Australas Phys Eng Sci Med (2015) 38:139–149

DOI 10.1007/s13246-015-0333-x

SCIENTIFIC PAPER

Feature extraction and classification for EEG signals using

wavelet transform and machine learning techniques
Hafeez Ullah Amin • Aamir Saeed Malik • Rana Fayyaz Ahmad •

Nasreen Badruddin • Nidal Kamel • Muhammad Hussain •

Weng-Tink Chooi

Received: 10 June 2014 / Accepted: 29 January 2015 / Published online: 4 February 2015
 Australasian College of Physical Scientists and Engineers in Medicine 2015

Abstract This paper describes a discrete wavelet trans- coefficients (D4), which represent the frequency range of
form-based feature extraction scheme for the classification 0.53–3.06 and 3.06–6.12 Hz, respectively. The findings of
of EEG signals. In this scheme, the discrete wavelet trans- this study demonstrated that the proposed feature extraction
form is applied on EEG signals and the relative wavelet approach has the potential to classify the EEG signals
energy is calculated in terms of detailed coefficients and the recorded during a complex cognitive task by achieving a
approximation coefficients of the last decomposition level. high accuracy rate.
The extracted relative wavelet energy features are passed to
classifiers for the classification purpose. The EEG dataset Keywords Discrete wavelet transform (DWT)  Machine
employed for the validation of the proposed method con- learning classifiers  Electroencephalography (EEG) 
sisted of two classes: (1) the EEG signals recorded during Cognitive task
the complex cognitive task—Raven’s advance progressive
metric test and (2) the EEG signals recorded in rest condi-
tion—eyes open. The performance of four different classi- Introduction
fiers was evaluated with four performance measures, i.e.,
accuracy, sensitivity, specificity and precision values. The Electroencephalography (EEG) is a neuroimaging tech-
accuracy was achieved above 98 % by the support vector nique for recording the brain’s electrical potentials, which
machine, multi-layer perceptron and the K-nearest neighbor are commonly used to study the dynamics of neural
classifiers with approximation (A4) and detailed information processing in the brain, and diagnose brain
disorders and cognitive processes. Large amounts of EEG
data are recorded and it is not possible to analyze EEG data
H. U. Amin  A. S. Malik (&)  R. F. Ahmad  N. Badruddin  visually [1]. Therefore, there is a strong demand to extract
N. Kamel relevant information from EEG recordings for the proper
Department of Electrical & Electronic Engineering, Centre for
evaluation and understanding of the desired cognitive
Intelligent Signal & Imaging Research (CISIR), Universiti
Teknologi PETRONAS, Bandar Seri Iskandar, 31750 Tronoh, processes. The main steps in the process of extracting
Perak, Malaysia relevant information from EEG recordings include pre-
e-mail: [email protected] processing, feature extraction and classification [2].
H. U. Amin Extracting relevant features is among the most critical and
e-mail: [email protected] significant steps for EEG data classification. The reason
behind this is that the feature extraction step has a direct
M. Hussain
Department of Computer Science, College of Computer and impact on the systems classification performance [3]. If the
Information Sciences, King Saud University, Riyadh 12372, extracted features are not expressive for a certain problem,
Saudi Arabia then the classification performance will not be satisfactory.
In such a case, the classification method may be highly
W.-T. Chooi
Advanced Medical and Dental Institute (AMDI), Universiti optimal for the problem but due to inadequate features, the
Sains Malaysia, Kepala Batas 13200, Penang, Malaysia method may not provide good classification results. Hence,

