[ieee 2012 ieee embs conference on biomedical engineering and sciences (iecbes 2012) - langkawi,...
TRANSCRIPT
![Page 1: [IEEE 2012 IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES 2012) - Langkawi, Malaysia (2012.12.17-2012.12.19)] 2012 IEEE-EMBS Conference on Biomedical Engineering](https://reader037.vdocuments.net/reader037/viewer/2022092710/5750a68e1a28abcf0cba76f1/html5/thumbnails/1.jpg)
EEG Spectral Analysis for Attention State
Assessment: Graphical Versus Classical
Classification Techniques
Ahmed Fathy, Ahmed Fahmy, Mohamed ElHelw
Center for Informatics Science
Nile University
Cairo, Egypt
Seif Eldawlatly
Computer and Systems Engineering Department
Faculty of Engineering, Ain Shams University
Cairo, Egypt
Abstract— Advances in Brain-computer Interface (BCI)
technology have opened the door to assisting millions of people
worldwide with disabilities. In this work, we focus on assessing
brain attention state that could be used to selectively run an
application on a hand-held device. We examine different
classification techniques to assess brain attention state. Spectral
analysis of the recorded EEG activity was performed to compute
the Alpha band power for different subjects during attentive and
non-attentive tasks. The estimated power values were used to
train a number of classical classifiers to discriminate among the
two attention states. Results demonstrate a classification
accuracy of 70% using both individual- and multi-channel data.
We then utilize a graphical approach to assess the causal
influence among EEG electrodes for each of the two attention
states. The inferred graphical representations for each state were
used as signatures for state classification. A classification
accuracy of 83% was obtained using the graphical approach
outperforming the examined classical classifiers.
Keywords-EEG; brain-computer interface; attention state.
I. INTRODUCTION
In the last few years, tablets and smartphones have
increasingly become essential devices for many people. These
devices mainly rely on touch screens technology that can be
controlled by direct interaction between the user and the
device. While this is considered a great achievement, it
however requires that the user uses his hands to interact with
the touch screen of these devices. This neglects users with
hand disabilities who cannot use touch screen-based smart
devices that mainly rely on physical touch.
One possible way to enable this large population to use
such devices is to use brain activity to interact with smart
devices. Recent advances in non-invasive Brain-Computer
Interfaces (BCIs) have demonstrated the efficacy of
monitoring electroencephalogram (EEG) signals and
subsequently translating them into actions that represent the
user’s intentions [1]. Successful examples of BCIs include
EEG-speller systems [2], brain-controlled games [3] and
wheel-chair control [4]. However, a number of limitations
prevented BCIs from being of commercial interest in addition
to being inconvenient for most users such as the high cost of
monitoring hardware, large size of amplifiers and using wet
electrodes for EEG monitoring. Recently, a number of
portable EEG headsets have been made commercially
available at reasonable prices that overcome the
aforementioned limitations. Successful applications have been
demonstrated using these portable units such as emotion
detection [5] and brain-controlled dialing application for
mobile phones [6].
In this paper, we propose the use of a portable EEG headset for assessing attention state. This component is part of a larger system aiming at enabling people with motor disability to fully interact with touch-screen devices using EEG activity. The attention assessment component presented in this paper will be used to run an application after a brain-controlled cursor has been moved to the application icon. We capitalize on the extensive research that has been carried out in the last two decades on using spectral analysis of EEG activity to estimate different physiological and psychological states [7, 8]. To assess attention state, we estimate the power of the Alpha waves (8 – 12 Hz) in the recorded EEG activity [9]. The abundance of Alpha waves has been shown to positively correlate with being idle, whereas the lack of Alpha waves indicates attention [10] as demonstrated in Fig. 1. We report
0 1 2
Time (sec)
Non-attentive
0 1 2
Time (sec)
Attentive
Figure 1. Sample data recorded on electrode O2 during (top) the non-attentive state and (bottom) the attentive state. The figure clearly indicates the abundance of the Alpha wave (8 – 12 cycles/sec) in the non-attentive state compared to the attentive state.
