Content uploaded by Simanto Saha
Author content
All content in this area was uploaded by Simanto Saha on Feb 11, 2018
Content may be subject to copyright.
Unifying Sensorimotor Dynamics in Multiclass Brain
Computer Interface
Simanto Saha 1, Khawza I. Ahmed 2, Raqibul Mostafa 3
Department of Electrical and Electronic Engineering
United International University, Dhaka-1209, Bangladesh
Email: simanto.saha.bd@ieee.org 1, khawza@eee.uiu.ac.bd 2, rmostafa@eee.uiu.ac.bd 3
Abstract—Unification of spatial brain dynamics in multiclass
brain computer interface (BCI) paradigm reduces computational
latencies by using lesser number of electrodes from the senso-
rimotor regions of the brain. We employ reduced number of
channels without compromising performance notably. We apply
three spatial filtering methods, i.e., Common Spatial Pattern (CSP),
Regularized Common Spatial Pattern (RCSP) and Joint Approx-
imate Diagonalization (JAD) as preprocessing. But, we emphasize
on selecting specific EEG montages for BCI development. We
achieve best 86.7% classification accuracy for subject k3b applying
CSP using only 12 channels from sensorimotor regions instead of
using 60 channels from the whole brain. Additionally, the average
classification accuracies are 64.4% and 61.4% using 60 channels
and 12 channels respectively. Also, the average computational
latencies are 6.24s and 1.23s in cases of 60 channels and 12 channels
respectively.
I. INTRODUCTION
The Brain Computer Interface (BCI) is an unorthodox direct
communication pathway between human brain and the computer
without any muscular stimulation. The applications of BCI
systems are on motor function rehabilitation, lie detection, brain
fingerprinting, mode assessment and mind control etc. The brain
poses signatures of any cortical activity as electrical signal. The
electroencephalography (EEG) is the process of recording elec-
trical activity of the brain by placing electrodes directly on the
scalp. The EEG, a noninvasive signal acquisition technique, is
very popular due to its low cost and mobility. But, multichannel
recording of EEG introduces computational burden. The aim of
this paper is to employ reduced number of channels utilizing
the knowledge of sensorimotor spatial dynamics for classifying
multiclasses of motor imagery (MI) tasks.
The MI task is the imagination of the motor task rather to do it
actual. According to Jeannerod, brain’s response to the conscious
MI task is functionally equivalent to the unconscious motor
execution (ME) [1]. That’s why, MI based BCI is very popular
as it seems to be a potential application in motor rehabilitation
for motor impaired subjects (i.e., people having Tetraplegia).
Recent experiment on BCI evinces the association between
electrophysiological and hemodynamic signatures during MIs
and it supports the Jeannerod’s statements about MIs empirically
[2]. However, the sensorimotor areas such as Primary Motor
Cortex (M1), Supplementary Motor Area (SMA) and Premotor
Cortex (PMC), etc., are mostly activated during MI tasks [3]-[5].
Besides, there are significant activation in parietal lobe as well.
But, the parietal lobe is associated with visual perception and
visual imagery (VI). There is a significant difference between
the visual imagery (VI) and the motor imagery (MI). VI is the
visualization of the specific task whereas the motor imagery
is the kinesthetic feelings of the specific task. There is still
controversy about the exact spatio-temporal characterization of
the brain regions. Meanwhile, our investigation is based on
the assumption about MI cortical sources from recent litera-
ture. The assumption is ”the source activation during the MI
happens mostly in the sensorimotor (i.e. M1, SMA, PMC etc.)
areas” [2]-[5]. Previously, we have shown that integration of
the sensorimotor dynamics can help selecting reduced number
of channels while optimizing the computational latencies [6].
In this paper, we report the same experimentation on multiclass
paradigm. Also, we have applied an additional preprocessing
method named Joint Approximate Diagonalization (JAD) along
with Common Spatial Pattern (CSP) and Regularized Common
Spatial Pattern (RCSP). In our previous experiment, we have
assumed the left hemisphere is responsible for two MI tasks,
i.e., right hand and right foot. But, we consider both hemisphere
in this experiment for multiclasses MIs, i.e. right hand, left hand,
foot and tongue.