123
140 Australas Phys Eng Sci Med (2015) 38:139–149

extracting suitable features from EEG signals to get high Related work
classification performance is mandatory.
Recently, a multi-disciplinary research area—brain In literature, the time domain, frequency domain, and
computer interface (BCI), involving researchers from wavelet-based feature extraction techniques for classifica-
neuropsychology, engineering, computer science, mathe- tion of EEG signals have been reported [9–11]. These
matics and neuroscience attracted a lot of interest as it has techniques use the time and frequency domain features in
the potential to provide control capabilities and commu- the classification models to determine the optimal feature
nication to people facing severe motor disabilities. BCI is a set and combine with classifiers that gives the highest
system that enables a physically disabled subject to utilize classification performance. Here, we present the recent
brain signals to control a device without using any muscle related work of the time domain, frequency domain and
activity. In other words, it uses the brain signals and wavelet-based feature extraction methods for classification
communicates with the external devices for control. Most of EEG signals in cognitive tasks. Time domain features
of the research in BCI and related research has been per- mainly include sample entropy [12], approximate entropy
formed using electroencephalogram (EEG) signals. Any [12], permutation entropy, fractal dimension, Hjorth
good practical implementation of the BCI system demands parameters [13], Hurst component [10], and Lyapunov
an efficient brain signal processing scheme that could exponent [14]. Frequency domain features include EEG
extract features and perform classification [4]. Several absolute power, relative power and power ratio in different
methods have been reported for feature extraction, which frequency bands [15]. The time–frequency analysis include
include time domain, frequency domain, and wavelet wavelet-based feature extraction and stockwell transform
transform- (WT) based features [3]. However, WT-based [16]. Hariharan et al. [16] have used the stockwell trans-
analysis is highly effective, because it deals better with the form for feature extraction and the support vector machine
non-stationary behavior of EEG signals than other meth- (SVM) for classification of EEG signals of different cog-
ods. Wavelet-based features, including wavelet entropy [5], nitive tasks. The authors have reported classification
wavelet coefficients [2], and wavelet statistical features accuracy between 84.72 and 98.95 %. Noshadi et al. [17]
(mean, median, and standard deviations) have been repor- have used empirical mode decomposition and both time
ted for normal EEG analysis as well as in clinical appli- and frequency domain features for cognitive task classifi-
cations [6, 7]. Details on the performance of time domain, cation. The authors have employed linear classifiers (k-
frequency domain and wavelet-based techniques employed nearest neighbor and linear discriminant analysis) and have
in EEG classification for cognitive tasks and/or BCI reported 97.78 % classification accuracy. Guo et al. [8]
applications are provided in the related work section and have used weighted SVM with immune feature and clas-
the classification accuracy of these techniques are provided sified cognitive tasks with 85.4–97.5 % accuracy. Zhang
in the discussion section. Most of these studies have et al. [18] have reported 72.4–76.4 % classification accu-
reported good results in discriminating cognitive tasks of racy using high frequency power and Fischer’s discrimi-
different workloads in simulated and/or real EEG record- nant classifier for EEG classification in cognitive tasks.
ings [8]. Hence, the experimental design used in these Hosni et al. [19] utilized the EEG power feature and SVM
studies could not find dynamics in the performance of a classifier with a radial basis function (RBF) kernel, and
participant in a unique task with a constant cognitive have classified three cognitive tasks with 70 % accuracy.
workload. Therefore, this study presents EEG signal clas- Xue et al. [20] have used the wavelet packet transform for
sification in a cognitive task with a constant workload from feature extraction with the RBF classifier, and have
a baseline task using wavelet-based feature extraction and achieved 85.3 % accuracy. Zhiwei and Minfen [21] have
machine learning classifiers. used the wavelet pack entropy feature and SVM classifier
The purpose of this study has been to extract suitable and have shown 87.5–93 % accuracy for discriminating a
wavelet-based features (relative wavelet energy) for the baseline task from a cognitive task. The above cited studies
classification of EEG signals. The EEG dataset used for the have used a database of seven subjects who performed five
validation of the proposed method consisted of two clas- tasks—baseline (eyes open) task, multiplication, visual
ses—EEG recordings in a complex cognitive task (class 1), counting, mental letter composing and geometric object
and EEG recordings in a rest condition—eyes open (class rotation. The database was originally reported by Keirn and
2). The paper is structured as follows: ‘‘Related work’’ Aunon [22] at Colorado State University. The database
section reviews the related previous studies of feature consisted of only seven subjects and the experimental
extraction methods, ‘‘Materials and methods’’ section cognitive tasks were simple. Further, the majority of the
describes the materials and methods, ‘‘Experimental results studies utilized very few subjects for classification; for
and discussion’’ section presents the results and discussion, example, Zhiwei and Minfen [21] have used only two
and ‘‘Conclusion’’ section concludes the paper. subjects’ data while Nai-Jen and Palaniappan [23] have

