2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
EEG Signal Analysis for BCI Application using
Fuzzy System
Thanh Nguyen, Saeid Nahavandi, Abbas Khosravi, Douglas Creighton, and Imali Hettiarachchi
Centre for Intelligent Systems Research (CISR), Deakin University
Geelong Waurn Ponds Campus, Victoria, 3216, Australia
Email: thanh.nguyen@deakin.edu.au
Tel: +613 52278281. Fax: +613 52271046
Abstract—An approach to EEG signal classification for braincomputer interface (BCI) application using fuzzy standard
additive model is introduced in this paper. The Wilcoxon test is
employed to rank wavelet coefficients. Top ranking wavelets are
used to form a feature set that serves as inputs to the fuzzy
classifiers. Experiments are carried out using two benchmark
datasets, Ia and Ib, downloaded from the BCI competition II.
Prevalent classifiers including feedforward neural network,
support vector machine, k-nearest neighbours, ensemble
learning Adaboost and adaptive neuro-fuzzy inference system
are also implemented for comparisons. Experimental results
show the dominance of the proposed method against competing
approaches.
Keywords: Wavelet transform; fuzzy standard additive model;
Wilcoxon test; EEG signal classification; motor imagery data.
I. INTRODUCTION
EEG signal analysis to understand brain electrical activity
is an important problem of a BCI system. Constructing a
usable and reliable BCI therefore requires an accurate and
effective classification of multichannel EEG signals.
Various techniques have been introduced for EEG signal
classification in the literature from low-cost methods such as
LDA [1-4], logistic regression [5-7], k-nearest neighbour [810], to computationally expensive techniques such as support
vector machine [11-14], artificial neural networks [15-17],
and Adaboost ensemble learning [1, 18]. These methods
however face a common drawback in handling the nonlinear,
noisy, embedded outlier nature of EEG signal data.
Consequently, fuzzy logic (FL), which has been well-known
as a powerful tool for uncertainty modelling, is used for
modelling EEG signals.
Five adaptive neuro-fuzzy inference system (ANFIS)
classifiers based on the inputs derived by wavelet transform
(WT) were designed in Güler and Übeyli [19] to classify five
classes of EEG signals. The ANFIS was built using
backpropagation gradient descent method integrated with the
least squares method.
Type-2 FL systems for EEG signal classification based on
features extracted by the power spectral density estimation
with a sliding window strategy were investigated in Herman
et al. [20-21]. Type-2 FL has demonstrated greater
uncertainty handling capability and flexibility than type-1 FL
and thus provided a promising potential to address nonstationary and highly variable EEG data.
The use of fuzzy SVM (FSVM) for differentiating EEGbased left and right motor imagery with wavelet features
https://doi.org/10.1109/IJCNN.2015.7280593
obtained in two sub-bands beta and mu rhythms was proposed
in Xu et al. [22]. The FSVM classifier was shown as an
effective method for identifying different metal tasks from
EEG signals.
In another approach, Yang et al. [23] exploited ANFIS
classifier to distinguish electrical status epilepticus during
sleep (ESES) and normal EEG signals. Permutation entropy
and sample entropy of the EEG signals are fed into the
ANFIS models. ANFIS was highlighted as a potential tool to
classify the background EEG from ESES patients and normal
control subjects.
EEG signal analysis in general requires the investigation of
a feature extraction to obtain useful information from data.
Time series autoregressive (AR) models, Fourier transform
(FT), time-frequency analysis and wavelet transform (WT)
are broadly employed for exploring prominent discriminant
features [24]. AR, FT and conventional time-frequency
techniques commonly assume that EEG signal is stationary.
However, this assumption is often violated in practice.
Consequently, for non-stationary transient signals like EEG,
WT is recommended rather than AR and FT. WT provides
combined information in time-frequency domain that can
enhance the performance of EEG classification [25].
This paper proposes a technique using WT and Wilcoxon
test for EEG feature extraction. As conventional methods
such as maximum variance (MV) and Kolmogorov-Smirnov
(KS) test [26] has a shortcoming, the use of Wilcoxon
statistics for selecting wavelet has a potential to improve the
EEG signal classification performance.
The arguments of the paper are structured as follows. The
next section presents fuzzy standard additive model with tabu
search learning (tabu-FSAM) as a classifier. Section II is
devoted for the main methodology that includes a feature
extraction by WT with the Wilcoxon test. Experimental
results are presented in Section IV. Section V conveys
conclusions and future research directions.