The optimal EEG source localization for MI tasks plays
crucial role while improving BCI performance and reducing
computational cost. Most of the relevant literature suggests
various signal processing and/or machine learning tools for
optimal channel selection. In [7], the ‘Laplacian Derivative
(LAD) of power averaged across frequency bands’ features
are used to select optimal channels for stroke patients. The
channels are selected based on minimum redundancy which is
important for channel optimization while eliminating already
selected channels. A CSP based channel selection is proposed
[8] for the first instance in which the actual EEG montages
were correlated with optimal CSP channels. Sparse Common
Spatial Pattern (SCSP) is proposed in [9] as a novel method
for optimal channel selection technique within a constraint of
optimal classification accuracy. While our experiment is ac-
counting to the sensorimotor spatial brain dynamics for MI,
SCSP show how regularization in spatial filtering method can
help optimize number of EEG channels. Recently, a study on
Filter Bank Common Spatial Pattern (FBCSP) along with head
model geometry is proposed in [10] for developing EEG based
BCI. As resemblance to our investigation, this method accounts
the anatomical research for selecting specific channels from the
specific region of interest. However, EEG gives good temporal
resolution and poor spatial resolution while high density EEG
modality can be used as an alternative neuroimaging technique
instead of fMRI/PET [11] for capturing good spatial details
as well. An Independent Component Analysis (ICA) based
optimal EEG source localization is explained in [12]. The
proposed algorithm requires higher computational cost which
may cause limitation in many specific BCI applications. In [13],
the EEG source mapping is shown using Fourier coefficients
based method which evinces the significant activation in the
sensorimotor cortex during MI tasks. Certainly, different signal
processing and machine learning tools play crucial role while
selecting optimal channels. But, we assume integration of spatial
brain dynamics can contribute in EEG source localization as
complement to advanced scientific tools.
II. PROP OSED EXP ER IME NTA L PARADIGM
The overview of the proposed experimental paradigm is shown
in Figure 1. Firstly, we apply Butterworth filter of order 10
with cutoff frequencies 8Hz and 40Hz. Then, we apply three
different spatial filtering methods, i.e., CSP, RCSP and JAD
to extract spatially distinguishable features for four MI tasks.
After spatial filtering, we choose 8 optimal channels, 2 for each
class. From each of the optimal channels, we extract wavelet
decomposition based subband energy and subband entropy up
to level 3. As a result, we have four subband energy and four
subband entropy, accumulating to 8 attributes for each channel.
Altogether, we consider 64 attributes to train the classifier.
Finally, we use the Two-Layer Feed-Forward Neural Network
(NN) as a classifier. The three phases of classifier are Training,
Validation and Testing. The number of trials in different phases
for three experiments are summarized in Table I.
Fig. 1. The Proposed Experimental Paradigm: The EEG signal is prepro-
cessed using (1) Bandpass Filter with cut off frequencies at 8Hz and 40Hz and (2)
spatial filtering methods (CSP/RCSP/JAD). The spatial filtering gives 8 optimal
channels, 2 channels per class. Then, wavelet decomposition based subband
energy and entropy are calculated for each of the optimal channels. Finally,
classification takes place using a Two-Layer Feed-Forward Neural Network
(NN).
TABLE I
EXP ER IM EN TAL SET UP S: N UMBER OFTRIALS INTRAINING, VALIDATION
AND TES TI NG
Competition
/ Dataset
Subjects Total Train Validation Test
k3b 360 160 80 120
III / IIIa l1b 240 100 60 80
k6b 236 100 60 76
We consider three distinct Cases corresponding to different
sets of EEG channels accounting to the assumption of sen-
sorimotor dynamics for MI tasks. The Case III is based on
the assumption ”the source activation during the MI happens
mostly in the sensorimotor (i.e., M1, SMA, PMC etc.) areas”.
The details of the Cases are summarized in Table II.