978-1-4673-1666-8/12/$31.00 ©2012 IEEE
2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012
888
![Page 2: [IEEE 2012 IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES 2012) - Langkawi, Malaysia (2012.12.17-2012.12.19)] 2012 IEEE-EMBS Conference on Biomedical Engineering](https://reader037.vdocuments.net/reader037/viewer/2022092710/5750a68e1a28abcf0cba76f1/html5/thumbnails/2.jpg)
the classification accuracy among two brain states: attentive and non-attentive, using different classification and feature extraction techniques.
II. METHODS
A. Subjects and Task
EEG activity was recorded from 3 healthy adult subjects (1 female and 2 males) using the wireless Emotiv EPOC neuroheadset - the research edition (Emotiv Systems Inc., San Francisco, USA). This neuroheadset has 14 electrodes located at positions AF3, F7, F3, FC5, T7, P7, O1, O2, P8, T8, FC6, F4, F8 and AF4 according to the international 10-20 system. Recorded EEG was sampled at 128 Hz.
Subjects were asked to perform 2 tasks: attentive and non-attentive. For the attentive task, subjects were asked to play 2 concentration-based online games for ~5 minutes and fully concentrate on the game and nothing else. For the non-attentive task, subjects were asked to close their eyes and relax without thinking about anything in particular for ~5 minutes.
B. Data Pre-processing
For each channel and each of the two tasks, raw signals
recorded using the EPOC neuroheadset were first divided into
1 sec epochs. The signals recorded within each epoch were
filtered using common average reference (CAR) spatial filter
[11]. This is done by computing the mean of all channels
within the considered epoch and subtracting this mean value
from each channel
N
jjii tx
Ntxty
1
1
(1)
where ix represents the raw signal recorded on electrode i in
epoch τ, iy represents the filtered signal and N is the total
number of channels. CAR spatial filter has been demonstrated to outperform other referencing techniques such as ear referencing [11]. Filtered signals were then subsequently thresholded to eliminate blinking and muscle artifacts (mean ± 3 standard deviation). Using Fast Fourier Transform (FFT), spectral analysis of the filtered signals within each epoch for each channel was performed to extract the frequency components corresponding to the Alpha wave in the range 8-12 Hz. Power spectral analysis was then performed to estimate the Alpha band power for each epoch on each channel.
Attention assessment performed using individual channels data was compared to that performed using multi-channel data. In order to efficiently combine Alpha band power computed for individual channels, we utilized Principal Component Analysis (PCA) to project the multi-channel data into a reduced-dimensions space where the most significant features are expressed [12]. Briefly, this is done by, first, subtracting the mean of the input training data for each epoch at each electrode and subsequently computing the covariance matrix. The eigenvectors of the covariance matrix are then computed and ordered based on their corresponding eigenvalues. To classify a new test epoch, it is first projected into the principal component
space identified using the training data before applying the classifier.
C. Classical Classification Techniques
The estimated Alpha band power data was then used to
train three different classical classifiers to discriminate among
two classes: attentive versus non-attentive. For each subject,
the Alpha band power data was divided into two datasets:
training dataset (80% of the data) to learn the parameters of the
examined classifiers and test dataset (20% of the data) to
compute the corresponding classification accuracy. The
classifiers examined are
1) Linear Discriminant Analysis (LDA) Classifier
A linear discriminant function can be expressed as [12]
0wz T sws (2)
where s is the input training data (Alpha band power data in
this case), w is the weights vector, w0 is the bias and z is the
output of the discriminant functions. If z > 0, s is classified as
belonging to class C1, whereas if z < 0, s is classified as
belonging to class C2. LDA aims at projecting the input data to
a reduced dimensions space such that the separation between
the projections of different classes is maximized while
minimizing the within-class variance.
2) Naive Bayes Classifier
Naive Bayes classifier can be categorized as a probabilistic
generative model classifier that assumes complete
independence among the input features to estimate the
likelihood of the data [12]
N
n
lnl CsC1
PrPr s
(3)
where sn represents the Alpha band power for channel n in this
case. Each of the probabilities ln CsPr is assumed to follow
a Gaussian distribution 2, nnnsN whose parameters are
estimated from the training data. To classify s, the posterior
probability slCPr is computed using Bayes’ rule
s
ss
Pr
PrPrPr
lll
CCC (4)
where s is classified as belonging to class C1 if s1Pr C is
greater than s2Pr C , and as belonging to C2 otherwise.
