Conference PaperPDF Available

Improving SSVEP-BCI performance using pre-trial normalization methods

Conference Paper

Improving SSVEP-BCI performance using pre-trial normalization methods

Abstract and Figures

A brain-computer interface (BCI) enables users to communicate through a computer using only their brain signals, by extracting brain signal features containing information representative of the user's intent, and can be used in a wide variety of areas such as entertainment, rehabilitation, or assistive technologies. In this paper we assessed two normalization methods which are aimed at improving the quality of the extracted features: Baseline-Corrected CCA (BC-CCA), and Scaled CCA. Both methods were found to be able to improve classification accuracy in conditions using frequencies with a large range, whilst BC-CCA was found to be the superior of the two, improving SSVEP detection accuracy by as much as 9.22%.
Content may be subject to copyright.
Improving SSVEP-BCI Performance Using Pre-Trial
Normalization Methods
James Henshaw
Department of Electronic and
Electrical Engineering
University of Sheffield
Sheffield, England
Email: jhenshaw1@sheffield.ac.uk
Wei Liu
Department of Electronic and
Electrical Engineering
University of Sheffield
Sheffield, England
Email: w.liu@sheffield.ac.uk
Daniela M. Romano
Sheffield Robotics
University of Sheffield
Sheffield, England
Email: d.m.romano@sheffield.ac.uk
Abstract—A brain-computer interface (BCI) enables users
to communicate through a computer using only their brain
signals, by extracting brain signal features containing information
representative of the user’s intent, and can be used in a wide
variety of areas such as entertainment, rehabilitation, or assistive
technologies. In this paper we assessed two normalization meth-
ods which are aimed at improving the quality of the extracted
features: Baseline-Corrected CCA (BC-CCA), and Scaled CCA.
Both methods were found to be able to improve classification
accuracy in conditions using frequencies with a large range, whilst
BC-CCA was found to be the superior of the two, improving
SSVEP detection accuracy by as much as 9.22%.
Keywordsbrain-computer interface, BCI, SSVEP, normaliza-
tion, EEG.
I. INTRODUCTION
A brain-computer interface (BCI) is operated using brain-
generated information, thereby providing the user with an alter-
native communication or control channel that does not require
the brain’s normal peripheral nerve and muscle pathways [1].
Operating a BCI is a multi-step process [2], which involves
acquiring a signal from the user’s brain, preprocessing it to
reduce artifacts, extracting useful features that can be used
to make inferences about the user’s current cognitive state,
and then using a method for translating these features into
a communication channel that can control or communicate
with an external device, and finally, providing feedback so
the user can see that the BCI is working (Fig. 1) and adjust
their approach if necessary. There are several methods for
sending control signals to operate a BCI [3], and one of the
most popular methods is steady-state visually evoked potentials
(SSVEPs [4]). SSVEPs are a phase-locked brain response
triggered by fixing the user’s gaze upon a repetitive visual
stimulus (RVS) such as a flashing light [5], or a reversing
pattern [6]. The SSVEP is produced by groups of neurons
which output a repetitive signal matching the RVS frequency,
maintained for the duration of the fixation period. Correctly
identifying the user’s SSVEP frequency allows a correct com-
mand to be sent to the BCI. The SSVEP can be detected
by recording brain activity from electrodes placed around the
occipital and parietal lobes [7]. SSVEP-BCIs are popular due
to their short training time, high classification rate [8], and the
fact that they can be detected using non-invasive neuroimaging
methods such as electroencephalography (EEG). They have
been used in a diverse range of BCI types, including BCI-
controlled exoskeletons [9], [10], wheelchairs [11], [12], and
robotic humanoids [13]. As with all other signal production
methods, the overall goal is to maximize the signal-to-noise
ratio (SNR) using various methods.
Currently, two of the most popular methods for SSVEP
feature extraction are: canonical correlation analysis (CCA,
[14], [15], [16], [17], [18]), which calculates the correlation
between the user’s EEG signal and the target frequencies, and
power spectral density analysis (PSDA, [19], [20], [21]), which
uses frequency components of the EEG signal as features for
classification. CCA has a high accuracy and does not need
training data. A notable characteristic of EEG is that the
EEG frequency components from low frequencies tend to have
higher power than those from high frequencies, making them
easier to detect, and leaving a signal naturally skewed in favour
of low frequency RVS. One way to minimize this bias is to
only use stimulus frequencies from the same range. However,
it would be preferable to adjust our feature extraction methods
in a way that gives a balanced result. Having access to more
stimulus frequencies means more unique commands can be
sent and thus with a higher information transfer rate. It is
noted in [22] that uneven distribution between classification
accuracy of classes leads to a skewed performance - the ideal
BCI will have an equal chance of selecting any command.
This skewness can be alleviated by normalization, also known
as feature scaling, which standardizes features based on some
relationship within or between groups of features for reducing
the impact of extreme values and/or the difference between
features of different classes. In this paper, we compare a
variety of methods aimed at improving the BCI’s ability to
send information by improving the quality of the features
produced during the feature extraction process. There have
been numerous papers exploring different aspects of the BCI
process. However, to the authors’ knowledge there is currently
very little research into SSVEP normalization methods.
