classify the data using a cross validation
approach. As we intend to test the capability of the
system to adapt to new subjects, we iteratively
remove one subject, train the classifier with the other
data, and test on the rejected subject. This method is
similar, in spirit, with the classical leave-one-out
cross-validation approach – but here we leave one
subject out, instead of only one example.
The success rate (SR) is defined as the trace of
the confusion matrix after Leave-One-Subject-Out
testing. Generally, SR is not significantly improved
Figure 3: SR rate, sorted from the worst to the best
subject. Black: no calibration (mean = 0.56); black dotted:
passive distribution calibration (mean = 0.56); magenta:
active distribution calibration (reference frequency
6.67Hz, mean = 0.56).
by the distribution calibration, except for the
calibration on rest data which slightly improved the
worst subject (at the expense of decreasing the SR
for the best subject). These classification results are
stable across frequencies, as is illustrated on the
confusion matrix of Figure 4.
Figure 4: Confusion Matrix (average of subjects).
horizontal axis: stimulation frequency. Vertical axis:
estimated command.
This confusion matrix corresponds to classification
of the data without calibration; the (not shown)
confusion matrices for all type of calibrations
investigated in this manuscript share the same
stability properties across frequencies.
5.2 Feature Calibration
Calibration based on the selection of the dominant
harmonic led to a general improvement of the SR,
where six of the seven subjects had SR above 0.55
(Figure).
FastBCICalibration-ComparingMethodstoAdaptBCISystemsforNewSubjects
667
Figure 5: SR rate, sorted from the worst to the best
subject. black: harmonic selection (mean = 0.58); black
dotted: harmonic selection combined with passive
distribution calibration (mean = 0.53); magenta: harmonic
selection combined with active distribution calibration
(reference frequency 6.67Hz, mean = 0.54).
Topographic selection calibration also led to a
significant improvement (average SR = 0.58 for
reference frequency 8.75 Hz), but this improvement
was not stable across frequencies.
Figure 6: SR rate, sorted from the worst to the best
subject. black: topographic selection, using a 4 Hz
reference signal (mean = 0.60); black dotted: topographic
selection combined with passive distribution calibration
(mean = 0.57); magenta: topographic selection combined
with active distribution calibration (mean = 0.59).
Combining harmonic selection and topographic
selection led to an improvement, which was this
time much more stable across reference frequencies
(SR = 0.57 for 10 Hz, SR = 0.58 for 6.67 Hz, and SR
= 0.60 for 4 Hz).
Whether using harmonic selection, topographic
selection, or both, subsequent distribution calibration
did not provide any improvement.
6 DISCUSSION
We investigated four different types of BCI system
calibration, based on:
- Distribution mapping, using a rest condition
signal as reference,
- Distribution mapping, using a rest condition
signal and an active signal as references,
- Subject dependent choice of electrodes,
- Subject dependent choice of SSVEP
harmonics.
We compared the classification results for the
detection of SSVEP peaks of these four calibration
methods.
For the choice of stimulation frequencies, we
embedded 20 Hz stimulation in the design of our
stimulation interface. The SSVEP responses that
were generated with this frequency were strong
enough to be detected. However, in a study by
Bakardjian et al. (2010), the best choice of
stimulation frequency for evoking the strongest
response is reported to be among 5.6 to 15.3 Hz. On
the other hand, there exist studies supporting the
usefulness of high frequency stimuli in generating
good SSVEP responses. (Wang et al., 2006,
Volosyak et al., 2010). Wang et al. (2006) also
employed high frequency stimuli in their
experimental design, and explained this increase in
the stimulus frequency bandwidth not only as a
factor to decrease time length of signal epochs for
detection of SSVEPs but also as a factor for
reducing the eyestrain effect caused by the flickers.
However, these effects may vary from subject-to-
subject, and we did not investigate further in this
direction.
We consistently show that distribution mapping
proved to be useless. It never improved the
classification rates, whether used alone or in
combination with the other two calibration methods.
The features values of resting state data don’t seem
to be comparable to the features value of non-active
state frequency. Depending on the subject, the
frequency and the features, they can be lower or
higher. Thus no way was found yet to find a
generalization rule for all subjects. This result could
be due to several reasons. First of all, 30 seconds of
data may be insufficient to extract a sufficiently
stable signature of the EEG activity. Using longer
epochs might provide better results. Second, owing
to the non-stationary nature of EEG (see e.g. Kaplan
et al., 2005), it might be necessary to monitor the
signal evolution along time (the data collection
lasted up to one hour), otherwise the reference data
used for calibration may not be a good reference
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
668
anymore.
Selection of harmonics and topography led to
much more clear improvements. This is to be
expected: this method seeks to adapt the system to
the specificities of each subject. It is noteworthy that
each subject has specific brain responses to SSVEP
(see e.g. Silberstein, et al., 1990), whether
topographically or frequency-wise. It is therefore not
surprising that an adaptation of the system to the
specificities of each subject leads to an improved
classification. The best calibration method between
those two, according to our results, is the selection
of the dominant harmonic in the SSVEP response.