3) Support Vector Machine (SVM)
The two aforementioned classifiers are linear classifiers. In
the case of non-linearly separable data, a transformation can
be performed to project the input data into a higher-
dimensional space in which the data is linearly separable. To
project the data into this new space, an inner-product kernel is
typically used such as the Euclidean kernel given by [13]
2
22
1exp),( jijiK ssss
(5)
2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012
889
![Page 3: [IEEE 2012 IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES 2012) - Langkawi, Malaysia (2012.12.17-2012.12.19)] 2012 IEEE-EMBS Conference on Biomedical Engineering](https://reader037.vdocuments.net/reader037/viewer/2022092710/5750a68e1a28abcf0cba76f1/html5/thumbnails/3.jpg)
where si and sj are two input data vectors, and σ2 is the kernel
variance.
Given the training set Miii t
1,
s , where si is i
th training
vector, ti is the corresponding desired response (1 for class C1
or -1 for class C2), and M is the number of epochs in the
training data, SVM attempts to find the Lagrangian multipliers
Mii 1
that maximize the objective function
M
i
M
j
jijiji
M
i
i KttQ1 11
,2
1)( ss
(6)
subject to the constraints
,...,,2,10
01
MiC
t
i
M
i
ii
(7)
where C is a positive parameter to allow SVM to identify non-
linear decision boundary.
D. Graphical Approach for Classification
An alternative classification approach to attention state
assessment that we propose here is to classify the recorded data
based on the inferred causal relationships between the recorded
electrodes. We utilize Dynamic Bayesian Networks (DBNs) to
infer such causal relationships. DBNs represent an extension to
Bayesian networks to model time-dependent causal
relationships between random variables [14].
In the spectrally analyzed EEG recordings context, a DBN
represents the causal relationships inferred from the Alpha
band power as B =<G, P>, where G is a directed acyclic
graph (DAG) and P is a set of conditional probabilities that
expresses the statistical dependence between the
simultaneously observed Alpha power (s1, s2,…., sn) [14]. Each
graph G consists of a set of nodes {vi(t)
}, where i = 1 to n, in
which each node corresponds to the Alpha band power on one
electrode si at time t, denoted by si(t)
. Each directed edge in G
indicates conditional dependence. Using DBN formulation, the
conditional probability 1:121 ,...,,Pr tt
ntt sss s
can be
factorized as a product of individual conditional probabilities
Pr(si(t)
|sπ(i)(1:t-1)
) given that the status of any variable si(t)
in a
DBN is determined by only its parents’ history, denoted by
sπ(i)(1:t-1)
. The parents’ status history is considered up to a
maximum Markov lag T
.Pr,...,,Pr1
:1:121
n
i
Tt
i
ti
Tttn
tt ssss
ss (8)
The structure of a DBN for a given dataset can be learned
using a score-based approach, where a criterion is first defined
by which an arbitrary Bayesian network structure can be
evaluated on a given dataset, then a search is carried out
through the space of all possible structures to find the graph
with the highest score [15]. DBN has been shown to
successfully discriminate different brain states at the individual
neuron level [16].
To infer attention state networks, the observed Alpha
power on each electrode was discretized to 3 uniform levels
(0, 1 and 2). The training data for each attention state was
divided to 4 datasets and DBN was then used to infer a
network for each of the 4 datasets with 2 Markov lags. To
classify the test data as attentive or non-attentive, DBN was
used to infer the corresponding networks (test networks). The
similarity between test networks and the networks inferred for
the training data was quantified by, first, representing each
inferred network as a 14 × 14 binary adjacency matrix A. Each
element A(i, j) takes the value ‘1’ if there is a connection
inferred from electrode i to electrode j and ‘0’ if there is no
connection. All adjacency matrices of the inferred networks
were vectorized and stacked together into one matrix.