II. PREVIOUS WO RK
A number of studies have used methods of normalizing
EEG signals for SSVEP detection. Nakashini and colleagues
[16] took CCA features from their target frequencies and
normalized them against CCA features from neighbouring fre-
quency bands, to help compensate for poorer classification with
higher frequency RVS. They found that these features could
Fig. 1. BCI block diagram
perform as well as (and sometimes outperform) the standard
CCA, and performance improved as the number of neighbours
increased. Castillo et al. [19] applied a similar method of
normalizing features against neighbouring frequencies using
PSDA, where they would normalize against a single value to
find the largest ratio. This led to a more accurate BCI and
had less variance than the PSDA. Despite a relatively low
SNR of the high-frequency visual input, Sakurada et al. [23]
created a high frequency SSVEP-BCI with good three-class
classification accuracy, normalizing all the RVS frequencies
against the inter-trial average of spectral power across the
fixation period, and also against competing frequencies. In
effect each normalized SSVEP amplitude was the baseline
corrected amplitude with the mean amplitude of the (baseline-
corrected) competitors subtracted from it. Diez et al. [24]
had participants operate a BCI-controlled navigation robot
using SSVEP features from high-frequency (f > 35 Hz)
RVS. These features were normalized against the periodogram
of baseline data collected prior to the study. There was no
direct comparison with other normalization methods as this
was a navigation study. However, all participants were able
to successfully operate the BCI using the baseline-corrected
features.
The previous literature illustrates that there are a variety
of different ways to improve SSVEP performance using nor-
malization techniques. However, the majority of studies focus
on PSDA-based techniques, whereas the current state-of-the-
art SSVEP-BCI algorithms use CCA and CCA-based methods.
Previous research has indicated that it is possible to improve
CCA performance using normalization methods, and also that
data from the pre-fixation period can be used to normalize
the SSVEP response across frequencies, albeit with PSDA.
Therefore, we hypothesize SSVEP-BCI performance can be
improved by using CCA data from the pre-fixation period.
III. METHODOLOGY
A. Participants
Participants were 17 students recruited using the university
mailing system (4 female, 13 male) with a mean age of 26.5
years old.
B. SSVEP Stimulus
An RVS was created and displayed on a separate computer,
using code written on MATLAB (MathWorks Inc.) plugin
Fig. 2. SSVEP stimulus screen layout
Fig. 3. Trial outline
Psychtoolbox ( [25], [26], [27]. Eight SSVEP stimulus fre-
quencies: 6.66, 7.5, 8.57, 10, 12, 15, 20, and 30 Hz, were
produced using the method outlined by Cecotti et al. [28], and
displayed on a 60 Hz screen in a 3x3 layout, as shown in Fig.
2.
C. Data Collection
Each participant’s EEG activity was recorded as they gazed
at the on-screen stimulus, using a Neuroelectrics1Enobio
20-channel EEG system with AgCl electrodes, referenced
to the right mastoid. In a single group of eight trials, the
participant was instructed (via the fixation cross) to gaze at
subsequent stimulus squares in a left-to-right, top-to-bottom
fashion, meaning the frequency values increased for each of
the eight trials. This pattern was repeated for all 30 groups of
trials, giving a total of 240 trials, which took 30 minutes per
participant. Each individual trial lasted seven seconds: a two-
second fixation period, followed by five seconds of SSVEP
stimulation (Fig. 3). Participants were given a one-minute
break every nine minutes. During recording, participants were
seated 60 cm away from the screen, in a room with reduced
natural light.
D. Standard CCA Feature Extraction
CCA is a method for identifying the underlying correla-
tion between two multidimensional variables, and has been
successfully used to perform unsupervised SSVEP detection
[14], [15]. For two multidimensional variables Xand Ywith
weighted linear combinations x=XTWXand y=YTWY,
CCA works by finding the weight vectors WXand WY
which maximise the correlation between xand y. This is
accomplished by solving the following optimisation problem:
1www.neuroelectrics.com
max
WX,WY
ρ(x,y) = E[xy]
pE[xx]E[yy]
=E[WT
XXY TWY]
qE[WT
XXXTWX]E[WT
YY Y TWY]
,
(1)
where E[x]represents the expected value of x, and ρis the
correlation value, which is maximised with respect to weight
vectors WXand WY, thereby calculating the canonical cor-
relation between Xand Y.
In the case of SSVEP detection, XRC×Sis a mul-
tidimensional EEG signal with Cchannels and Ssamples.
YfR2Nh×Sis a multidimensional set of reference signals
based on stimulus frequency f, with 2Nhindividual sine waves
and Ssamples, where Nhis the number of harmonics. The
sine waves are assembled into a matrix [14]:
Yf=
sin(2πf t)
cos(2πf t)
...
sin(2πNhf t)
cos(2πNhf t)
,(2)
where tis the time in seconds. By performing CCA on X
and Yffor all f, the stimulation frequency with the maximal
canonical correlation value can be identified, which is selected
as the estimated SSVEP frequency.