123
Australas Phys Eng Sci Med (2015) 38:139–149 141

used only four subjects’ data. Additionally, the database Consent form and ethics approval
had a variable cognitive load in different tasks.
Many other studies have worked on EEG classification This research study was approved by the Research Coor-
in cognitive tasks using different databases recorded by the dination Committee of Universiti Teknologi PETRONAS,
authors or adopted from past studies. Such as, Lin and Malaysia [28]. All the participants had signed the informed
Hsieh [24] have classified cognitive tasks using EEG power consent form before starting the experiment. The consent
features with neural network classifier and have reported document had a brief description of this research study
78.31 % accuracy. Rodrıguez-Bermudez et al. [25] have concerning humans.
employed time, frequency and wavelet-based features with
the SVM classifier and reported 67.96–80.71 % accuracy. Experimental tasks
This study has used four subjects’ EEG data and a linear
classifier for discrimination of cognitive tasks. Karkare Eyes open task
et al. [26] have used a scaling exponent and classified two
groups who performed complex cognitive tasks using an In this task, there was no cognitive task to be performed.
artificial neural network, and have reported the classifica- The participants were instructed to sit relaxed and try to
tion performance at over 80 % accuracy. These studies think of nothing in particular. To maintain the concentra-
reported low classification accuracy and most of them have tion of the participants, they were asked to focus their
used non-linear classifiers, such as neural network and attention on a point displayed at the center of a computer
kernel-based SVM, which are time consuming in building screen during the EEG recording. The EEG recording of
models for classification. The standard psychometric cog- this task was used as a baseline signal.
nitive task, e.g., Raven the progressive metric test, has been
reported by Jahidin et al. [27]. They have utilized the EEG Raven’s advance progressive metric (RAPM)
power feature with the neural network classifier and
achieved the 88.89 % accuracy. However, this study clas- RAPM is a standard psychometric test used to measure the
sified within the group classification for high and low intellectual ability. It consists of two sets (I&II). Set-I is
cognitive processes. From the literature, we have found a used for practice, which contains 12 problems; Set-II
gap for this study, i.e., efficient feature extraction and consists of 36 problems used to measure the general cog-
classification for EEG signals for an offline dataset as well nitive ability. Each problem is a diagrammatic structure
as applicable for online EEG applications. with some missing information and with eight multiple
choices to complete the diagram’s missing part. Each
correct answer has a score of ‘1’ and a score of ‘0’ is
assigned for each incorrect answer. The score range is 36
Materials and methods
and the administration time for Set-I and Set-II is 10 and
40 min, respectively (for more details about the RAPM,
This section describes the details of the materials and
see [29, 30]).
methods used during this study, which include the experi-
mental tasks, dataset description, discrete wavelet trans-
Dataset
form and relative energy computation, description of
classifiers and the discussion of their performance
The dataset consisted of eight healthy volunteers’ EEG
parameters.
data, which were recorded while performing the RAPM
test and eyes open condition [28, 31]. The details about the
Participants procedure of this RAPM, EEG data recording and pre-
processing can be found in our previous studies [28, 31].
All of the eight healthy participants were graduate students For feature extraction and classification, the dataset was
in Universiti Teknologi PETRONAS. They participated organized into two classes. The class 1 represented all of
voluntarily in this study. All of them were male, right the eight participants’ EEG data recorded during the
handed and aged between 24 and 32 years (28.6 ± 4.20) RAPM task and the class 2 represented all of the eight
[28]. At the time of the experiment, they were free from participants’ EEG data recorded in the eyes open condition.
any medication, drugs, neurological disorder, or head In the RAPM task, each participant solved 36 problems.
injury that may have affected the experimental results. Thus, each participant was observed 36 times (maximum)
They had normal or corrected to normal vision. Previously, in the RAPM task, resulting in 36 instances corresponding
they had not experienced the cognitive task used in this to a single participant. There were some un-attempted (not
study. answered) problems with a few participants. The missing

123
142 Australas Phys Eng Sci Med (2015) 38:139–149

problems in all of the participants were excluded and we The maximum level of decomposition is specified
were left with a total of 280 instances for class 1 (8 par- depending on the principal frequency components in the
ticipants 9 36 problems = 288, excluding missing prob- given signal [2]. The coefficients of the DWT are referred
lems ) 280 instances). Similarly, each participant’s eyes to as the dot product of the original time series and the
open EEG recording (class 2) was segmented according to designated basis functions. The approximation coefficients
the corresponding numbers of the problem solved in the Ai and the detailed coefficients Di in the ith level are
RAPM task (class 1). Hence, we have balanced both the denoted as [2]:
classes in terms of instances for classification, i.e., 280 1 X
instances for each class. Ai ¼ pffiffiffiffiffi xðnÞ  uj;k ðnÞ ð3Þ
M n
Discrete wavelet transform (DWT) 1 X
Di ¼ pffiffiffiffiffi xðnÞ  wj;k ðnÞ ð4Þ
M n
The DWT is widely used for the time–frequency analysis
of biomedical signals [2, 32], especially in an EEG signal where k = 0, 1, 2, …, 2j-1 and M is the length of the EEG
analysis due to its non-stationary characteristics. The DWT time series in the discrete points.
employs extensive time windows for low frequencies and
short time windows for higher frequencies, resulting in Relative and total wavelet sub-band energy
good time–frequency analysis. The DWT decomposition of
a signal uses successive high pass and low pass filtering of The wavelet energy at each decomposition level i = 1,…,
the time series and two down samplers by 2. The high pass L is computed as follows:
filter g(n) is the discrete mother wavelet and the low pass
X
N  
filter h(n) is its mirror version [33]. The mother wavelet of ED i ¼ Dij 2 ; i ¼ 1; 2; 3; . . .; L ð5Þ
the Daubechies wavelet (db4) and the corresponding scal- j¼1
ing function are shown in Fig. 1.
X
N  
The output of the first high pass and low pass filters are EA i ¼ Aij 2 ; i¼L ð6Þ
referred to as the approximation and detailed coefficients, j¼1
represented by A1 and D1, respectively. The A1 is further
disintegrated and the procedure is repeated till the specified The ‘L’ is the maximum level of decomposition. Hence,
number of decomposition levels is reached (see Fig. 2) [32, 33]. from Eqs. 5 and 6, the total energy can be defined as:
!
The dilation function uj;k ðnÞ is dependent on the low XL