II. FUZZY SYSTEM WITH TABU SEARCH LEARNING
A. Fuzzy Standard Additive Model (FSAM)
The FSAM system :
→
consists of if-then fuzzy
rules, which have been proved to be able to uniformly
approximate continuous and bounded measurable functions in
a compact domain [27-30]. Any type of if-part fuzzy sets,
which are denoted by , can be employed based on this
approximation theorem. The theorem also facilitates the
choice of any then-part fuzzy sets, denoted by
,
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
because the FSAM system utilizes only the centroid
and
volume of to calculate the output ( ) where ∈
is
the input vector.
∑
( )=
( )
=
∑
∑
( )
( )
Fuzzy rules in the word form “If =
then = ”
constitutes fuzzy rule patches of the form ×
×
to cover the graph of an approximand . If-part fuzzy set
is characterized by a joint set function : →[0, 1]
( )=
that factors across
input components as:
( ) … ( ). Then-part fuzzy set
is described by
a membership function :
→ [0, 1] having a volume (or
area) and centroid [29].
The FSAM output ( ) can be represented as a convex
sum of centroids of then-part fuzzy sets:
( )=∑
( )
where ( ) are considered as the convex weights:
( )=
( )
∑
( )
(a)
(b)
Fig. 1. (a) A structure of FSAM system [30]. Each input
sample is fed into each fuzzy rule that fires to some
membership degree to calculate output F(x). (b) Fuzzy rules
define fuzzy patches to cover the approximand in the inputoutput space.
Fig. 1 illustrates the parallel configuration of FSAM, which
requires an exponential explosion of fuzzy rule number to
cover the function’s graph. FSAM system normally needs
fuzzy rules to approximate the function :
→
in a compact domain [30].
https://doi.org/10.1109/IJCNN.2015.7280593
Learning is a vital process of FSAM to construct a
knowledge-based system that comprises if-then fuzzy rules.
The FSAM learning process conventionally includes two
basic steps: unsupervised learning for constructing if-then
fuzzy rules and supervised learning for tuning rule parameters
[31, 32]. We propose the use of a meta-heuristic learning
process, i.e. tabu search, to find the optimal if-then fuzzy rule
structure for classification. Details of the learning process are
presented in the following subsection.
B. Tabu Search for Fuzzy System Learning
Tabu search is a meta-heuristic algorithm that was
introduced by Glover [33]. It has emerged as a competent
technique for solving difficult optimization problems. A main
feature of tabu search is the employment of special strategies
to exploit adaptive memory. Memory-based strategies allow
tabu search to penetrate complexities that often confound
other solving approaches. They also enable the
implementation of searching procedures that are able to
explore the solution space of objective functions
economically and effectively [34, 35].
Similar to ordinary local or neighbourhood search, tabu
search begins from an initial point (solution) and proceeds
repeatedly from one point to another point until a pre-set
termination criterion is met. The basic principle is to avoid
entrainment in cycles by restricting moves that take to
previously visited points in the solution space.
We apply tabu search to optimize the rule structure of
FSAM system. Fuzzy rules are coded by binary digits where
1 means the rule is selected and 0 represents the case of
omitting the rule. This application of tabu search becomes
solving a problem where solution is a binary vector. The
optimization problem is carried out with the objective that
minimizes the error between the FSAM estimated outputs and
real values. Fig. 2 presents the tabu search pseudo code used
in our methodology.
III. TABU FUZZY SYSTEM WITH WAVELETS FEATURES FOR
EEG SIGNAL CLASSIFICATION
Fuzzy systems in general or FSAM in particular often
encounters a huge challenge in computation if there are many
inputs. High-dimensional data would decline the performance
of FSAM. Therefore, a dimension reduction or feature
extraction tool can be implemented before FSAM is executed.
This is particularly important as the EEG data are often
assembled in noisy and high-dimensional nature.
The approach to EEG signal analysis introduced in this
paper is exhibited in Fig. 3. The raw EEG signal data are
divided into different channels before processing. For each
channel, the WT is executed separately to obtain the
information contained in the EEG signals. The wavelet
coefficients are then ranked based on the Wilcoxon test
statistic to select the most discriminative coefficients of each
channel. In the next step, these selected coefficients are
combined to produce a feature set that serves as inputs to the
tabu-FSAM classifier. The following presents in detail WT
and the employment of the Wilcoxon test for wavelet
coefficient selection.