TABLE II
CHA NN EL SE LEC TI ON UTILIZING SENSORIMOTOR DYN AM IC S (NUMBER OF
CHANNELS)
#Channels Spatial Brain Areas
Case I 60 Whole Brain
Case II 12 Sensorimotor Areas + Parietal Lobe
(Both Hemisphere)
Case III 12 Sensorimotor Areas (Both Hemisphere)
(Both Hemisphere)
A. About The Datasets: BCI Competition III Dataset IIIa
The data was recorded for four class MI (left hand, right
hand, foot and tongue movement) following by a cue for each
trial. During data acquisition, a 64-Channel EEG amplifier from
Neuroscan, using the left mastoid for reference and the right
mastoid as ground. The EEG was sampled with 250 Hz, it was
filtered between 1 Hz and 50 Hz with Notch Filter on. Finally,
sixty EEG channels were recorded for different MI tasks [14]-
[15].
The subject sat on a relaxing chair with arms resting and asked
to perform one of the four MI tasks. The order of the cues was
random. The experiment consist of several runs (at least 6) with
40 trails each. For each trial, the first 2s were quiet, at t=2s an
acoustic stimulus was given to subject indicating the beginning
of each trial, then a + is shown. From t=3s, an arrow to the left,
right, up and down was displayed for 1s. At the same time, the
subject was asked to imagine a left hand, right hand, tongue or
foot movement according to the cue until the cross disappeared
at t=7s. Each of the four cues displayed 10 times within each run
in a randomized order. The data was recorded for three subjects.
Fig. 2. Time-Frequency (T-F) representation of single trial MI tasks for C3,
Cz and C4 channels: the T-F subplots are generated using Continuous Wavelet
Transform (CWT) and each subplot represents the distribution of power along
both time and frequency axes as percentiles of total power from concatenated
C3, Cz and C4 for each trial, corresponding to one of the four MI tasks. Each
trial has 3sec of data within 8Hz to 40Hz.
B. Spatial Filtering as Preprocessing
In this paper, we have applied different spatial filtering
methods such as Common Spatial Pattern (CSP), Regularized
Common Spatial Pattern (RCSP) and Joint Approximate Diag-
onalization (JAD) as preprocessing techniques. The CSP gives
spatially optimal projection of multichannel EEG data in which
the maximal discriminative features between two classes present.
In case of RCSP, regularization of different parameters are
introduced which show robustness to outliers and perform better
in case of small sample settings [16]. As CSP/RCSP refers in
two class paradigm, we have used one versus all paradigm so
that we can extract discriminative features for four MI classes.
However, we have also used CSP by JAD algorithm which can be
easily extended to multi-class paradigm. The following sections
describe briefly about CSP, RCSP and JAD algorithms.
The rest of this subsection briefly describes about CSP, RCSP
and JAD algorithms. The write-up on CSP and RCSP has been
taken from our previous work [6]. It is to be noted that these
are all established algorithms. However, since we are using all
these algorithms in our work, we are briefly describing these for
the sake of the completeness of this paper and the convenience
of the readers.
1) Common Spatial Pattern with Regularization: At first, the
CSP method was applied to EEG regarding MI hand movements
by Ramoser in [17]. CSP is very sensitive to outliers and
artefact as reported in [16] and [18]. As EEG signals have very
challenging SNR, more robust methods were needed to improve
performance. We used RCSP which is more robust compared to
CSP [16], [19].
EEG signal is represented by Eand of size N×P, where N
is the number of channels and Pis the number of samples per
trial. In traditional CSP algorithm, the sample based covariance
matrix estimation is required. The sample covariance matrix of
a trial Eis normalized with the total variance as [16], [17].
S=EET
trace hEETi(1)
where Tdenotes the transpose of a matrix.
Considering Mtrials are available for training in each class
for a subject, indexed by mas E(c,m), m = 1,2, ......, M
¯
Sc=1
M
M
X
m=1
S(c,m)(2)
where c∈ {1,2}represents either of the two classes of the trial
depending on each of the MI tasks.
The discriminative spatial patterns in CSP are calculated based
on this sample based covariance matrix estimation based on . The
next section will introduce regularization in CSP.
2) Covariance Matrix Estimation With Regularization:: Reg-
ularization works by biasing the covariance estimation away
from their sample based values towards more physically plau-
sible values, which reduces the variance of the sample based
estimates while tending to increase bias [16]. This is done using
one or more regularization parameters (i.e., βand γin this
paper).