pass filter, and the wavelet function wj;k ðnÞ is follows the ETotal ¼ EDi þ EAL ð7Þ
i¼1
high pass filter, which is denoted as follows.
  The normalized energy values represent the relative
uj;k ðnÞ ¼ 2j=2 h 2j n  k ð1Þ
wavelet energy.
 
wj;k ðnÞ ¼ 2j=2 g 2j n  k ð2Þ Ej
Er ¼ ð8Þ
ETotal
where n = 0, 1, 2,…, M-1; j = 0, 1, 2,…, J-1; k = 0, 1, 2,…,
2j-1; J = log2(M); and M is the length of the signal [34]. where Ej ¼ EDi¼1;...;L or EAi¼L

Fig. 1 Mother wavelet and

scaling function (db4)

123
Australas Phys Eng Sci Med (2015) 38:139–149 143

Fig. 2 DWT sub-band

decomposition

Feature extraction using DWT Classification methods

The relevant information extraction from raw signals is a A classifier is a technique that utilizes various independent
critical step in EEG pattern classification due to its direct variable values (features) as input and predicts the corre-
influence on classification performance. The illustrative sponding class to which the independent variable belongs
representation of the proposed feature extraction scheme is [12]. In the EEG signal analysis, the features can be any
presented in Fig. 3. The band pass (1–48 Hz) EEG signal kind of extracted information from the signal, such as
was decomposed into sub-band frequencies by using the energy, entropy, power etc. and the class can be the type of
discrete wavelet transformation with the Daubechies wave- task or the stimulus used during the recording. A classifier
let of order 4 up to level 4. The approximate and detailed has a number of parameters that need to be learned from
coefficients were computed (see Fig. 4 as an example). training data. The learned classifier is a model of the
Table 1 represents one channel’s sub-band percentage rela- association between the features and the classes. For
tive energy and its frequency range of a single subject. The example, for a given feature x of a class y, the classifier is a
total and relative sub-band energies were computed from the function f that predicts the class y = f(x). After the learn-
extracted wavelet coefficients. The relative wavelet energy ing, the classifier is able to predict new instances that have
ErD1 ; ErD2 ; . . .; ErA4 was calculated using Eq. 8. not been used in the training data. Thus, the performance of
The relative energy features were computed for all of the classifier is tested on a different set of instances.
the participants and all of the channels’ data. Accordingly, To demonstrate the effectiveness of the proposed feature
the feature matrix of relative energy for a single participant extraction scheme in cognitive function classification, the
in each EEG task and each sub-band (detailed or approx- SVM, multi-layer perceptron (MLP), K-nearest neighbor
imation) became as follows: (K-NN) and Naı̈ve Byes classifiers were used. The SVM
! used a kernel trick to transform the data points into a higher
Relative Energy Feature MatrixðFr Þ ¼ ½ErA4ð280128Þ  ð9Þ
dimensional space and then separated them by a hyper-
where the number of channels was 128, number of plane with a maximal margin. The MLP is a neural net-
instances in each class was 280, D1*D4 and A4 were the work-based method, which is commonly used for per-
detailed and approximation coefficients. Accordingly, the forming a different variety of detection and estimation
ED1ð280128Þ represented the relative energy feature matrix tasks. The K-nearest neighbor works to find a testing
of the first detailed coefficients for all of the eight partic- sample’s class by the majority class of the k nearest
ipants in each class. training samples. The Naı̈ve Bayes is a simple and efficient

Fig. 3 Proposed scheme for

feature extraction and Discrete Wavelet Sub-Bands Relative
classification of EEG signal Decomposition Energy Calculation
EEG Signals Coefficients
WT Features

Cognitive Task Classifiers

Classification Performance
ML Classifiers Machine Learning (ML)
Comparison Classifiers