Principal component analysis (PCA) was then applied to this
matrix to extract significant features from the inferred
networks by projecting the adjacency matrices into a p-
dimension network space that accounts for most of the
variance in the networks [16]. The distance D(Al, Am) between
a pair of matrices Al and Am was defined as
mlml qqAAD , (9)
where ql and qm are the projections of Al and Am in the p-
dimension network space, respectively, and ||.|| is the
Euclidean distance. The number of principal components p
was set to 2. A test network was classified as belonging to the
attention state with minimum distance between the test
network and the corresponding training networks.
III. RESULTS
We first examined the performance of each of the three
classical classifiers in discriminating among the two brain
states: attentive or non-attentive. First, we tested each
classifier on the Alpha band power computed for individual
channels. Fig. 2 illustrates the classification accuracy for each
classifier averaged across subjects (mean ± SD). As can be
seen, the three classifiers performed equally well with
maximum average accuracy obtained on channel O2 (LDA:
67.8±5.6%; Naive Bayes: 67.2±4.9%; SVM: 65.2±3.1%).
The relatively higher accuracy obtained on the occipital and
parietal electrodes (O2, P7 and P8) compared to other
electrodes is consistent with previous studies that demonstrated
the increased amplitude of the Alpha wave on the occipital and
parietal areas when idle and the diminished amplitude when
attentive [10].
AF3 AF4 F3 F4 F7 F8 FC5 FC6 O1 O2 P7 P8 T7 T80
20
40
60
80
Channel
Cla
ssific
atio
n A
ccu
racy (
%)
LDA
NB
SVM
Figure 2. Classification accuracy obtained for individual channels using
Linear Discriminant Analysis (LDA), Naive Bayes (NB) and Support Vector Machine (SVM) classifiers.
2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012
890
![Page 4: [IEEE 2012 IEEE EMBS Conference on Biomedical Engineering and Sciences (IECBES 2012) - Langkawi, Malaysia (2012.12.17-2012.12.19)] 2012 IEEE-EMBS Conference on Biomedical Engineering](https://reader037.vdocuments.net/reader037/viewer/2022092710/5750a68e1a28abcf0cba76f1/html5/thumbnails/4.jpg)
1 2 3 4 5 6 7 8 9 10 11 12 13 140
20
40
60
80
Number of PCs
Cla
ssific
atio
n A
ccu
racy (
%)
LDA
NB
SVM
Figure 3. Classification accuracy obtained for the multi-channel data for different number of principal components (PCs).
We also examined the performance of the same classifiers
when applied to multi-channel data. This was done by first
using PCA to extract the most significant features from the
input multi-channel data. Fig. 3 illustrates the performance of
each classifier for different number of principal components
(PCs). Similar to the results obtained using individual channels
data, the performance of the three examined classifiers was not
significantly different. However, the results indicate a slight
improvement in the classification accuracy compared to using
individual channels data when using 3 PCs (LDA: 70±7.7%;
Naive Bayes: 69.5±4.5%; SVM: 69.6±9%).
As an alternative approach, we used graphical
representation of the causal influence among the recording
electrodes as signatures of the attention state. Using Dynamic
Bayesian Network (DBN), we inferred causal connections for
both attentive and non-attentive task. Fig. 4 illustrates sample
networks inferred for both attention states for the training data.
The significant difference between the networks across the two
states indicates the feasibility of using this approach as means
to classify the networks inferred for the test data. A larger
average classification accuracy of 83% was obtained using
DBN compared to the other approaches as illustrated in Fig. 5.
IV. CONCLUSION
We investigated the use of different classification
techniques in assessing brain attention state. Using affordable
wireless EEG headset, we recorded EEG activity during
performing attentive and non-attentive tasks. Data recorded on
individual channels were pre-processed and subsequently used
to train each of the examined classifiers. Results demonstrate
the superiority of graphical representations to classical
classification techniques to discriminate among attentive and
non-attentive brain states with acceptable accuracy. The
examined methods can be further extended to examine other
spectral bands (such as Theta and Beta waves and the
AF3 AF4
F3 F4
F7 F8
FC5 FC6
O1 O2
P7 P8
T7 T8
Non-attentive State
AF3 AF4
F3 F4
F7 F8
FC5 FC6
O1 O2
P7 P8
T7 T8
Attentive State
Figure 4. Sample networks for non-attentive and attentive states. Each node corresponds to 1 electrode. Each directed edge indicates a causal relationship.