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
Inputs:
- Objective function: The objective function that uses the binary vector input
- Dim: The length of the binary solution vector
- Pre-set maximum number of switches: The number of times the algorithm
searches a better point in its direct neighbouring before the algorithm
terminates
1:
2:
3:
4:
5:
Create random starting Point that is a binary vector
Calculate its objective using Objective function
Initiate the tabu list
Set the number of switches equal to zero
Do
6: Choose a random index that is less than Dim
7: Choose a different index if it is already present in the tabu list
8: Add the index to the tabu list
9: Create New Point by changing the value of current Point at the index
from 0 to 1 or vice versa
10: Calculate New Point’s objective
11: If Objective of New Point is less than that of the current Point
12: Replace current Point by New Point: Point = New Point
13: Reset the number of switches to zero
14: End If
15: If tabu list is full (its length equal to Dim)
16: Break
17: End If
18: Increase number of switches by 1
19: While number of switches is less than its pre-set maximum
20: Return Point that is the binary solution
Fig. 2. Tabu search pseudo code
A. WT for Feature Extraction
WT represents a signal in a time-frequency fashion [36].
WT eliminates the requirement of signal stationarity that
often applies to conventional methods. Once the wavelets (the
mother wavelet) ( ) is fixed, translations and dilations of
the mother wavelet can be formed
,( , ) ∈
R × R . It is convenient to take special
Fig. 3. Tabu FSAM classifier with wavelets features
https://doi.org/10.1109/IJCNN.2015.7280593
and
such as
=
2 and = 2
where and are integers. One of the
simplest wavelets is the Haar wavelet. Haar functions can
uniformly approximate any continuous function. Dilations
( )=
and translations of the function , which is
. (2 − ), characterize an orthogonal basis in
( ). Therefore any element in ( ) may be defined as a
linear expression using these basis functions [37].
WT has been employed for a number of problems such as
those of medical data analysis, e.g. see [38-41].
We employ the four-level Haar wavelets for feature
extraction applied to processing EEG signal data. Haar
wavelets are employed because of their compact support and
orthogonality by which the discriminant features of data
samples can be characterized by a few representative wavelet
coefficients [26].
Once the transformation is completed, a procedure to select
significant wavelet coefficients that best discriminate the
different classes is performed. A conventional approach to
this procedure is the maximum variance (MV) criterion. MV
selects coefficients (features) that have greatest variance.
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
Quiroga et al. [26] argued that coefficients with the largest
variance do not necessarily show the best discrimination
among classes. Accordingly, selected coefficients should have
the largest deviation from normality for the best
discrimination. For this end, Quiroga et al. [26] suggested
using the Lilliefors modification version of a KolmogorovSmirnov (KS) statistical test. Given the dataset , the
comparison between the data cumulative distribution function
( ) with a Gaussian distribution ( ) is investigated. The
(| ( ) − ( )|) [26].
test statistic is then measured by
Nevertheless, the KS test follows an unsupervised strategy
that does not emphasize the difference or the discrimination
of the classes. It is important to note that even in a single
class, features may still present a large deviation from
normality. If this context occurs, the KS test may nominate
these features although they do not refer to the difference
among the classes. Thus, the information used by KS test may
not be appropriate to guarantee good discrimination
properties of a feature passing the test. We introduce a
method using the Wilcoxon test to select elite wavelet
coefficients for classification. Unlike the MV or KS test, the
Wilcoxon method provides information about the equality of
population locations of the classes. It involves a supervised
approach that takes into account class labels to separate
features of different classes. The following subsection
scrutinizes backgrounds of the Wilcoxon method.
B. Wilcoxon Method
Wilcoxon rank sum test is a nonparametric test, which
evaluates the equality of population locations (medians). The
null hypothesis is that two populations enclose identical
distribution functions whereas the alternative hypothesis
refers to the case two distributions differ regarding the
medians. The normality assumption regarding the differences
between the two samples is not required. That is why this test
is used instead of the two sample t-test in many applications
when the normality assumption is concerned.
The main phases of the Wilcoxon ranking test are
summarized below [42, 43]:
1) Assemble all observations of the two-class populations
and sort them in the ascending order.
2) The Wilcoxon statistic is calculated by the sum of all the
ranks corresponding with the observations from the smaller
group.
3) The hypothesis decision is made based on the p-value,
which is found in the table of Wilcoxon rank sum distribution
or using statistical packages.