123
144 Australas Phys Eng Sci Med (2015) 38:139–149

Fig. 4 Representation of the

A4 and D1–D4 components of
one participant’s EEG signal at
the F3 scalp location during a
cognitive task

tenfold cross validation method divided the dataset into ten

Table 1 Frontal F3 channel’s sub-band percentages’ relative energy
and their frequency range subsets of equal size and used nine subsets for the classifier
training and one subset for the classifier testing. This process
Levels Wavelet Wavelet Frequency
energy (%) coefficients bands (Hz) repeated ten times, each time leaving out one of the subsets
from the training, which was used for testing. This method
1 0.48 D1 23.5–48 has the advantage that it utilized all of the instances in the
2 8.47 D2 12.25–23.5 dataset for both training and testing. The classifiers’ per-
3 10.70 D3 6.12–12.25 formances were computed using the most commonly used
4 13.63 D4 3.06–6.12 parameters, are the accuracy, sensitivity, specificity, preci-
4 70.96 A4 0.53–3.06 sion, and Kappa statistic [38]. These parameters are defined
D1 detailed coefficient of level 1, D2 detailed coefficient of level 2, as follows.
D3 detailed coefficient of level 3, D4 detailed coefficient of level 4, Total no: of correctly classified instances
A4 approximation level 4 Accuracy ¼
Total numbers of instances
 100
statistical method, which is based on Bayes’ theorem. For ð10Þ
more details about SVM, MLP, K-NN and Naı̈ve Bayes
(see [1, 2, 35–37]. True Positive
Sensitivity ¼  100 ð11Þ
True Positive þ False Negative
True Negative
Experimental results and discussion Specificity ¼  100 ð12Þ
True Negative þ False Positive
In this section, we present the experimental results for True Positve
Precision ¼  100 ð13Þ
validation, and discuss them. We start with the experi- True Positive þ False Positive
 
mental set-up used for these experiments. Po  PCe
KappaðkÞ ¼   ð14Þ
1  PCe
Experimental setup
where Po represents the probability of the overall agree-
The classifiers were trained and tested for these extracted ment of the label assignments between the classifier and the
features using the tenfold cross validation method. The true process, and PCe denotes the chance agreement over the

123
Australas Phys Eng Sci Med (2015) 38:139–149 145

labels—sum of the proportion of instances assigned to a such as sensitivity, specificity, precision and Kappa sta-
class multiplies the proportion of true labels of that class in tistic were prominent. From these results, it seems that the
the dataset. relative wavelet energy in the low frequency band
(0.53–3.06 Hz) and the above low frequency band
Classification results (3.06–6.12 Hz) was a useful feature to classify the EEG
brain patterns in eyes open and the complex cognitive task
The extracted relative wavelet energy features of D1–D4 (i.e., RAPM).
and A4 were classified using SVM with RBF kernel, MLP
with five hidden layers, K-NN with k = 1, and Naı̈ve Ba-
Discussion
yes classifiers for both of the EEG conditions, i.e., eyes
open and cognitive task. This classification process was
Comparison with existing techniques
implemented for the extracted features from all of the
decomposition levels (D1–D4, and A4). However, the
A direct comparison of the results with the previous research
classification results were not prominent in all of the
in EEG signals was hard due to the variety of EEG datasets,
decomposition levels. The highest classification perfor-
wavelet types, decomposition levels, participants’ variabil-
mance was found in the relative energy of the approxi-
ity, and the cognitive tasks used. However, a brief compar-
mation coefficients and detailed coefficients of level 4,
ison with the previous related studies is presented here. The
which reflected the low frequency (0.53–3.06 Hz) and
information about the dataset, feature extraction methods,
above low frequency (3.06–6.12 Hz) dominations in the
cognitive tasks, machine learning algorithm and the classi-
cognitive task (see Tables 2 and 3). A representative signal
fication performance reported in previous studies are pre-
of 8 s from both the experimental tasks at the F3 electrode
sented in Table 4. The list of studies in Table 4 have used the
is presented in Fig. 5. The amplitude differences can be
time domain, frequency domain, autoregressive (AR) coef-
observed in both the 0.5–3 and 3–6 Hz frequency bands of
ficient and/or wavelet transform-based features for EEG
the two experimental tasks.
classification in a cognitive task as mentioned in ‘‘Related
The SVM classifier achieved 98.75 % accuracy and the
work’’ section. The majority of the studies have used non-
MLP and K-NN classifiers achieved 98.21 %, accuracy in
linear classifiers (e.g., ANN and kernel-based SVM), which
the classification using the relative energy of the approxi-
are complex in nature and time consuming to build the
mation A4 coefficients. In the detailed D4 coefficients, the
classification model. In the case of using very few instances
SVM and MLP achieved 98.21 and 98.57 % accuracy,
in the classification as mentioned in a few studies in Table 4,
respectively, as shown in Tables 2 and 3. The accuracy of
it may causes the over fitting problem in classification [39].
the Naı̈ve Bayes classifier was also found to be above
In this work, we used 280 instances for each class in the
80 %. The values of the other performance parameters,