One of the main advantages of using CCA in SSVEP-BCIs
is that it can be used without any training data. In order to
retain these benefits, this study is focused on normalization
methods that can classify commands without the use of train-
ing data. For convenience, CCA without any normalization
methods applied will be referred to as Standard CCA from
this point onwards. As noted in Section I, SSVEPs elicited by
higher frequency RVS are harder to detect. As such, two new
methods aimed at improving the accuracy of Standard CCA
through the use of pre-trial normalization have been proposed.
E. Proposed Methods: Pre-Trial Normalization
Three CCA methods are compared in this paper; they are
described below along with plots of their correlation values
across time for a single participant (values are averaged across
30 trials).
Standard CCA: This is CCA without any normalization
method applied (Fig. 4)
Baseline-Corrected CCA: BC-CCA subtracts baseline corre-
lation values from the Standard CCA correlation scores at the
target time (Fig. 5);
Scaled CCA: This method divides the Standard CCA correla-
tion scores at the target time by the baseline correlation values
(Fig. 6).
The first step of applying either normalization method
requires data from the pre-trial fixation period, during which
no RVS is displayed on-screen (Fig. 3). A baseline correlation
score is calculated across the pre-fixation period by calculating
Fig. 4. Standard canonical correlation coefficients
Fig. 5. Baseline-corrected canonical correlation coefficients
the maximum canonical correlation for each class multiple
times using an overlapping window. Taking the mean of these
scores gives a single value for each class, which will be termed
the “baseline ρ” for convenience. This can be calculated for
each class using:
baselineρ =1
K
K
X
i=1
ρ(wi,Yf)(3)
where K is the total number of time windows used, and wi
represents the ith time window of data.
Later in the trial, the baseline ρcan be used to perform
normalization against the Standard CCA correlation scores
using either BC-CCA:
BC CCA =ρ(x,Yf)baselineρ (4)
or Scaled CCA:
scaledCCA =ρ(x,Yf)
baselineρ (5)
with the maximum value across classes selected as the classi-
fier output.
Fig. 6. Scaled canonical correlation coefficients
TABLE I. FR EQU ENC Y CO MBI NATIO N GRO UP S
Condition Freq. 1 Freq. 2 Freq. 3 Freq. 4 Range
Low 6.66 Hz 7.5 Hz 8.57 Hz 12 Hz 5.34 Hz
Medium 8.57 Hz 12 Hz 15 Hz 20 Hz 11.43 Hz
Wide Range 6.66 Hz 8.57 Hz 12 Hz 30 Hz 23.43 Hz
F. Method Application
Normalization requires calculating the correlation coef-
ficients for each class several times during a single trial.
To achieve this, the EEG data was downsampled to 250
Hz and separated into analysis windows using MATLAB
plugin FieldTrip [29]. Each analysis window contained one
second of data, filtered from 1-49 Hz using a zero-phase
Butterworth band-pass filter with two seconds of data padding
on either side. The ρvalues of each class were calculated
for every analysis window. The start points of the analy-
sis windows, that is, the left corners, were positioned as
follows: the windows for calculating baseline ρwere offset
to t∆ = [2,1.8,1.6,1.4,1.2] seconds, relative to
t0(stimulus onset). These overlapping one-second windows
effectively covered most of the two-second period between
the previous trial and stimulus onset of the current trial. The
analysis windows for calculating Standard CCA was offset to
t∆ = 1 second relative to t0, in order to avoid the “dead
time” [30], the period occurring after RVS onset but before
the SSVEP response reaches maximum effectiveness.
Offline analysis tests were conducted using four different
frequencies, which would provide enough degrees of freedom
to control many simple games or assistive devices. The stim-
ulation frequencies were separated into three conditions: Low
Frequency, Medium Frequency, and Wide Range condition
(Table I). These three conditions allowed for combinations
of RVS frequencies that had no inter-frequency interference
within the first three harmonics. Analysis included all 30 trials
for each class, giving a total of 120 trials per condition. Each
trial had Standard CCA, Scaled CCA, and BC-CCA applied
to it.
IV. RES ULT S
Each participant had their data (120 trials per condition,
three conditions) analysed using the Standard CCA, Scaled
CCA, and BC-CCA methods (Fig. 7). The highest accura-
cies were found in the Low Frequency condition (mean =
72.53%), followed by the Medium Frequency condition (mean
TABLE II. ME AN ACC UR ACY ACR OSS CONDITIONS
Condition Standard CCA (%) BC-CCA (%) Scaled CCA (%)
Low 73.48 72.99 71.13
Medium 64.46 71.72 70.20
Wide Range 57.84 67.06 64.41
Fig. 7. Mean classification accuracy
TABLE III. CLASSIFICATION ACCURACY (WIDE RANGE CONDITION)
Participant Standard CCA (%) BC-CCA (%) Scaled-CCA (%)
1 33.33 34.17 33.33
2 75.83 87.50 89.17
3 50 50.83 49.17
4 75 87.50 83.33
5 35 40 44.17
6 61.67 70.83 70.83
7 76.67 96.67 90
8 50.83 57.50 51.67
9 55.83 64.17 66.67
10 35.83 41.67 42.50
11 26.67 34.17 30.83
12 61.67 74.17 74.17
13 72.50 85.83 86.67
14 70 87.50 80.83
15 74.17 94.17 85
16 70 80 70
17 58.33 53.33 46.67
Mean 57.84 67.06 64.41
= 68.79%), with the lowest accuracies found in the Wide
Range condition (mean = 63.11%). Standard CCA has a
very slightly improved performance in the Low Frequency
condition (+0.49%); however, both Scaled CCA and BC-CCA
outperformed it in the other conditions (Table II), with BC-
CCA outperforming it by 7.26% in the Medium Frequency
condition, and by 9.22% in the Wide Range condition. A
closer look at the Wide Range condition (Table III) shows that
this effect is fairly consistent across participants, with only
one user performing better using Standard CCA. Separating
participants into performance-based groups using Tan et al.’s
[31] threshold for acceptable BCI control accuracy (>70%
accuracy) produces 11 higher accuracy participants versus 6
lower accuracy participants. The performance of these groups
in the Wide Range condition suggests that the majority of
improvements are made by the more accurate participants (Fig.