In the application of the Wilcoxon test for wavelet
coefficient selection, the absolute values of the standardized
Wilcoxon statistics are employed to rank coefficients. Note
that the Haar wavelets are orthogonal. This ensures that the
higher ranking coefficients are more prominent [44].
IV. EXPERIMENTAL RESULTS
Experiments in this study are deployed using the two
widely-used Ia and Ib datasets, which are downloaded from
the BCI Competition II. The data were generated by
Birbaumer et al. [45].
https://doi.org/10.1109/IJCNN.2015.7280593
In the Ia dataset, number of training samples is 268 where
135 trials are of class “1” and 133 samples are of class “2”,
which correspond to moving a cursor up and down. Number
of testing trials of this dataset is 293.
In the Ib dataset, number of training samples is 200 where
classes “1” and “2” both have the same number of trials at
100. Number of testing samples in this dataset is 180 trials.
Fig. 4a&b exhibit noisy, embedded outliers, non-stationary
and multidimensional characteristics of the EEG signals
recorded in the Ia and Ib datasets. The first step is to separate
the EEG signals into various channels for ease of processing.
Data of individual channels show different spectra with rather
clear separation points. More specifically, Fig. 4a graphically
indicates 6 channels in the Ia dataset whereas there are visibly
7 distinct channels in the Ib dataset depicted by Fig. 4b.
(a)
(b)
Fig. 4. Plots of trials of different channels of the (a) Ia dataset
and (b) Ib dataset
After the raw EEG data are separated into different
channels, the Haar wavelet transform at level 4 is
implemented for each channel. Then some filter approaches
are applied to remove coefficients with low absolute values,
little variation, small ranges or low entropy. These
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
coefficients are generally not of interest because they have a
low potential to discriminate the classes. Taking into account
these features may enhance noise into the process.
Fig. 5a&b display the distributions of wavelet coefficients
obtained by employing WT on channel 1 of the Ia dataset and
channel 6 of the Ib dataset respectively. The original signal is
a sum of the coarse approximation component A4 and four
detail components D1-D4. Each component corresponds to a
particular frequency bandwidth. The blue triangular marks
indicate the most discriminative coefficients selected through
the statistical test based on the Wilcoxon method.
Alternatively, the blue diamond marks specify the coefficient
chosen by the KS test.
The feature set in the Ia dataset consists of 6 features
corresponding to 6 channels. Likewise, there are 7 features
resulting from 7 channels of the Ib dataset.
(a)
(a)
(b)
Fig. 5. Wavelet coefficients of (a) channel 1 of the Ia dataset
and (b) channel 6 of the Ib dataset
Distributions of the first feature (i.e., wavelet coefficient of
channel 1) of the feature set for the Ia and Ib datasets are
illustrated in Fig. 6a&b respectively. It can be seen that there
is a disturbance and vague distinction of two classes in both
datasets. Modelling these data requires powerful uncertainty
handling tools in which fuzzy logic, i.e. FSAM, is an
example.
https://doi.org/10.1109/IJCNN.2015.7280593
(b)
Fig. 6. Distribution of the first feature (Channel 1) of the (a)
Ia dataset, (b) Ib dataset
Once the feature set has been generated, one is able to
establish FSAM systems for classification. The number of
inputs is equivalent to the number of features of the training
feature set. The centres of the antecedent Gaussian fuzzy sets
are set equivalent to the values of the features. In the
consequent part, the centres of the fuzzy sets are assigned
either “1” or “2” depending on the samples representing for
class “1” or “2” correspondingly.
Accuracy, F1 score statistics (F-measure), Gini coefficient
and mutual information are metrics used to evaluate
classification performance in the experiments. The F-measure
is a single measurement of a classification method’s
usefulness. The F-measure takes into account both the
“Precision” and “Recall” of the classification procedure to
calculate the evaluating score as the harmonic mean of
“Precision” and “Recall” expressed by:
F measure 2
Precision Recall
Precision Recall
The higher the F-measure, the superior is the predictive power
of the classification technique. A score of 1 (or 100%) means
the classification procedure is perfect.
Gini coefficient (index) is an empirical measure of
performance of classification methods based on the area
under a receiver operating characteristic curve (AUC). It is a
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
linear rescaling of AUC: 2 ∗
– 1. The greater the Gini
index, the better performance is the classifier.