Now, the regularized average spatial covariance matrix for
each class is defined as
d
Xc(β, γ ) = (1 −γ)b
Ωc(β) + γ
Ntrace hb
Ωc(β)i·I(3)
Here, βand γare the two regularization parameters (0 ≤β, γ ≤
1) and Iis an identity matrix. b
Ωc(β)is comprised of covariance
matrix for the trials from the specific subjects as well as generic
trials and given by
b
Ωc(β) = (1 −β).Sc+β.b
Sc
(1 −β).M +β. c
M(4)
In that case, Scis the sum of the sample covariance matrices
for all Mtraining trials for class c:
Sc=
M
X
m=1
S(c,m)(5)
Also, b
Scis the sum of the sample covariance matrices for c
M
generic training trials with covariance matrix E(c, ˆm)for class c:
b
Sc=
c
M
X
bm=1
S(c,b
m)(6)
Here, Scand b
Scare normalized and analogous to the sample
covariance matrix mentioned in (1). The objective of b
Scis to
reduce the variance in the covariance matrix estimation and may
produce more reliable results.
The regularization parameter βcontrols the shrinkage of the
training samples covariance matrix estimate to the pooled esti-
mate. The other regularization parameter γcontrols the degree
of shrinkage towards a multiple of the identity matrix, with
the average eigenvalue of ˆ
Ωc(β)as the multiplier. This second
shrinkage has the effect of decreasing the larger eigenvalues
while increasing the smaller ones. Actually, this works well in
case of small training samples set as in [16].
3) Feature Extraction In RCSP: The regularized composite
spatial covariance is formed and factorized as [16]
d
X(β, γ ) = d
X1(β, γ ) + d
X2(β, γ ) = b
Ub
∧b
UT(7)
Here, b
Udenotes the matrix eigenvectors and b
∧denotes the
diagonal matrix of corresponding eigenvalues. The eigenvalues
are assumed to be sorted in descending order throughout this
paper [16], [17].
Finally, the projection matrix is formed as [16]
c
W=b
BTb
∧−1/2b
UT(8)
where b
Bdenotes the matrix of eigenvectors for the whitened
spatial covariance matrix and defined as
b
B=b
∧−1/2b
UTd
Xc(β, γ )b
Ub
∧−1/2(9)
In RCSP, an input trial E is projected as [16]
b
X=c
WTE(10)
To get the most discriminative features for both classes, the
optimal channels are to be selected from the leftmost and
rightmost channels. For example, the first channel represents
the most distinguished features for class 1 and the last channel
represents the most distinguished features for class 2. As channel
selection converges to the central channel of the b
X, the features
becomes poor which may hardly distinguish different classes.
It is to be noted that, RCSP becomes traditional CSP when
β=γ= 0.
In this experiment, we have used different combination of
values of βand γas follows
β= (0,0.001,0.01,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9)
γ= (0,0.01,0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9)
We have tested all possible combination of values from the
above sets. These values are selected from previous experiments
described in [16] and [8].
4) Joint Approximate Diagonalization (JAD): In this section,
we assume a two class paradigm i.e. C={c1, c2}and briefly
discuss the CSP by JAD algorithm [20], [21]. The CSP algorithm
can be solved as optimization problem
w∗=argmax
w∈RN
{wTRx|c1w
wTRx|c2w}(11)
with Rx|c1,Rx|c2the covariance matices of x given c1,c2
respectively. Since (11) is in the form of well known Rayleigh
quotient, solutions to (11) are given by eigenvectors of the
generalized eigenvalue problem
Rx|c1w=λRx|c2w(12)
The eigenvectors of (12) corresponds to the desired spatial
filter. Furthermore, for a given eigenvector w∗the corresponding
eigenvalue determines the value of the cost function
λ∗=w∗TRx|c1w∗
w∗TRx|c2w∗(13)
The eigenvalues are thus a measure for the quality of the ob-
tained spatial filters, i.e. the eigenvalue associated with a spatial
filter expresses the ratio of the variance between conditions of
the component of the EEG/MEG data extracted by the spatial
filter. Pre-processing is then usually done by combining the L
eigenvectors of (12) with the smallest/largest eigenvalues to form
W∈RNX L and computing ˆx=WTx.