Table 2 Classification results of the relative wavelet energy of the level 4 approximate coefficients (A4) for the cognitive task
Classifier Accuracy (%) Sensitivity (%) Specificity (%) Precision (%) AUC Kappa statistic

SVM 98.75 100 97.50 97.60 0.98 0.97

MLP 98.21 100 96.40 96.60 0.98 0.96
K-NN 98.21 99.60 96.80 96.90 0.98 0.96
Naı̈ve Bayes 83.57 75.00 92.10 90.50 0.93 0.67
SVM support vector machine, MLP multi-layers perceptron, K-NN k-nearest neighbors, AUC area under the receiver operating characteristic
curve

Table 3 Classification results of the relative wavelet energy of the level 4 detailed coefficients (D4) for the cognitive task
Classifier Accuracy (%) Sensitivity (%) Specificity (%) Precision (%) AUC Kappa statistic

SVM 98.21 99.60 96.80 96.90 0.98 0.96

MLP 98.57 99.60 97.60 97.60 0.98 0.97
K-NN 97.14 98.90 95.40 95.50 0.97 0.94
Naı̈ve Bayes 83.03 84.60 81.40 82.00 0.89 0.66
SVM support vector machine, MLP multi-layers perceptron, K-NN k-nearest neighbors, AUC area under the receiver operating characteristic
curve

123
146 Australas Phys Eng Sci Med (2015) 38:139–149

classification. Such a high number of instances in classifi- contributor to the P300 component of ERP [47]. Gennady
cation have not been reported in cognitive task classification et al. [48] reviewed the delta band relationship with cog-
using EEG. Hence, we have used both the linear (e.g., nitive processing and have confirmed that delta is linked
K-NN) as well as the non-linear classifiers (e.g., ANN) along with the cognitive process. Similarly, the theta rhythms are
with the 10-cross validation scheme. The benefit of the the most intensively studied in the cognitive neuroscience
10-cross validation process is that all instances in the sample aimed at correlating the theta rhythms with cognitive
are used for both training and validation exactly once [39]. processing [49–51]. Particularly, theta in the frontal
Therefore, in the presence of a high number of instances, the regions is critical for attentional and cognitive processing
use of multiple classifiers in the present study made it in ERP tasks [50]. Most of these studies reported signifi-
comparable with previous studies in terms of classification cant increase in delta and theta power in the cognitive
performance. The classification results of this study in both tasks. This may be the reason, that we achieved high
linear and non-linear classifiers were found to be better than classification accuracy in the low frequency bands
related studies which used similar classifiers and the same (0.53–3.06 and 3.06–6.12 Hz) for discriminating the
nature of the cognitive tasks. RAPM and baseline—eyes open task. Hence, the results of
this study reflect the previous studies’ findings in cognitive
EEG low frequencies with cognitive neuroscience neuroscience research.
perspective

The EEG low frequency bands (delta and theta) have been Conclusion
reported by the cognitive neuroscientists as cognitive
rhythms, and have been linked with cognitive and attention This paper has presented the use of relative discrete
demanded tasks [43–45]. Especially, the event-related wavelet energy along with machine learning algorithms for
potential (ERP) studies have reported the most significant the classification and the quantitative analysis of sponta-
findings of the delta band related to cognitive processing neous EEG signals recorded during complex cognitive
[44], i.e., the associations of the P300 component with the task. The EEG signals were split into sub-bands using
cognitive process [46]. This relationship has been widely DWT with Daubechies (db4) wavelets and the sub-bands’
reported in the cognitive neuroscience literature. In brief, relative energies were computed for all of the 128 channels
the delta band has been considered as the primary of each subject’s EEG recording. For classification, four

Fig. 5 Representative signal of low frequencies (delta and theta bands) at the F3 electrode position (Red color shows the signal of the cognitive
task, and the blue color represents the signal of the eyes open task)

123
Australas Phys Eng Sci Med (2015) 38:139–149 147

Table 4 Summary of existing feature extraction and classification techniques for EEG in cognitive tasks
Ref Year Subjects EEG scalp Cognitive task Feature Classifier Accuracy
(instances) electrodes (%)