The mutual information (MI) between estimated and true
label is calculated by:
( ̂, )
,
=∑ ̂ ∑
( ̂, ) log ( ̂) ( )
where ( ̂ , ) is the joint probability distribution of estimated
and true class labels
and , and ( ̂) and ( ) are the
marginal probability distributions of and correspondingly
[46].
For comparisons, the following procedures: feedforward
neural network (FFNN), support vector machine (SVM), knearest neighbours (kNN), ensemble learning Adaboost and
adaptive neuro-fuzzy inference system (ANFIS) are also
executed. Table 1 & 2 present results of the proposed tabuFSAM method and the comparable techniques deployed on
the Ia and Ib datasets respectively. With nondeterministic
classifiers (i.e. FFNN, ANFIS, tabu-FSAM), average results
over 30 independent trials are presented. For these methods,
standard deviation statistics are also displayed adjacent to the
means. Note that results are all measured in percentage.
We also report here the best results of the competition on
the
two
datasets,
which
can
be
seen
at
http://www.bbci.de/competition/ii/results/index.html.
The
winners of the Ia and Ib datasets respectively were Mensh et
al. [25] and Bostanov [47].
Mensh et al. [25] used gramma-band power of EEG signals
for BCIs due to its correlation with high-level mental states.
Although most BCIs are deployed using the mu and beta
rhythms, the authors recognized that most of the meaningful
frequency information for EEG signal analysis in the Ia
dataset is in the gamma band, with basically none below 24
Hz. The discriminant analysis was utilized efficiently in this
dataset despite its linearity limitation.
Bostanov [47] in another approach used the continuous WT
and Student’s two-sample t-statistic for EEG data feature
extraction. The method performs totally automated
recognition and quantification of event-related brain potential
components in the time-scale plane. The classical linear
discriminant analysis is then employed for the classification.
Mensh et al. [25] obtained the greatest accuracy on the Ia
dataset at 88.7%. However, with the same method, the
authors were just able to obtain 43.9% accuracy on the Ib
dataset. On the other hand, the method of Bostanov [47]
derived the best performance on the Ib dataset with the
accuracy at 54.4%. This method however could just produce
the accuracy at 82.6% on the Ia dataset. These statistics reveal
the fact that none of the two competition-winner methods
performs effectively on both Ia and Ib datasets.
In contrast, the proposed tabu-FSAM clearly outperforms
both of the competition-winner methods in both datasets.
Tabu-FSAM obtains 90.20% and 57.28% accuracy in the Ia
and Ib dataset respectively.
Table 1. Performance on the Ia dataset
F-measure
Gini index
MI
Accuracy
The best result of competition by Mensh et al. (2004)
88.70
The accuracy obtained by Bostanov (2004)
82.60
SVM
85.71
68.67
40.42
84.30
kNN
81.57
67.83
41.34
83.96
Adaboost
83.97
65.92
35.19
82.94
FFNN
80.17 (± 5.65)
62.49 (± 9.90)
32.47 (± 9.81)
81.26 (± 4.95)
ANFIS
86.91 (± 1.86)
73.88 (± 3.97)
44.35 (± 5.08)
86.94 (± 1.99)
Tabu-FSAM
90.07 (± 0.42)
80.40 (± 0.81)
53.82 (± 1.27)
90.20 (± 0.40)
Table 2. Performance on the Ib dataset
F-measure
Gini index
MI
Accuracy
The best result of competition by Bostanov (2004)
54.40
The accuracy obtained by Mensh et al. (2004)
43.90
SVM
54.84
6.67
0.32
53.33
kNN
56.22
10.00
0.72
55.00
Adaboost
55.14
7.78
0.44
53.89
FFNN
55.42 (± 3.64)
8.07 (± 5.60)
0.70 (± 0.66)
54.04 (± 2.80)
ANFIS
57.49 (± 1.71)
10.70 (± 3.16)
0.91 (± 0.48)
55.35 (± 1.58)
Tabu-FSAM
57.53 (± 1.17)
14.56 (± 2.13)
1.57 (± 0.45)
57.28 (± 1.07)
https://doi.org/10.1109/IJCNN.2015.7280593
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
(a)
(b)
Fig. 7. Accuracy obtained by features selected by KS test and Wilcoxon test in (a) Ia dataset, (b) Ib dataset
The comparisons among the classifiers also highlight the
advantage of the proposed tabu-FSAM against the
competitive classifiers. The dominance of tabu-FSAM is
shown not only in accuracy but also in other performance
measures, i.e. F-measure, Gini index and MI. Among
nondeterministic classifiers, tabu-FSAM also yields more
stable results with smaller standard deviations compared to
FFNN and ANFIS (see Table 1 & 2).