8, +11.81%), with less change attributed to the less accurate
participants (Fig. 9, +4.45%).
Fig. 8. Higher accuracy participants (wide condition)
Fig. 9. Lower accuracy participants (wide condition)
V. DISCUSSION
This study has investigated the problem of whether it is
possible to further improve CCA performance without the use
of training data, through the application of pre-trial normaliza-
tion methods. The results show that it is indeed possible, and its
effectiveness is dependent upon the RVS frequencies selected
for use. Both Scaled CCA and BC-CCA were found to be
effective in some cases: for low frequencies BC-CCA (-0.49%)
and Scaled CCA (-2.35%) were outperformed by Standard
CCA; however, for medium frequencies, BC-CCA (+7.26%)
and Scaled CCA (+5.74%) both had a higher accuracy than
Standard CCA, and both outperformed it again for wide range
frequencies (BC-CCA = +9.22%, Scaled CCA = +6.57%).
As shown by the plots of each method’s canonical coef-
ficient values across time (Figs. 4, 5, and 6), BC-CCA and
Scaled CCA appear to minimize the difference between the
CCA coefficients, thereby making it more likely that weaker
SSVEP responses such as at 20 and 30 Hz can be correctly
detected. However, it is unclear why BC-CCA appears to
perform better than Scaled CCA on a fairly consistent basis.
As it is a baseline correction method, BC-CCA preserves
the changes of each frequencies correlation score over time,
relative to itself; it simply equalizes their value at t0. Whereas,
Scaled CCA effectively applies a penalty to frequencies with
a high baseline ρ, and applies that to their correlation score
at every time point which should theoretically allow weaker
frequencies a stronger response. This should give some insight
into why the methods perform differently, although further
work is required to determine which situations are preferable
for each method.
Future work should look at whether training data can be
used to further improve the results of BC-CCA and Scaled
CCA, and test their effectiveness with a larger number of
frequencies. A more structured approach to selecting the pre-
trial fixation period may reduce the computations required for
real-time control.
VI. CONCLUSION
BC-CCA and Scaled CCA were both found to be effective
normalization methods, mitigating the decrease in BCI perfor-
mance seen as the distance between frequencies increases, thus
allowing a greater range of visual stimulus frequencies to be
selected. Of all the methods investigated, BC-CCA was found
to be the most effective.
ACK NOW LE DG EM EN T
The authors would like to thank Sheffield Robotics for
funding this research.
REFERENCES
[1] G. Pfurtscheller, D. Flotzinger, and J. Kalcher, “Brain-
Computer Interface-a new communication device for hand-
icapped persons,” pp. 293–299, 1993. [Online]. Available:
http://dx.doi.org/10.1006/jmca.1993.1030
[2] R. P. N. Rao, Brain-Computer Interfacing: An Introduction. Cambridge
University Press, 2013.
[3] J. Henshaw, W. Liu, and D. Romano, “Problem Solving Using Hybrid
Brain-Computer Interface Methods : a Review,” in Cognitive Infocom-
munications (CogInfoCom), 2014 5th IEEE Conference on, 2014, pp.
215–219.
[4] G. R. Mcmillan, G. L. Calhoun, M. S. Middendorf, J. H. Schnurer, D. F.
Ingle, and V. T. Nasman, “Direct brain interface utilizing self-regulation
of steady-state visual evoked response (SSVER),Rehabilitation Engi-
neering and Assistive Technology Society of North America, no. Figure
1, pp. 693–695, 1995.
[5] C. S. Herrmann, “Human EEG responses to 1-100 Hz flicker: Res-
onance phenomena in visual cortex and their potential correlation to
cognitive phenomena,Experimental Brain Research, vol. 137, no. 3-4,
pp. 346–353, 2001.
[6] H. Bakardjian, T. Tanaka, and A. Cichocki, “Optimization of SSVEP
brain responses with application to eight-command Brain-Computer
Interface.” Neuroscience letters, vol. 469, no. 1, pp. 34–38, jan 2010.
[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/19931592
[7] G. R. Burkitt, R. B. Silberstein, P. J. Cadusch, and a. W. Wood,
“Steady-state visual evoked potentials and travelling waves.” Clinical
neurophysiology : official journal of the International Federation of
Clinical Neurophysiology, vol. 111, no. 2, pp. 246–258, 2000.