Extending CSP to multi-class paradigms is either done by
performing two-class CSP on different combinations of classes
(e.g. by computing CSPs for all combinations of classes or by
computing CSP for one class versus all other classes), or by Joint
Approximate Diagonalization (JAD). Since the first approach
(One vs All) is discussed in the previous section, we will focus
JAD here.
Given EEG/MEG data from Mdiffernt classes, the goal of
CSP by JAD is to find a transformation W∈RNXN that
diagonalizes the covariance matrices Rx|ci, i.e.
WTRx|ciW=Dci , i = 1, ....., M (14)
with Dci ∈RNX N diagonal matrices. The idea of using
JAD for multi-class CSP lies in the fact that CSP for two
classes can be understood as diagonalizing two covariance
matrices. More precisely, if the eigenvectors of the generalized
eigenvalue problem (12) are combined in a matrix W, then
WTRx|ciW=Dci , i = 1, ....., M. It then seems plausible to
extend CSP to multi-class paradigms by finding a transformation
Wthat approximately diagonalizes multiple covariance matrices.
A total of Lcolumns of the obtained matrix Ware then taken
as the desired spatial filters. For details on JAD algorithm, see
[20] and [21].
III. RES ULT S
The optimal EEG source selection is crucial for (a) reducing
the number of channels and thus optimizing computational
latencies and (b) compensating the effect of outliers by elim-
inating undesired channels producing outliers. We investigate
if the integration of sensorimotor brain dynamics can reduce
the number of channels, thus the computational latencies. Our
experimental objective is to emphasize the spatial brain dynam-
ics for EEG channel selection as complementary to advanced
signal processing and machine learning tools. The classification
accuracies for three different cases are reported in Table III.
The corresponding comparison of computational latencies for
all cases is shown in Table IV.
TABLE III
CLASSIFICATION ACCURACIES(%): BCI COMPETITION III DATAS ET IIIA
Subject Method Case I Case II Case III
CSP 78.3% 81.7% 86.7%
k3b RCSP 92.5% 75.8% 84.2%
JAD 87.5% 82.5% 82.5%
CSP 65.0% 52.5% 47.5%
l1b RCSP 66.2% 52.5% 62.5%
JAD 60.0% 50.0% 56.2%
CSP 42.1% 35.5% 36.8%
k6b RCSP 56.6% 47.4% 53.9%
JAD 32.9% 19.7% 42.1%
Avg. 64.6% 55.3% 61.4%
TABLE IV
COM PU TATION AL LATE NC IE S (IN S EC S): COMPARISON BETW EE N STU DI ES
Method Case-I Case-II Case-III
k3b CSP/RCSP 14.09 2.97 2.93
JAD 1.54 0.11 0.13
l1b CSP/RCSP 9.48 2.06 2.02
JAD 1.47 0.09 0.12
k6b CSP/RCSP 9.41 2.00 2.02
JAD 1.46 0.10 0.16
Avg. 6.24 1.22 1.23
The classification accuracies are subject specific and varies
across subjects (Table III). The brain dynamics deviate across
subjects. Even, the ability to stimulate MI tasks differ from
subject to subject. However, we consider the effect of selecting
different number of channels from different brain regions. The
average classification accuracies are 64.6%, 55.3% and 61.4%
respectively for Case I,Case II and Case III. Although the
number of channels (i.e., 12 channels from the sensorimotor
areas instead of 60 channels from the whole brain) are sig-
nificantly reduced in Case III, the performance degrades quite
insignificantly. The classification accuracy for subject k3b using
CSP is 86.7% in Case III which is much higher than 78.3%
achieved in Case I. Also, the classification accuracy for subject
k6b using JAD is 42.1% in Case III which is also much
higher than classification accuracy 32.9% in Case I. Sometimes,
the classification accuracies in Case III outperform what we
achieved in Case I, may occur due to lesser effect of outliers
from reduced number of channels. So, this study evinces the
feasibility of integrating the knowledge of sensorimotor dynam-
ics towards a BCI system while employing reduced number
of channels. In Table IV, the comparison of subject specific
computational latencies are shown. The three different cases are
identical elsewhere except the number of actual EEG channels
used. The average computational latencies for three different
cases are 6.4s, 1.22s and 1.23s respectively. Certainly, Case I
requires more computation for 60 channels than for 12 channels
in Case II and in Case III. The computational complexities differ
in preprocessing phase only. We measure the computational
latencies from the input of bandpass filter to the output of the
spatial filtering. The computational latencies are observed for
10 times, i.e., 10 simulation runs and are then averaged. As
conclusion, computational cost can be reduced by employing less
number of EEG channels from specific brain areas considering
sensorimotor dynamics.