Fig. 7 presents the accuracy comparisons when performing
classifiers using features selected by KS test and Wilcoxon
test. Noticeably, the performance of applications of the
Wilcoxon features is superior to those of the KS features
through all classifiers. This is understandable as the KS test
selects features without a reference to the class labels (i.e., an
unsupervised approach). Inversely, the Wilcoxon method
takes into account the class labels in examining the features
and also provides the information about the equality of
population locations of the classes so that it is a more efficient
feature selection.
V. CONCLUSIONS
A method for EEG data classification using tabu-FSAM
has been introduced in this paper. The noisy, nonlinear and
outlier-embedded nature of EEG signals is modelled
proficiently by FSAM whose if-then fuzzy rule structure is
optimized by the tabu search algorithm. The paper also
presents an approach to supervised EEG signal feature
extraction based on WT decomposition and the Wilcoxon
statistics. Well-known methods such as MV and KS tests for
selecting discriminative wavelet coefficients are used to
examine features in an unsupervised approach without
reference to the class labels. This does not guarantee the
separability of the feature set. The proposed method in this
study suggests using the Wilcoxon test, which performs in a
supervised strategy. The Wilcoxon test separates data samples
according to the class labels and select features by evaluating
the population locations of the classes.
Experimental results on two benchmark datasets
downloaded from the BCI competition II demonstrate the
superiority of the Wilcoxon feature selection against the KS
test approach. The tabu-FSAM designed for classification
also shows great performance dominance to other comparable
https://doi.org/10.1109/IJCNN.2015.7280593
classifiers, including FFNN, SVM, kNN, Adaboost and
ANFIS. More noticeably, the tabu-FSAM in combination
with wavelets outperforms the two winning methods of the
BCI competition II in both benchmark Ia and Ib datasets by
1.50% and 2.88% respectively.
As feature extraction plays an essential role in determining
the classification performance, future research would
investigate alternative feature extraction methods. Wavelet
packet transform (WPT) is an example where it yields a
broader range of potentials for signal analysis than WT. WPT
allows wavelet detail components to be decomposed to
acquire further its approximation and detail information
components. The higher frequency components, which may
store important information of the signal, therefore can be
examined in the WPT. On the other hand, BCI behaviours
may involve multiple actions that cause the classification task
to be more complicated. As the present study focuses on the
capability of tabu-FSAM for a binary classification, designing
fuzzy systems for multi-class problems would be worth a
further exploration.
ACKNOWLEDGMENT
This research is supported by the Australian Research
Council (Discovery Grant DP120102112) and the Centre for
Intelligent Systems Research (CISR) at Deakin University.
[1]
[2]
[3]
[4]
[5]
REFERENCES
Sabeti, M., Katebi, S. D., Boostani, R., & Price, G. W. (2011). A new
approach for EEG signal classification of schizophrenic and control
participants. Expert Systems with Applications, 38(3), 2063-2071.
Vidaurre, C., Kawanabe, M., von Bunau, P., Blankertz, B., & Muller,
K. R. (2011). Toward unsupervised adaptation of LDA for brain–
computer
interfaces. IEEE
Transactions
on
Biomedical
Engineering, 58(3), 587-597.
Zhang, R., Xu, P., Guo, L., Zhang, Y., Li, P., & Yao, D. (2013). Zscore linear discriminant analysis for EEG based brain-computer
interfaces. PloS One, 8(9), e74433.
Chen, M., Fang, Y., & Zheng, X. (2014). Phase space reconstruction
for improving the classification of single trial EEG. Biomedical Signal
Processing and Control, 11, 10-16.
Hosseinifard, B., Moradi, M. H., & Rostami, R. (2013). Classifying
depression patients and normal subjects using machine learning
techniques and nonlinear features from EEG signal. Computer
Methods and Programs in Biomedicine, 109(3), 339-345.
2015 International Joint Conference on Neural Networks (IJCNN), Killarney, Ireland, pp. 1-8.
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
Prasad, P., Halahalli, H., John, J., & Majumdar, K. (2014). Single-trial
EEG classification using logistic regression based on ensemble
synchronization. IEEE Journal of Biomedical and Health Informatics,
18(3), 2014.