[20] 2003 4 6 Five tasks as reported by Wavelet packet RBF 85.30

Keirn and Aunon [22] network
[40] 2004 4 6 Five tasks as reported by DWT ANN 74.40–82.30
Keirn and Aunon [22]
[4] 2006 7 6 Five tasks as reported by AR coefficients SVM, ANN 53.98–56.07
Keirn and Aunon [22] and ELM
[21] 2007 2 6 Five tasks as reported by Wavelet Packet Entropy up to seven SVM 87.5–93.0
Keirn and Aunon [22] levels
[19] 2007 7 6 Three tasks classified Frequency Bands Power and AR SVM with 70
(baseline, mental letter, coefficients RBF
multiplication) from kernel
Keirn and Aunon [22]
[41] 2009 9 9 – Wavelet relative energy ANN 95.2
[42] 2009 7 6 Five tasks as reported by Empirical mode decomposition, time ANN and 87.35–91.17
Keirn and Aunon [22] and frequency domain features LDA
[6] 2009 4 6 Five tasks (Baseline or DWT with db4 up to level 5 K-NN 81.48–89.58
relaxed state,
multiplication, visual
counting, mental letter
composing geometric
object rotation) as
reported by Keirn and
Aunon [22]
[24] 2009 3 8 Three cognitive tasks Power spectral density feature ANN and 65.90–68.35
(imagination of left hand SVM
movement, imagination
of right hand movement,
and generation of words)
[8] 2011 7 6 Five tasks as reported by Immune feature SVM 85.4–97.5
Keirn and Aunon [22]
[27] 2014 (120) 9 Raven’s progressive metric Frequency Band power ANN 88.89
This 2014 8(560) 128 Baseline (eyes open), Wavelet relative energy SVM 98.75
work RAPM MLP 98.57
K-NN 98.21
Naı̈ve Bayes 83.57
SVM support vector machine, MLP multi-layers perceptron, K-NN k-nearest neighbors, AUC area under the receiver operating characteristic
curve, AR autoregressive, RBF radial basis function, ANN artificial neural network, LDA linear discriminant analysis, DWT Discrete wavelet
transform, RAPM Raven’s advance progressive metric, db4 Daubechies wavelet order 4, 9 unknown. The values in brackets ‘()’ indicate number
of instances and values without bracket or outside bracket denote number of subjects only in the 3rd column from side of the above table

different classifiers (SVM, MLP, K-NN and Naı̈ve Bayes) the localization characteristics of the wavelet transforma-
were employed and their performance was evaluated for tion [5]. The low frequencies, especially in the range of the
cognitive task discrimination. The classification results of delta band, are perceived in cognitive neuroscience as the
SVM and MLP demonstrated above 98 % accuracy with primary contributor to the cognitive processing. Hence, the
features extracted using A4 (0.53–3.06 Hz) and D4 proposed feature scheme has the clinical significance to be
(3.06–6.12 Hz) sub-bands. The wavelet energy is a useful applied on real time EEGs in BCI applications for severe
feature to classify the EEG signals corresponding to com- motor disabled patients to control external devices using
plex cognitive tasks, and it will be helpful for EEG clas- cognitive power. This may be implemented in future work.
sification in clinical applications, such as epilepsy,
depression, and stress diagnosis as it is capable of identi- Acknowledgments This research work has been supported by Uni-
versity Research Internal Funding (URIF: 0153AA-B26), Universiti
fying variations in non-stationary EEG signals because of