[8] B. Graimann, B. Allison, and G. Pfurtscheller, Brain-computer inter-
faces: Revolutionizing human-computer interaction. Springer Science
& Business Media, 2010.
[9] A. J. McDaid, S. Xing, and S. Q. Xie, “Brain controlled robotic
exoskeleton for neurorehabilitation,” 2013 IEEE/ASME International
Conference on Advanced Intelligent Mechatronics: Mechatronics for
Human Wellbeing, AIM 2013, pp. 1039–1044, 2013.
[10] T. Sakurada, T. Kawase, K. Takano, T. Komatsu, and K. Kansaku, “A
BMI-based occupational therapy assist suit: Asynchronous control by
SSVEP,” Frontiers in Neuroscience, vol. 7, no. 7 SEP, pp. 1–10, 2013.
[11] S. M. T. M¨
uller and T. F. Bastos-filho, “Brain-computer Interface
Based on Visual Evoked Potentials to Command Autonomous Robotic
Wheelchair,” Journal of Medical and Biological Engineering, vol. 30,
no. 6, pp. 407–416, 2010.
[12] G. Vanacker, J. Del R. Mill´
an, E. Lew, P. W. Ferrez, F. G. Moles,
J. Philips, H. Van Brussel, and M. Nuttin, “Context-based filtering for
assisted brain-actuated wheelchair driving,” Computational Intelligence
and Neuroscience, vol. 2007, pp. 1–12, 2007.
[13] C. J. Bell, P. Shenoy, R. Chalodhorn, and R. P. N. Rao, “Control of a
humanoid robot by a noninvasive brain-computer interface in humans.”
Journal of neural engineering, vol. 5, no. 2, pp. 214–20, jun 2008.
[Online]. Available: http://www.ncbi.nlm.nih.gov/pubmed/18483450
[14] Z. Lin, C. Zhang, W. Wu, and X. Gao, “Frequency Recognition Based
on Canonical Correlation Analysis for SSVEP-Based BCIs,” IEEE
Transactions on Biomedical Engineering, vol. 53, no. 12, pp. 315–323,
2006. [Online]. Available: http://link.springer.com/10.1007/978-3-319-
12436-0
[15] G. Bin, X. Gao, Z. Yan, B. Hong, and S. Gao, “An online multi-channel
SSVEP-based brain-computer interface using a canonical correlation
analysis method.” Journal of neural engineering, vol. 6, no. 4, p.
046002, 2009.
[16] M. Nakanishi, Y. Wang, Y.-t. Wang, Y. Mitsukura, and T. Jung, “En-
hancing Unsupervised Canonical Correlation Analysis-Based Frequency
Detection of SSVEPs by Incorporating Background EEG,” in 36th
Annual International Conference of the IEEE Engineering in Medicince
and Biology Society, 2014, pp. 3053–3056.
[17] Y. Zhang, G. Zhou, J. Jin, X. Wang, and A. Cichocki, “FRE-
QUENCY RECOGNITION IN SSVEP-BASED BCI USING MULTI-
SET CANONICAL CORRELATION ANALYSIS,” International Jour-
nal of Neural Systems, vol. 24, no. 4, p. 1450013, 2014.
[18] S. N. Carvalho, T. B. Costa, L. F. Uribe, D. C. Soriano, G. F. Yared,
L. C. Coradine, and R. Attux, “Comparative analysis of strategies for
feature extraction and classification in SSVEP BCIs,” Biomedical Signal
Processing and Control, vol. 21, pp. 34–42, 2015. [Online]. Available:
http://linkinghub.elsevier.com/retrieve/pii/S1746809415000877
[19] J. Castillo, S. Muller, E. Caicedo, and T. Bastos, “Feature
extraction techniques based on power spectrum for a SSVEP-BCI,
in 2014 IEEE 23rd International Symposium on Industrial
Electronics (ISIE), 2014, pp. 1051–1055. [Online]. Available:
http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=6864758
[20] O. Friman, I. Volosyak, and A. Gr¨
aser, “Multiple channel
detection of steady-state visual evoked potentials for brain-
computer interfaces.” IEEE transactions on bio-medical engineering,
vol. 54, no. 4, pp. 742–50, 2007. [Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/17405382
[21] S. Resalat and S. Setarehdan, “An Improved SSVEP Based BCI
System Using Frequency Domain Feature Classification,” American
Journal of Biomedical . . . , vol. 3, no. 1, pp. 1–8, 2013. [Online].
Available: http://article.sapub.org/10.5923.j.ajbe.20130301.01.html
[22] P. Yuan, X. Gao, B. Allison, Y. Wang, G. Bin, and S. Gao,
“A study of the existing problems of estimating the information
transfer rate in online brain-computer interfaces.” Journal of neural
engineering, vol. 10, no. 2, p. 026014, 2013. [Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/23448963
[23] T. Sakurada, T. Kawase, T. Komatsu, and K. Kansaku, “Use
of high-frequency visual stimuli above the critical flicker
frequency in a SSVEP-based BMI,” Clinical Neurophysiology,
vol. 126, no. 10, pp. 1972–1978, 2015. [Online]. Available:
http://dx.doi.org/10.1016/j.clinph.2014.12.010
[24] P. F. Diez, V. A. Mut, E. Laciar, and E. M. A.