Li, Y., & Wen, P. P. (2014). Modified CC-LR algorithm with three
diverse feature sets for motor imagery tasks classification in EEG
based brain computer interface. Computer Methods and Programs in
Biomedicine, 113(3), pp. 767-780.
Wang, D., Miao, D., & Xie, C. (2011). Best basis-based wavelet
packet entropy feature extraction and hierarchical EEG classification
for epileptic detection. Expert Systems with Applications, 38(11),
14314-14320.
Guo, L., Rivero, D., Dorado, J., Munteanu, C. R., & Pazos, A. (2011).
Automatic feature extraction using genetic programming: An
application to epileptic EEG classification. Expert Systems with
Applications, 38(8), 10425-10436.
Acharya, U. R., Molinari, F., Sree, S. V., Chattopadhyay, S., Ng, K.
H., & Suri, J. S. (2012). Automated diagnosis of epileptic EEG using
entropies. Biomedical Signal Processing and Control, 7(4), 401-408.
Garrett, D., Peterson, D. A., Anderson, C. W., & Thaut, M. H. (2003).
Comparison of linear, nonlinear, and feature selection methods for
EEG signal classification. IEEE Transactions on Neural Systems and
Rehabilitation Engineering, 11(2), 141-144.
Li, Y., & Wen, P. P. (2011). Clustering technique-based least square
support vector machine for EEG signal classification. Computer
Methods and Programs in Biomedicine, 104(3), 358-372.
Vatankhah, M., Asadpour, V., & Fazel-Rezai, R. (2013). Perceptual
pain classification using ANFIS adapted RBF kernel support vector
machine for therapeutic usage. Applied Soft Computing, 13(5), 25372546.
Joshi, V., Pachori, R. B., & Vijesh, A. (2014). Classification of ictal
and
seizure-free
EEG
signals
using
fractional
linear
prediction. Biomedical Signal Processing and Control, 9, 1-5.
Ting, W., Guo-zheng, Y., Bang-hua, Y., & Hong, S. (2008). EEG
feature extraction based on wavelet packet decomposition for brain
computer interface. Measurement, 41(6), 618-625.
Guo, L., Rivero, D., Dorado, J., Rabunal, J. R., & Pazos, A. (2010).
Automatic epileptic seizure detection in EEGs based on line length
feature and artificial neural networks. Journal of Neuroscience
Methods, 191(1), 101-109.
Özbay, Y., Ceylan, R., & Karlik, B. (2011). Integration of type-2 fuzzy
clustering and wavelet transform in a neural network based ECG
classifier. Expert Systems with Applications, 38(1), 1004-1010.
Ahangi, A., Karamnejad, M., Mohammadi, N., Ebrahimpour, R., &
Bagheri, N. (2013). Multiple classifier system for EEG signal
classification with application to brain–computer interfaces. Neural
Computing and Applications, 23(5), 1319-1327.
Güler, I., & Übeyli, E. D. (2005). Adaptive neuro-fuzzy inference
system for classification of EEG signals using wavelet
coefficients. Journal of Neuroscience Methods, 148(2), 113-121.
Herman, P., Prasad, G., & McGinnity, T. M. (2008). Design and online evaluation of type-2 fuzzy logic system-based framework for
handling uncertainties in BCI classification. In 30th Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBS), pp. 4242-4245.
Herman, P., Prasad, G., & McGinnity, T. M. (2008). Designing a
robust type-2 fuzzy logic classifier for non-stationary systems with
application in brain-computer interfacing. In IEEE International
Conference on Systems, Man and Cybernetics (SMC), pp. 1343-1349.
Xu, Q., Zhou, H., Wang, Y., & Huang, J. (2009). Fuzzy support vector
machine for classification of EEG signals using wavelet-based
features. Medical Engineering & Physics, 31(7), 858-865.
Yang, Z., Wang, Y., & Ouyang, G. (2014). Adaptive neuro-fuzzy
inference system for classification of background EEG signals from
ESES patients and controls. The Scientific World Journal, 2014, 1-8.
McFarland, D. J., Anderson, C. W., Muller, K., Schlogl, A., &
Krusienski, D. J. (2006). BCI meeting 2005-workshop on BCI signal
processing: feature extraction and translation. IEEE Transactions on
Neural Systems and Rehabilitation Engineering, 14(2), 135.
Mensh, B. D., Werfel, J., & Seung, H. S. (2004). BCI competition
2003-data set Ia: combining gamma-band power with slow cortical
potentials
to
improve
single-trial
classification
of
https://doi.org/10.1109/IJCNN.2015.7280593
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
electroencephalographic signals. IEEE Transactions on Biomedical
Engineering, 51(6), 1052-1056.