123
148 Australas Phys Eng Sci Med (2015) 38:139–149

Teknologi PETRONAS; the Fundamental Research Grant Scheme 19. Hosni SM, Gadallah ME, Bahgat SF, AbdelWahab MS (2007)
(Ref: FRGS/1/2014/TK03/UTP/02/1), Ministry of Education (MOE), Classification of EEG signals using different feature extraction
Malaysia and by NSTIP strategic technologies programs, Grant number techniques for mental-task BCI. In: International conference on
(12-INF2582-02), in the Kingdom of Saudi Arabia. computer engineering & systems, 2007. ICCES’07, pp 220–226
20. Xue J-Z, Zhang H, Zheng C-x, Yan X-G (2003) Wavelet packet
transform for feature extraction of EEG during mental tasks. In:
International conference on machine learning and cybernetics,
References 2003, pp 360–363
21. Zhiwei L, Minfen S (2007) Classification of mental task EEG
1. Übeyli ED (2009) Combined neural network model employing signals using wavelet packet entropy and SVM. In: 8th Interna-
wavelet coefficients for EEG signals classification. Digit Signal tional conference on electronic measurement and instruments,
Proc 19:297–308 2007. ICEMI’07, pp 3-906–3-909
2. Orhan U, Hekim M, Ozer M (2011) EEG signals classification 22. Keirn ZA, Aunon JI (1990) A new mode of communication
using the K-means clustering and a multilayer perceptron neural between man and his surroundings. IEEE Trans Biomed Eng
network model. Expert Syst Appl 38:13475–13481 37:1209–1214
3. Iscan Z, Dokur Z, Demiralp T (2011) Classification of electro- 23. Nai-Jen H, Palaniappan R (2004) Classification of mental tasks
encephalogram signals with combined time and frequency fea- using fixed and adaptive autoregressive models of EEG signals.
tures. Expert Syst Appl 38:10499–10505 In: 26th Annual international conference of the IEEE engineering
4. Liang N-Y, Saratchandran P, Huang G-B, Sundararajan N (2006) in medicine and biology society, 2004. IEMBS’04, pp 507–510
Classification of mental tasks from EEG signals using extreme 24. Lin C-J, Hsieh M-H (2009) Classification of mental task from
learning machine. Int J Neural Syst 16:29–38 EEG data using neural networks based on particle swarm opti-
5. Rosso OA, Blanco S, Yordanova J, Kolev V, Figliola A, Schür- mization. Neurocomputing 72:1121–1130
mann M et al (2001) Wavelet entropy: a new tool for analysis of 25. Rodrı́guez-Bermúdez G, Garcı́a-Laencina PJ, Roca-González J,
short duration brain electrical signals. J Neurosci Methods Roca-Dorda J (2013) Efficient feature selection and linear dis-
105:65–75 crimination of EEG signals. Neurocomputing 115:161–165
6. Yazdani A, Ebrahimi T, Hoffmann U (2009) Classification of 26. Karkare S, Saha G, Bhattacharya J (2009) Investigating long-
EEG signals using Dempster Shafer theory and a k-nearest range correlation properties in EEG during complex cognitive
neighbor classifier. In: 4th International IEEE/EMBS conference tasks. Chaos Solitons Fractals 42:2067–2073
on neural engineering, pp 327–330 27. Jahidin A, Ali MM, Taib MN, Tahir NM, Yassin IM, Lias S
7. Garry H, McGinley B, Jones E, Glavin M (2013) An evaluation (2014) Classification of intelligence quotient via brainwave sub-
of the effects of wavelet coefficient quantisation in transform band power ratio features and artificial neural network. Comput
based EEG compression. Comput Biol Med 43:661–669 Methods Programs Biomed 114:50–59
8. Guo L, Wu Y, Zhao L, Cao T, Yan W, Shen X (2011) Classification 28. Amin HU, Malik AS, Subhani AR, Badruddin N, Chooi W-T
of mental task from EEG signals using immune feature weighted (2013) Dynamics of scalp potential and autonomic nerve activity
support vector machines. IEEE Trans Magn 47:866–869 during intelligence test. In: Lee M et al (eds) Neural information
9. Taghizadeh-Sarabi M, Daliri MR, Niksirat KS (2014) Decoding processing, vol 8226. Springer, Berlin, pp 9–16
objects of basic categories from electroencephalographic signals 29. Raven J (2000) The Raven’s progressive matrices: change and
using wavelet transform and support vector machines. Brain stability over culture and time. Cogn Psychol 41:1–48
Topogr 1–14 30. Kunda M, McGreggor K, Goel A (2012) Reasoning on the
10. Acharya UR, Sree SV, Ang PCA, Yanti R, Suri JS (2012) Raven’s advanced progressive matrices test with iconic visual
Application of non-linear and wavelet based features for the representations. In: 34th Annual conference of the cognitive
automated identification of epileptic EEG signals. Int J Neural science society pp 1828–1833
Syst 22:1250002 31. Amin HU, Malik AS, Badruddin N, Chooi W-T (2013) EEG
11. Gandhi T, Panigrahi BK, Anand S (2011) A comparative study of mean power and complexity analysis during complex mental
wavelet families for EEG signal classification. Neurocomputing task. In: ICME International conference on complex medical
74:3051–3057 engineering (CME) pp 648–651
12. Richman JS, Moorman JR (2000) Physiological time-series 32. Jahankhani P, Kodogiannis V, Revett K (2006) EEG signal
analysis using approximate entropy and sample entropy. Am J classification using wavelet feature extraction and neural net-
Physiol Heart Circ Physiol 278:H2039–H2049 works. In: IEEE John Vincent Atanasoff 2006 international
13. Vidaurre C, Krämer N, Blankertz B, Schlögl A (2009) Time symposium on modern computing, 2006. JVA’06. pp 120–124
domain parameters as a feature for EEG-based brain–computer 33. Subasi A (2007) EEG signal classification using wavelet feature
interfaces. Neural Netw 22:1313–1319 extraction and a mixture of expert model. Expert Syst Appl
14. Übeyli ED (2010) Lyapunov exponents/probabilistic neural net- 32:1084–1093
works for analysis of EEG signals. Expert Syst Appl 37:985–992 34. Gonzalez RC, Woods RE (2002) Digital image processing, 2nd
15. Thatcher RW, North D, Biver C (2005) EEG and intelligence: edn. Prentice Hall, SL
relations between EEG coherence, EEG phase delay and power. 35. Hausfeld L, De Martino F, Bonte M, Formisano E (2012) Pattern
Clin Neurophysiol 116:2129–2141 analysis of EEG responses to speech and voice: influence of
16. Hariharan M, Vijean V, Sindhu R, Divakar P, