Upload
mrh
View
214
Download
2
Embed Size (px)
Citation preview
Abstract—EEG-Based mental task classification is an
approach to understand the processes in our brain which lead
to our thoughts and behavior. Different mental tasks have been
used for this purpose and we have chosen relaxation and
imagination for our study. As well as normal conscious state,
we have considered mental tasks performed in hypnosis which
is defined as a state of consciousness with high concentration.
To assess nonlinear dynamics, we have considered fractal
dimension in addition to frequency features. HMM classifiers
have been used for classification. Results show the most
important features in EEG signal related to mentioned mental
tasks as well as differences between normal and hypnotic states
of the brain.
I. INTRODUCTION
T has been shown that people can control their brain’s
rhythms by performing a specific mental task. Therefore,
EEG-based mental task studies have been selected as an
approach to understand brain activity. EEG signal is
recorded and after preprocessing and feature extraction it is
classified to a number of predefined classes of mental tasks
which are mostly movement imagery, relaxation, complex
problem solving, mental letter composing, visual counting,
geometric figure rotation [1]. Study of mental tasks' specific
EEG pattern and provided models can help us in realizing
different activities of brain's subsystems and their
interaction.
Hypnosis is an altered state of consciousness which is
with high concentrated attention [2]. Based on clinical
observations, different theories have been suggested about
hypnosis. In addition, hypnosis has been studied using EEG
signal processing in works such as classification of mental
tasks as a hypnotic one or a normal one [3], or assessing the
depth of hypnosis and hypnotizability of subjects [4].
High concentration gained through hypnosis can remove
problems such as distraction from performing the asked
mental task. To find out how hypnosis can help us in mental
task studies, we have compared mental task classification in
hypnosis and normal states. Classification of two mental
Manuscript received May 2, 2005; revised Jul 15, 2005. This work was
supported in part by the Neuropsychiatry Research, Research Center
for Intelligent Signal Processing, Tehran, Iran.
S. Solhjoo is with the Biomedical Engineering Faculty, Amirkabir
University of Technology, Tehran, Iran (corresponding author, phone: +98-
9173134754; fax: +98-21-8063547; e-mail: [email protected]).
A. Motie Nasrabadi is with the Biomedical Engineering Department,
Engineering Faculty, Shahed University, Tehran, Iran (e-mail:
M.R. Hashemi Golpayegani is with the Biomedical Engineering Faculty,
Amirkabir University of Technology, Tehran, Iran (e-mail:
tasks has been studied: 1. Relaxation, 2. Imagination.
Frequency and chaotic features have been extracted from
EEG signal, while modeling and classifying of raw signal
have been studied as well. HMM (Hidden Markov Model)
-based classifiers have been used for classification purpose.
II. DATA COLLECTION AND PROCESSING
A. Subjects and Mental Tasks
We have used EEG signals recorded during hypnotism
sessions in RCISP by Abootalebi et al. [3]. EEG signals we
have used were of 5 men with hypnotizability scores of 4 or
5 according to SHCS1 standard test, i.e. high hypnotizability.
All subjects were right-handed and right-eared (with right
hand and ear dominant). Both tasks of relaxation and
imagination were performed by the subjects in both normal
and hypnotic states. The subject’s eyes were closed, and he
was relaxed on a comfortable chair during the recordings.
First step was relaxation for 2 minutes. Then in imagination
task, the subject was asked to imagine himself in front of a
blackboard in a classroom. Then he should draw a circle on
the blackboard and write the number “1” in it. Then it was
erasing that number and writing the next number repeatedly,
in an increasing manner, for two minutes. After these two
steps, hypnotism process was started according to SHCS.
Relaxation and the same imagination were performed in
hypnosis as well. Each hypnotic task was performed by
subjects in a period of 1 or 2 minutes [3].
TABLE I: AGE AND SHCS HYPNOTIZABILITY SCORES OF SUBJECTS [3]
Subject HH1 NDR1 RNP1 ASZ1 MRJ1
Age 36 32 23 30 32
Hypnotizability
Score5 4 5 4 4
B. Data Acquisition and Preprocessing
Electrodes were placed according to 10-20 standard. EEG
signal was sampled at 256 Hz and filtered using an elliptic
filter with band-width of 0.5-30 Hz and then down-sampled
to 128 Hz.
For each subject, a 1-minute segment of each task was
selected to be processed. Each 1-minute segment was
divided into 12 5-second segments. Segments with eye
blinking and movement artifacts were deleted manually.
1 Stanford Hypnotic Clinical Scale
EEG-Based Mental Task Classification in Hypnotized and Normal
Subjects
Soroosh Solhjoo, Ali Motie Nasrabadi, Mohammad Reza Hashemi Golpayegani
I
Proceedings of the 2005 IEEEEngineering in Medicine and Biology 27th Annual ConferenceShanghai, China, September 1-4, 2005
0-7803-8740-6/05/$20.00 ©2005 IEEE. 2041
Against the movement imagery mental tasks, that the
involving brain segments are almost specified, in the tasks
mentioned here, there is no especial segment of brain
specified to be responsible. Therefore, while data just from 2
or 3 EEG channels are enough to be considered in motor
imagery classification [5], data from all channels were used
in this study. We have applied a surface Laplacian (SL)
filter to reveal the actual signal produced only by local
sources below each electrode [6]. Doing this we decrease
data to signals only from 6 SL channels.
C. Extracted Features
Using raw EEG signal and trying to classify the
mentioned mental tasks with HMM classifiers led to low
percentage of correct classification for most of subjects in
hypnosis state, while results in normal state were
satisfactory. For only one of the subjects (MRJ1), we
reached good result (100% classification accuracy) in
hypnosis. So we turn to using frequency and chaotic features
for our classification purpose.
Parameters that represent chaotic behavior may be divided
to two categories. The First category indicates dynamic
behavior. Maximum Lyapunov exponent (MLE) is of this
category. These parameters state how the system behaves on
the nearby trajectories. The second category emphasizes the
geometric property of basin of attraction. Fractal dimension
(FD) is of this category. These dimension show geometrical
property of attractors and is also computed very fast [4].
Our goal was to associate each 5-second segment data as
a trial to its corresponding class. To do this, features were
extracted form each 1-second segment with 50% overlap,
and sequence of 9 extracted features was considered as the
feature vector of a 5-second segment, which was to be
modeled and classified.
Features were as follows:
1. Power spectral density components;
2. Petrosian Fractal Dimension;
3. Higuchi Fractal Dimension.
Power spectral density components of each SL channel
for frequencies 2 to 30 Hz with the frequency resolution of 2
Hz (15 feature for each of 6 SL channel) are extracted.
Segments of 0.5 second were averaged. A Hamming
window was applied to each segment, and the overlapping
between the segments was 50% [5].
Petrosian’s algorithm uses a quick estimate of the FD:
)*4.0
(loglog
log
1010
10
Nn
nn
nFD
where n is the length of the sequence, and N is the
number of sign changes in derivative of the signal [4].
In Higuchi’s algorithm, k new time series are constructed
from the signal x(1), x(2),…, x(N) under study:
)(),...,2(),(),( kk
mNmxkmxkmxmxx k
m
where m = 1, 2, …, k and k indicate the initial time value,
and the discrete time interval between points, respectively.
For each of the k time seriesk
mx , the length Lm(k) is
computed by:
kk
mN
Nkimxikmx
kL im
)1())1(()(
)( 1
where N is the total length of the signal x. An average
length is computed as the mean of the k lengths )(kLm (for
m = 1, 2, …, k). This procedure is repeated for each k
ranging from 1 to kmax, obtaining an average length for
each k. In the curve of ))(ln( kL versus )/1ln( k , the slope of
the best fitted line to this curve is the estimate of the fractal
dimension [4].
III. DATA ANALYSIS AND RESULTS
A. Classifiers
Classifying 1-second segments of data using different
features and classifiers led to low classification accuracy. So
we considered the temporal trend of EEG signal and
segments of 5-second length were to be classified.
We have used HMM classifiers. HMM models with
different number of states (ranging from 1 to 8) and different
number of Gaussian mixtures for each state (ranging from 1
to 8) were used to model feature vector sequences.
Classification was done according to log-likelihood
measure. Results are the best classification accuracy
percentage of different model structures.
According to low number of data segments (12 5-second
segments for each task/subject), we used LOO (Leave One
Out) cross validation method [7]. In this method, the
training and test sessions repeat to the number of data
segments available, and then classification accuracy
percentage is specified. In each session, one segment is left
out as the test data, and other segments are considered as the
training data.
For some of train and test sessions we did LOO method
for 5 times to test our classifiers for effect of initial
conditions. Low and in most of cases no variance of results
showed that their performance is not dependent on this
factor.
For each trial (5-second segment) there is a sequence of 9
feature vectors. Using HMM classifiers, this data is used as
a sequence of data with length 9 and dimension 90 (in the
case of the frequency features).
2042
B. Results
At first we considered HMM classifiers applied to
frequency features. Results are in a good range comparing to
our previous studies on raw signal data, but dimension of
feature vectors (90) is large which leads to large and
complex HMM models that need long time to be trained.
TABLE II: MENTAL TASK CLASSIFICATION ACCURACY ACCORDING TO
FREQUENCY FEATURES USING HMM CLASSIFIER FOR 3 SUBJECTS
Subjects Hypnosis Normal
HH1 83.33 % 87.50 %
NDR1 68.18 % 100 %
RNP1 87.50 % 100 %
Decreasing the length of feature vector did not lead to
better results, so we turned to test other features.
To consider nonlinear dynamics of EEG signal, we
selected fractal dimension. Fractal dimension was calculated
for each 1-second segment for each of 6 SL channels,
providing a feature vector of dimension 6. Sequence of 9
feature vector for a 5-second interval was to be modeled and
classified.