LDA NB SVM DBN0
20
40
60
80
100
Max. A
vg. A
ccura
cy (
%)
Figure 5. Comparing the maximum average classification accuracy of the multi-channel data (from Fig. 3) to that obtained using DBN.
corresponding Theta-Beta ratio) that have been shown to
correlate with attention. The approaches presented in this paper
will be used in the context of a brain-smart device interface to
run an application on a touch-screen device based on the
attention level of the user.
REFERENCES
[1] J. R. d. Millán, et al., "Combining brain-computer interfaces and
assistive technologies: state-of-the-art and challenges," Front. Neurosci.,
vol. 4, p. 161, 2010.
[2] E. W. Sellers and E. Donchin, "A P300-based brain–computer interface: initial tests by ALS patients," Clin Neurophysiol, vol. 117, pp. 538–548,
2006.
[3] A. Nijholt, D. O. Bos, and B. Reuderink, "Turning shortcomings into challenges: Brain-computer interfaces for games," Entertain Comput,
vol. 1, pp. 85–94, 2009.
[4] K. Tanaka, K. Matsunaga, and H. O. Wang, "Electroencephalogram-Based Control of an Electric Wheelchair," IEEE Transactions on
Robotics, vol. 21, pp. 762–766, 2005.
[5] Y. Liu, O. Sourina, and M. K. Nguyen, "Real-time EEG-based human emotion recognition and visualization," in Int. Conf. on Cyberworlds,
Singapore, 2010, pp. 262–269.
[6] A. Campbell, et al., "NeuroPhone: brain-mobile phone interface using a wireless EEG headset," in ACM. MobiHeld, New Delhi, India, 2010, pp.
3–8.
[7] B. Hamadicharef, et al., "Learning EEG-based spectral-spatial patterns for attention level measurement," in IEEE International Symposium on
Circuits and Systems (ISCAS2009), Taipei, Taiwan, 2009, pp. 1465–
1468. [8] H. Laufs, et al., "Electroencephalographic signatures of attentional and
cognitive default modes in spontaneous brain activity fluctuations at
rest," Proc. Natl. Acad. Sci. U. S. A., vol. 100, pp. 11053–11058, 2003. [9] G. Buzsáki, Rhythms of the brain New York: Oxford University Press,
2006.
[10] J. J. Foxe and A. C. Snyder, "The role of alpha-band brain oscillations as a sensory suppression mechanism during selective attention," Front.
Psychology, vol. 2, p. 154, 2011.
[11] D. J. McFarland, L. M. McCane, S. V. David, and J. R. Wolpaw, "Spatial filter selection for EEG-based communication,"
Electroencephalogr. Clin. Neurophysiol., vol. 103, pp. 386-394, 1997.
[12] C. Bishop, Pattern recognition and machine learning. New York: Springer, 2006.
[13] C. Burges, "A Tutorial on Support Vector Machines for Pattern
Recognition," Data Mining and Knowledge Discovery, vol. 2, pp. 1-47,
1998.
[14] K. Murphy, "Dynamic Bayesian Networks: Representation, Inference and Learning," PhD thesis, UC Berkeley, Computer Science Division,
2002.
[15] A. J. Hartemink, D. K. Gifford, T. Jaakkola, and R. Young, "Using graphical models and genomic expression data to statistically validate
models of genetic regulatory networks," in Pacific Symposium on
Biocomputing (PSB01), 2001, pp. 422-433. [16] S. Eldawlatly and K. G. Oweiss, "Millisecond-Timescale Local Network
Coding in the Rat Primary Somatosensory Cortex," PLoS ONE, vol. 6, p.
e21649, 2011.
2012 IEEE EMBS International Conference on Biomedical Engineering and Sciences | Langkawi | 17th - 19th December 2012
891