Quiroga, R. Q., Nadasdy, Z., & Ben-Shaul, Y. (2004). Unsupervised
spike detection and sorting with wavelets and superparamagnetic
clustering. Neural Computation, 16(8), 1661-1687.
Kosko, B., (1994). Fuzzy systems as universal approximators. IEEE
Transactions on Computers, 43(11), 1329-1333.
Kosko, B., (1996). Fuzzy Engineering, Prentice Hall.
Nguyen, T., Khosravi, A., Creighton, D., & Nahavandi, S. (2015).
Classification of healthcare data using genetic fuzzy logic system and
wavelets. Expert Systems with Applications, 42(4), 2184-2197.
Mitaim, S. and Kosko, B., (1996). What is the best shape for a fuzzy
set in function approximation?. Proc. 5th IEEE Int. Conf. Fuzzy
Systems (FUZZ-IEEE-96), 2, 1237-1243.
Dickerson, J. A., & Kosko, B. (1996). Fuzzy function approximation
with ellipsoidal rules. IEEE Transactions on Systems, Man, and
Cybernetics, Part B: Cybernetics, 26(4), 542-560.
Mitaim, S. and Kosko, B. (2001). The shape of fuzzy sets in adaptive
function approximation. IEEE Transactions on Fuzzy Systems, 9(4),
637-656.
Glover, F. (1986). Future paths for integer programming and links to
artificial intelligence. Computers and Operations Research, 13, 533549.
Glover, F. (1989). Tabu search, Part I. ORSA Journal on Computing,
1(3), 190-206.
Glover, F., Laguna, M. (1997). Tabu Search, Kluwer Academic
Publishers, Boston.
DeVore, R. A., & Lucier, B. J. (1992). Wavelets. Acta Numerica, 1(1),
1-56.
Nguyen, T., Khosravi, A., Creighton, D., & Nahavandi, S. (2015).
EEG signal classification for BCI applications by wavelets and
interval type-2 fuzzy logic systems. Expert Systems with Applications,
42(9), 4370-4380.
Li, D., Pedrycz, W., & Pizzi, N. J. (2005). Fuzzy wavelet packet based
feature extraction method and its application to biomedical signal
classification. IEEE Transactions on Biomedical Engineering, 52(6),
1132-1139.
Bozhokin, S. V., & Suslova, I. B. (2014). Analysis of non-stationary
HRV as a frequency modulated signal by double continuous wavelet
transformation method. Biomedical Signal Processing and Control,
10, 34-40.
Tan, Y., Li, G., Duan, H., & Li, C. (2014). Enhancement of medical
image details via wavelet homomorphic filtering transform. Journal of
Intelligent Systems, 23(1), 83-94.
Torbati, N., Ayatollahi, A., Kermani, A. (2014). An efficient neural
network based method for medical image segmentation. Computers in
Biology and Medicine, 44, 76-87.
Deng, L., Pei, J., Ma, J., & Lee, D. L. (2004). A rank sum test method
for informative gene discovery. Proceedings of the Tenth ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, pp. 410-419.
Lehmann, E. L., & D'Abrera, H. J. (2006). Nonparametrics: Statistical
Methods Based on Ranks. New York: Springer.
Nguyen, T., Khosravi, A., Creighton, D., & Nahavandi, S. (2015).
EEG data classification using wavelet features selected by Wilcoxon
statistics. Neural Computing and Applications, In press, doi:
10.1007/s00521-014-1802-y.
Birbaumer, N., Flor, H., Ghanayim, N., Hinterberger, T., Iverson, I.,
Taub, E., Kotchoubey, B., Kübler, A., & Perelmouter, J. (1999). A
brain-controlled spelling device for the completely paralyzed. Nature,
398, 297-298.
Piedra-Fernández, J. A., Cantón-Garbín, M., & Wang, J. Z. (2010).
Feature selection in AVHRR ocean satellite images by means of filter
methods. IEEE Transactions on Geoscience and Remote Sensing,
48(12), 4193-4203.
Bostanov, V. (2004). BCI competition 2003-data sets Ib and IIb:
feature extraction from event-related brain potentials with the
continuous wavelet transform and the t-value scalogram. IEEE
Transactions on Biomedical Engineering, 51(6), 1057-1061.