At first, we considered fractal dimension calculated based
on Petrosian algorithm. In this case, we did not get good
results as they can found in TABLE III. In some cases
results are worse than using of raw data.
TABLE III: MENTAL TASK CLASSIFICATION ACCURACY ACCORDING TO
PETROSIAN FRACTAL DIMENSION USING HMM CLASSIFIER FOR 3 SUBJECTS
Subjects Hypnosis Normal
HH1 68.18 % 75.83 %
NDR1 77.27 % 58.33 %
MRJ1 95.00 % 53.64 %
Considering Higuchi fractal dimension instead, showed
very good results as it can be seen in TABLE IV.
TABLE IV: MENTAL TASK CLASSIFICATION ACCURACY ACCORDING TO
HIGUCHI FRACTAL DIMENSION USING HMM CLASSIFIER FOR 4 SUBJECTS
Subjects Hypnosis Normal
HH1 86.36 % 95.83 %
NDR1 72.73 % 79.17 %
RNP1 91.67 % 100 %
ASZ1 68.18 % 95.83 %
Considering low correct classification percentage for
classifying of hypnotic mental tasks of subject ASZ1, we
used asymmetry signal of Laplacian channels and then
Higuchi fractal dimension was extracted from that signal.
Asymmetry signal was produced as (R-L)/(R+L), where R
and L are EEG signal from right and left Laplacian
channels, respectively [1, 8]. Sequence of 9 Higuchi fractal
dimensions of these three new signals are applied to
classifiers. Results of using such a feature are shown in
TABLE V, for classification of two data sets of hypnotic
mental tasks done by HH1 and ASZ1.
As it can be seen there is a very high increase in
performance of HMM classifier for ASZ1 (20%), while
there is a very low decrease for HH1 (5%).TABLE V: COMPARISON OF MAXIMUM CLASSIFICATION ACCURACY
PERCENTAGE, DECREASING THE NUMBER OF FEATURES USING
ASYMMETRY DATA
Subjects Higuchi Higuchi_Asymmetry
HH1_Hypnosis 86.36 % 81.82 %
ASZ1_Hypnosis 68.18 % 87.50 %
IV. DISCUSSION AND CONCLUSION
We have applied HMM classifiers to some different
frequency and chaotic feature vectors, to classify mental
tasks performed in normal and hypnotic states. HMM
classifiers with different number of states and Gaussian
mixtures were studied.
It has been shown that considering temporal trend of data
and long term study of EEG signals leads to better results.
To reach good results, we turn to use fractal dimension
features.
While Petrosian fractal dimension led to worse results
compared to raw data studies, using Higuchi fractal
dimension led to very good results, as we saw about 20%
increase in accuracy percentage in classification of hypnotic
and normal tasks of HH1 using Higuchi instead of Petrosian
fractal dimension. As well such an increase can be seen in
NDR1 normal mental task classification, while there is only
4.5% decrease in hypnotic mental task classification of that
subject.
The best results were of Higuchi fractal dimension
extracted from 3 asymmetry channels’ signals which were
produced from 6 Laplacian channels’ signals.
REFERENCES
[1] Z. A. Keirn, Jorge I. Aunon, "A new mode of communication between
man and his surroundings," IEEE Transaction on Biomedical
Engineering, vol. 37, No. 12, December 1990
[2] J.H. Gruzelier, "A Working Model of the Neurophysiology of
Hypnotic Relaxation," INABIS '98, 5th Internet World Congress for
Biomedical Sciences at McMaster University, Canada, Dec 7-16th.
Invited Symposium. Available at URL
http://www.mcmaster.ca/inabis98/woody/gruzelier0814/index.html
[3] V. Abootalebi, "Higher Order Spectra Study of EEG Signal to Assess
Hypnotizability," M.Sc. Thesis Report, Bioelectric Department,
Electrical Engineering Faculty, Sharif University of Technology, Jan.
2001
[4] A.M. Nasrabadi , M.R. Hashemi Golpaygani, M.A. Khalilzadeh, A.
Sharifi, “Comparison between Linear and Nonlinear EEG Signal
Processing during Different Mental Activities,” Scientific Journal of
Amirkabir, No. A55, pp. 592-600, Summer 2003
[5] S. Solhjoo, M.H. Moradi, "Mental Task Recognition: A comparison
between some of classification methods," Proceedings of
BIOSIGNAL 2004 International EURASIP Conference, pp. 24-26,
2004
[6] J. Mourino et al., "Spatial filtering in the training process of a brain
computer interface," Proceeding of the 23rd Annual EMBS
International Conference of IEEE, Vol.2 pp.639-642, 2001
[7] J.Larsen, C. Guotte, “On Optimal Data Split for generalization
estimation and model selection,” Proceeding of the Neural Networks
for Signal Processing IX, pp. 23-25, Aug. 1999
[8] H. Ehrlichman and M. S. Wiener, "EEG Asymmetry during covert
mental activity," Psychophysics., vol. 17, pp. 228-235, 1980
2043