Perona, “Mobile robot navigation with a self-paced brain-
computer interface based on high- frequency SSVEP,
Robotica, vol. 32, no. November 2013, pp. 695–709, 2014.
[Online]. Available: http://www.scopus.com/inward/record.url?eid=2-
s2.0-84888194023&partnerID=tZOtx3y1
[25] D. H. Brainard, “The Psychophysics Toolbox.” Spatial vision, vol. 10,
no. 4, pp. 433–436, 1997.
[26] D. G. Pelli, “The VideoToolbox software for visual psychophysics:
transforming numbers into movies.” pp. 437–442, 1997.
[27] M. Kleiner, D. Brainard, D. Pelli, A. Ingling, R. Murray, and C. Brous-
sard, “What ’ s new in Psychtoolbox-3 ? Foreword ; - ),Perception,
vol. 36, no. 14, pp. 1–89, 2007.
[28] H. Cecotti, I. Volosyak, and A. Gr¨
aser, “RELIABLE VISUAL STIM-
ULI ON LCD SCREENS FOR SSVEP BASED BCI,” in The 2010
European Signal Processing Conference (EUSIPCO-2010), 2010, pp.
919–923.
[29] R. Oostenveld, P. Fries, E. Maris, and J. M. Schoffelen, “FieldTrip:
Open source software for advanced analysis of MEG, EEG, and
invasive electrophysiological data,” Computational Intelligence and
Neuroscience, vol. 2011, 2011.
[30] M. Sengelmann, A. K. Engel, and A. Maye, “Maximizing
information transfer in SSVEP-based brain computer
interfaces,” IEEE Transactions on Biomedical Engineering,
vol. 9294, no. c, pp. 1–1, 2016. [Online]. Available:
http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=7460915
[31] D. Tan and A. Nijholt, Human-computer Interaction, 2010.
... Their method calculates the CCA correlation coefficients for background EEG activity within neighbouring frequency bands, and summed them before dividing the CCA correlation coefficient at the target frequency by this, creating a ratio value similar to signal-to-noise-ratio (SNR). Testing the technique on an offline EEG dataset (n = 13) for eight frequencies (8)(9)(10)(11)(12)(13)(14)(15) Hz, ∆ f = 1 Hz), this method was found to significantly increase accuracy when compared to standard CCA (84.89% vs. 80.08%), and improved detection accuracy for the higher frequency SSVEPs more when compared to the lower frequency SSVEPs. Castillo et al. [5] devised a similar method in their study (n = 19), where they used PSD features to create a ratio, this time using signal amplitude and 'error', where error is the absolute difference between the peak frequency in the power spectrum and the amplitude at the target frequency. ...
... Three CCA-based methods were compared: Standard CCA, and CCA with two proposed normalisation methods [9]. The CCA-based normalisation methods we proposed used the pre-trial fixation period to calculate a correlation baseline for normalisation, termed 'baseline ρ', which was calculated using: ...
... computer) operatable commands [1]. The state-of-the art BCI is based on the idea of developing an artificial, muscle-free communication channel that acts as a natural communication channel between the brain and a machine [2,3]. Applications of BCI systems are widespread and vary from the fields of neuroscience, rehabilitation, cognitive infocommunications (CogInfoCom) [4] to entertainment, and defence [5]. ...
Article
The lower-limb representation area in the human sensorimotor cortex has all joints very closely located to each other. This makes the discrimination of cognitive states during different motor imagery tasks within the same limb, very challenging; particularly when using electroencephalography (EEG) signals, as they share close spatial representations. Following that more research is needed in this area, as successfully discriminating different imaginary movements within the same limb, in form of a single cognitive entity, could potentially increase the dimensionality of control signals in a braincomputer interface (BCI) system. This report presents our research outcomes in the discrimination of left foot-knee vs. right foot-knee movement imagery signals extracted from EEG. Each cognitive state task outcome was evaluated by the analysis of eventrelated desynchronization (ERD) and event-related synchronization (ERS). Results reflecting prominent ERD/ERS, to draw the difference between each cognitive task, are presented in the form of topographical scalp plots and average time course of percentage power ERD/ERS. Possibility of any contralateral dominance during each task was also investigated. We have compared the topographical distributions and based on the results we were able to distinguish between the activation of different cortical areas during foot and knee movement imagery tasks. Currently, there are no reports in the literature on discrimination of different tasks within the same lower-limb. Hence, an attempt towards getting a step closer to this has been done. Presented results could be the basis for control signals used in a cognitive infocommunication (CogInfoCom) system to restore locomotion function in a wearable lower-limb rehabilitation system, which can assist patients with spinal cord injury (SCI).
... CogInfoCom covers several disciplines appearing in applications and research areas also. CogInfoCom is available technology from socio-cognitive ICT [55][56][57][58][59][60][61][62][63][64][65] to cognitive aided engineering [66][67][68][69][70][71][72][73][74][75][76][77] and its related aspects in terms of online collaborative systems and virtual reality solutions [78][79][80], teaching-learning [81][82][83][84][85][86] and human cognitive interfaces such as braincomputer interfaces (BCIs) [87][88][89][90][91][92] and medicals [93][94][95][96]. ...