Fig. 3. Visualization of subband Energy-Entropy (sEE) in 2D plot: the
features for four classes of MI tasks which are extracted from a spatially filtered
channel are shown in different markers along with rectangles representing the
Mean±SD.
This experiment has been performed in a personal workstation
with the following configurations: CPU-2.5GHz, RAM-4GB,
Windows 7 Home Premium, MATLAB 2013b.
IV. CONCLUSION
Multi-channel EEG recordings require simultaneous process-
ing of high dimensional signals and thus cause high com-
putational needs. But, all EEG channels are not significantly
important for specific cognitive tasks. Even, undesired channels
sometimes introduce outliers and degrade the performance. In
this experiment, we consider spatial brain dynamics to select
some specific channels from the sensorimotor areas of the brain
for multi-classes of MI tasks. In contrast to other channel
selection literature, we emphasize only on concepts of brain
dynamics to select some specific channels. However, the channel
selection method using sensorimotor dynamics doesn’t remain
to be universal across sessions and subjects. Because, the brain
dynamics varies significantly across subjects and across sessions
even for a particular subject. Our experiment doesn’t necessarily
show any generic phenomenon regarding spatial brain reasoning,
rather it evinces that the integration of brain dynamics can
sometimes play important role for optimal channel selection
as complementary to advanced scientific tools. There is need
of fundamental reasoning of MI brain dynamics for localizing
cortical sources for specific tasks. Many literature show signif-
icant reasoning of MI spatial dynamics [1]-[5]. Due to inter-
subject variabilities in cognitive features, the generic localization
of optimal sources of MI tasks is difficult. In that cases, multi-
modality EEG/fMRI can be employed to capture spatial locations
of the optimal EEG channels and then to utilize EEG channels
from that areas for extracting distinguishable temporal features.
ACKNOWLEDGMENT
The authors would like to thank the organizers of BCI
Competition III for providing the dataset available for research.
REFERENCES
[1] M. Jeannerod, ”Mental imagery in the motor context”, Neuropsychologia,
vol. 33, no. 11, pp. 1419-1432, 1995.
[2] C. Zich, S. Debener, C. Kranczioch, M. Bleichner, I. Gutberlet and M. De
Vos, ”Real-time EEG feedback during simultaneous EEGfMRI identifies the
cortical signature of motor imagery”, NeuroImage, vol. 114, pp. 438-447,
2015.
[3] M. Lotze, P. Montoya, M. Erb, E. Hlsmann, H. Flor, U. Klose, N. Birbaumer
and W. Grodd, ”Activation of Cortical and Cerebellar Motor Areas during
Executed and Imagined Hand Movements: An fMRI Study”, Journal of
Cognitive Neuroscience, vol. 11, no. 5, pp. 491-501, 1999.
[4] M. Lotze and U. Halsband, ”Motor imagery”, Journal of Physiology-Paris,
vol. 99, no. 4-6, pp. 386-395, 2006.
[5] A. Solodkin, P. Hlustik, E. E. Chen, and Steven L. Small, ”Fine Modula-
tion in Network Activation during Motor Execution and Motor Imagery”,
Cerebral Cortex, vol. 14, no. 11, pp. 1246-1255, 2004.
[6] S. Saha and K. I. Ahmed, ”Efficient event related oscillatory pattern
classification for EEG based BCI utilizing spatial brain dynamics,” Electrical
and Computer Engineering (ICECE), 2014 International Conference on,
Dhaka, 2014, pp. 707-710.