However, the reader should keep in mind that those
two methods are based on very different approaches.
Harmonic selection used 5 minutes of data, whereas
topography selection used only 1 minute of data, but
still led to some significant improvements. Our
results therefore also confirm the interest of
selecting the channels, which was already pointed by
Wang et al. (2006).
ACKNOWLEDGEMENTS
In this project Hovagim Bakardjian was supported
by an International Neuroinformatics Coordinating
Facility grant (june 2011), and by the #3862
fellowship of the Fondation Pierre Gilles de Gennes.
REFERENCES
Bakardjian, H., Tanaka, T., Cichocki, A. 2010,
Optimization of SSVEP brain responses with
application to eight-command Brain–Computer
Interface. Neuroscience Letters, 469(1):34-38.
Bin, G., Gao, X., Yan, Z., Hong, B., Gao, S., 2009. An
online multi-channel SSVEP based brain–computer
interface using a canonical correlation analysis
method. J. Neural Eng. 6 (4) 6: 046002
Friman, O., Volosyak, I., Graser, A., 2007. Multiple
channel detection of steady-state visual evoked
potentials for brain–computer interfaces. IEEE Trans.
Biomed. Eng. 54 (4), 742–750.
Kaplan A. Y., Fingelkurts A. A., Fingelkurts A. A.,
Borisov S. V., Darkhovsky B. S., 2005, Nonstationary
nature of the brain activity as revealed by EEG/MEG:
Methodological, practical and conceptual challenges.
Signal Processing, 85(11):2190-2212.
Krauledat, M., Tangermann, M., Blankertz, B., Muller, K.
R., 2008. Towards zero training for brain–computer
interfacing. PLoS ONE 3(8), e2967.
Lin, Z., Zhang, C., Wu, W., Gao, X., 2006. Frequency
Recognition Based on Canonical Correlation Analysis
for SSVEP-Based BCIs. IEEE Trans. Biomed Eng.
532610-4.
Lotte, F., 2011. Generating Artificial EEG Signals To
Reduce BCI Calibration Time, 5th International
Brain-Computer Interface Workshop
Nicolelis M., 2011, Beyond Boundaries: The New
Neuroscience of Connecting Brains With Machines-
And How It Will Change Our Lives. Times books.
Parini, S., Maggi, L., Turconi, A. C., and Andreoni, G.,
2009. A Robust and SelfPaced BCI System Based on a
Four Class SSVEP Paradigm: Algorithms and
Protocols for a High-Transfer-Rate Direct Brain
Communication. Comput.Intelli. Neurosci ID 864564.
Silberstein, R. B., Schier, M. A., Pipingas, A., Ciorciari,
J., Wood, S. R., Simpson, D. G., 1990. Steady-state
Visually evoked potential topography associated with
a visual vigilance task. Brain Topogr. 3, 337–347.
Shishkin S. L., Nikolaev A. A., Nuzhdin Y. O., Zhigalov
A. Y., Ganin I. P., Kaplan A., Y. 2011. Calibration of
the P300 BCI with the single-stimulus protocol. Proc.
of the Fifth International BCI Conference Graz
University of Technology, Austria. 256-259
Tomita, Y., Gaume, A., Bakardijian, H., Maurice, M.,
Cichocki, A., Yamaguchi, Y., Drefus, G., Vialatte, F.,
2011. Concatenation Method for high temporal
resolution SSVEP-BCI. International Conference on
Neural Computation Theory and Applications.
Vialatte, F. B., Maurice, M., Dauwels, J., Cichocki, A.,
2010. Steady-state visually evoked potentials: Focus
on essential paradigms and future perspectives,
Progress in Neurobiology 90 418–438
Volosyak, I., Malechka, T., Valbuenaand, D., Graser, A.,
2010. A Novel Calibration Method for SSVEPBased
Brain Computer Interfaces.
18th European Signal
Processing Conference, 939-943.
Wang, Y., Wang, R., Gao, X., Hong, B., Gao, S., 2006. A
practical VEP-based brain–computer interface. IEEE
Trans. Neural Syst. Rehabil. Eng. 14 (2), 234–239.
Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,
Pfurtscheller, G., Vaughan, T. M., 2002. Brain–
computer interfaces for communication and control.
Clin. Neurophysiol. 113 (6), 767–791.
Zhang, Y., Jin, J., Qing, X., Wang, B., Wang, X., 2012.
LASSO based stimulus frequency recognition model
for SSVEP BCIs, Biomedical Signal Processing and
Control (7)104– 111
FastBCICalibration-ComparingMethodstoAdaptBCISystemsforNewSubjects
669