Article
Full-text available
Brain research is one of the most significant research areas of the last decades, in which many developments and modern engineering technologies are applied. The electroencephalogram (EEG)-based brain activity observation processes are very promising and have been used in several engineering research fields. Objective: The main goal of this research was to develop a Brain-Computer Interface (BCI) system to observe the level of vigilance calculated by Think Gear-ASIC Module (TGAM1) technology and to evaluate the output with learning efficiency tests applied in cognitive neuroscience. Methods: The performance of the BCI system is evaluated in a comparative study. The BCI system was tested by thirty-two test subjects and the attention level output was compared by the Psychology Experiment Building Language's (PEBL's) Corsi block test (P CORSI) and PEBL's Ebbinghaus procedure (P EBBINGHAUS) tasks. Results: Using the BCI, we have shown statistically significant results between the BCI and the conventional cognitive neuroscience tests. The correlation between the tests and the average attention of the BCI was slightly strong for P CORSI Total Score (r=.63, p<.01 (2-tailed) and slightly strong for P EBBINGHAUS Total Correct (r=-.71, p<.01 (2-tailed). The average level of attention measured by the BCI system during the P CORSI test was 49.00%, while in case of the P EBBINGHAUS test it was 52.41% on all samples. Conclusion: The developed BCI system has a significant correlation with P CORSI and P EBBINGHAUS cognitive neuroscience tests. The BCI system is capable of observing attentional vigilance continuously. Significance: The developed BCI system is applicable to observe vigilance level in real-time while the level of attention depends on activities.
... These human-machine interface systems have appeared in many applications and research areas, basically called the field of cognitive infocommunications (CogInfoCom). CogInfoCom, involves the speechability, communication, linguistic and behavioural interaction analysis, face and gesture recognition, brain-computer interface (BCI), human cognitive interfacevirtual and real avatars dealing with the complex mixture of human and artificial cognitive capabilities in human computer interaction processes [48][49][50][51][52][53][54][55][56][57][58][59][60][61][62][63]. The socio-cognitive ICT field includes collective knowledge, cognitive networks and their intelligent capabilities. ...
Article
Full-text available
New technological advances of the 20 th and 21 st Centuries provide several new opportunities for engineers in the future. Modern engineering methods feasible by these new technical solutions, however, cannot meet expectations in all fields and situations without adequate knowledge and competence. Spreading new and modern engineering methods as well as acquiring new applications is a crucial issue in education. In brain research, which is one of the most significant research areas of the past decades, many new results and meters have appeared that could be used in engineering methods, too. Based on brain activity observation, new meters open up new horizons in engineering applications. Electroencephalogram-based brain activity observation processes are very promising and have been used in several engineering research primarily for implementation of control tasks. In this paper, an EEG-based engineering research work is demonstrated, which supports the acquirement of practical knowledge and can measure cognitive ability with a device capable of brain activity observation. In the engineering research task, a brain-computer interface (BCI) had to be developed for the measurement of the average level of attention. The results of the BCI have been compared and contrasted to the results of two tests applied in cognitive psychology, the PEBL Continuous Performance Test (pCPT) and the PEBL Test of Attentional Vigilance (pTOAV). It can be stated that the results of the procession developed in this research and the results of the pCPT and the pTOAV tests are in correlation.
Article
Full-text available
A brain–computer interface (BCI) is a system for commanding a device by means of brain signals without having to move any muscle. One kind of BCI is based on Steady-State Visual Evoked Potentials (SSVEP), which are evoked visual cortex responses elicited by a twinkling light source. Stimuli can produce visual fatigue; however, it has been well established that high-frequency SSVEP (>30 Hz) does not. In this paper, a mobile robot is remotely navigated into an office environment by means of an asynchronous high-frequency SSVEP-based BCI along with the image of a video camera. This BCI uses only three electroencephalographic channels and a simple processing signal method. The robot velocity control and the avoidance obstacle algorithms are also herein described. Seven volunteers were able to drive the mobile robot towards two different places. They had to evade desks and shelves, pass through a doorway and navigate in a corridor. The system was designed so as to allow the subject to move about without restrictions, since he/she had full robot movement's control. It was concluded that the developed system allows for remote mobile robot navigation in real indoor environments using brain signals. The proposed system is easy to use and does not require any special training. The user's visual fatigue is reduced because high-frequency stimulation is employed and, furthermore, the user gazes at the stimulus only when a command must be sent to the robot.
Conference Paper
Full-text available
The latest advancements in hybrid brain-computer interface (hBCI) research are reported in this paper. Specifically we look at the latest pure hBCI systems, which use two or more different brain-computer interface (BCI) signals to allow the user to interact with a computer system. A special emphasis is placed on investigating how problems encountered by standard BCI systems are now being tackled by hBCI methods.