[7] H. Yang, C. Guan, C. C. Wang and K. K. Ang, ”Maximum dependency and
minimum redundancy-based channel selection for motor imagery of walking
EEG signal detection,” 2013 IEEE International Conference on Acoustics,
Speech and Signal Processing, Vancouver, BC, 2013, pp. 1187-1191.
[8] Yijun Wang, Shangkai Gao and Xiaornog Gao, ”Common Spatial Pattern
Method for Channel Selelction in Motor Imagery Based Brain-computer
Interface,” 2005 IEEE Engineering in Medicine and Biology 27th Annual
Conference, Shanghai, 2005, pp. 5392-5395.
[9] M. Arvaneh, C. Guan, K. K. Ang and C. Quek, ”Optimizing the Channel
Selection and Classification Accuracy in EEG-Based BCI,” in IEEE Trans-
actions on Biomedical Engineering, vol. 58, no. 6, pp. 1865-1873, June
2011.
[10] A. Zaitcev, G. Cook, W. Liu, M. Paley and E. Milne, ”Feature extraction
for BCIs based on electromagnetic source localization and multiclass Filter
Bank Common Spatial Patterns,” 2015 37th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan,
2015, pp. 1773-1776.
[11] H. Vikram Shenoy, A. P. Vinod, and C. Guan, Cortical source local-ization
for analysing single-trial motor imagery EEG, in2015 IEEEInternational
Conference on Systems, Man, and Cybernetics (SMC)(accepted), Oct. 2015
[12] M. G. Wentrup, K. Gramann, E. Wascher and M. Buss, ”EEG Source
Localization for Brain-Computer-Interfaces,” Conference Proceedings. 2nd
International IEEE EMBS Conference on Neural Engineering, 2005., Ar-
lington, VA, 2005, pp. 128-131.
[13] S. Haufe et al., ”Localization of class-related mu-rhythm desynchronization
in motor imagery based Brain-Computer Interface sessions,” 2010 Annual
International Conference of the IEEE Engineering in Medicine and Biology,
Buenos Aires, 2010, pp. 5137-5140.
[14] Blankertz, B. (2005). BCI competition III webpage. Webpage. URL:
http://www. bbci. de/competition/iii.
[15] B. Blankertz et al., ”The BCI competition 2003: progress and perspectives
in detection and discrimination of EEG single trials,” in IEEE Transactions
on Biomedical Engineering, vol. 51, no. 6, pp. 1044-1051, June 2004.
[16] H. Lu, H. L. Eng, C. Guan, K. N. Plataniotis and A. N. Venetsanopoulos,
”Regularized Common Spatial Pattern With Aggregation for EEG Clas-
sification in Small-Sample Setting,” in IEEE Transactions on Biomedical
Engineering, vol. 57, no. 12, pp. 2936-2946, Dec. 2010.
[17] H. Ramoser, J. Muller-Gerking and G. Pfurtscheller, ”Optimal spatial
filtering of single trial EEG during imagined hand movement,” in IEEE
Transactions on Rehabilitation Engineering, vol. 8, no. 4, pp. 441-446, Dec
2000.
[18] F. Lotte and C. Guan, ”Regularizing Common Spatial Patterns to Improve
BCI Designs: Unified Theory and New Algorithms,” in IEEE Transactions
on Biomedical Engineering, vol. 58, no. 2, pp. 355-362, Feb. 2011.
[19] Xinyi Yong, R. K. Ward and G. E. Birch, ”Robust Common Spatial Patterns
for EEG signal preprocessing,” 2008 30th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, Vancouver, BC,
2008, pp. 2087-2090.
[20] M. Grosse-Wentrup* and M. Buss, ”Multiclass Common Spatial Patterns
and Information Theoretic Feature Extraction,” in IEEE Transactions on
Biomedical Engineering, vol. 55, no. 8, pp. 1991-2000, Aug. 2008.
[21] A. Ziehe, P. Laskov, G. Nolte, and K.-R. Muller, ”A Fast Algorithm for
Joint Diagonalization with Non-orthogonal Transformations and its appli-
cation to Blind Source Separation, Journal of Machine Learning Research,
Vol. 5, pp. 777-800, 2004.