Conference Paper
Full-text available
This paper presents a comparison among three methods for Steady-State Visually Evoked Potentials (SSVEP) detection. These techniques are based on Power Spectral Density Analysis (PSDA) and Canonical Correlation Analysis (CCA). The first method estimates the signal-to-noise ratio of the power spectrum in each stimulus frequency using PSDA, which is called Traditional-PSDA. The second analysis estimates the relation between the difference of the stimulus frequency and its neighbor frequencies, using the power spectrum in these neighbor frequencies, and seeks the neighbor frequency which present the lowest relation value. This technique is referred to Ratio-PSDA. The third and final technique called Hybrid-PSDA-CCA. The performances of the methods were evaluated using a database of electroencephalogram (EEG) signals. The EEG signals were recorded from 19 volunteers, from which six people present disabilities. They were stimulated with visual stimuli flickering at 5.6, 6.4, 6.9 and 8.0 Hz. The system performance was evaluated considering the accuracy and the Information Transfer Rate (ITR) for each stimulus frequency. The results showed that the Hybrid-PSDA-CCA method achieved the best result with an average accuracy of 91.44%.
Article
Compared to the different brain signals used in BCI paradigms, the SSVEP features a high signal to noise ratio, enabling reliable and fast classification of neural activity patterns without extensive training requirements. Objective: In this article we present methods to further increase the ITRs of SSVEP-based BCIs. Methods: Starting with stimulus parameter optimizations methods, we develop an improved approach for use of the Canonical Correlation Analysis (CCA) and analyze properties of the SSVEP when the user fixates a target and during transitions between targets. Results: These transitions show a negative effect on the system's ITR which we trace back to delays and dead times of the SSVEP. Conclusion: Using two classifier types adapted to continuous and transient SSVEPs and two control modes (fast feedback and fast input), we present a simulated online BCI implementation which addresses the challenges introduced by transient SSVEPs. Significance: The resulting system reaches an average ITR of 181 Bits/min and peak ITR values of up to 295 Bits/min for individual users.
Article
The idea of interfacing minds with machines has long captured the human imagination. Recent advances in neuroscience and engineering are making this a reality, opening the door to restoration and augmentation of human physical and mental capabilities. Medical applications such as cochlear implants for the deaf and neurally controlled prosthetic limbs for the paralyzed are becoming almost commonplace. Brain-computer interfaces (BCIs) are also increasingly being used in security, lie detection, alertness monitoring, telepresence, gaming, education, art, and human augmentation. This introduction to the field is designed as a textbook for upper-level undergraduate and first-year graduate courses in neural engineering or brain-computer interfacing for students from a wide range of disciplines. It can also be used for self-study, and as a reference by neuroscientists, computer scientists, engineers, and medical practitioners. Key features include: Essential background in neuroscience, brain recording and stimulation technologies, signal processing, and machine learning Detailed description of the major types of BCIs in animals and humans, including invasive, semi-invasive, non-invasive, stimulating, and bidirectional BCIs In-depth discussion of BCI applications and BCI ethics Questions and exercises in each chapter Supporting website with annotated list of book-related links.
Article
Brain-computer interface (BCI) systems based on electroencephalography have been increasingly used in different contexts, engendering applications from entertainment to rehabilitation in a non-invasive framework. In this study, we perform a comparative analysis of different signal processing techniques for each BCI system stage concerning steady state visually evoked potentials (SSVEP), which includes: 1) feature extraction performed by different spectral methods (bank of filters, Welch's method and the magnitude of the short-time Fourier transform); 2) feature selection by means of an incremental wrapper, a filter using Pearson's method and a cluster measure based on the Davies-Bouldin index, in addition to a scenario with no selection strategy; 3) classification schemes using linear discriminant analysis (LDA), support vector machines (SVM) and extreme learning machines (ELM). The combination of such methodologies leads to a representative and helpful comparative overview of robustness and efficiency of classical strategies, in addition to the characterization of a relatively new classification approach (defined by ELM) applied to the BCI-SSVEP systems.
Article
This study presents a new steady-state visual evoked potential (SSVEP)-based brain-machine interface (BMI) using flickering visual stimuli at frequencies greater than the critical flicker frequency (CFF). We first asked participants to fixate on a green/blue flicker (30-70Hz), and SSVEP amplitude was evaluated. Participants were asked to indicate whether the stimulus was visibly flickering and to report their subjective level of discomfort. We then assessed visibly (41, 43, and 45Hz) vs. invisibly (61, 63, and 65Hz) flickering stimulus in an SSVEP-based BMI. Visual fatigue was assessed via the flicker test before and after operation of the BMI. Higher frequency stimuli reduced participants' subjective discomfort. Participants successfully controlled the SSVEP-based BMI using both the visibly and invisibly flickering stimuli (93.1% and 88.0%, respectively); the flicker test revealed a decrease in CFF (i.e., visual fatigue) under the visible condition only (-5.7%, P<0.001). The use of high-frequency visual stimuli above the CFF led to high classification accuracy and decreased visual fatigue in an SSVEP-based BMI. High-frequency flicker stimuli above the CFF were able to induce SSVEPs and may prove useful in the development of BMI-based assistive products. Copyright © 2014 International Federation of Clinical Neurophysiology. Published by Elsevier Ireland Ltd. All rights reserved.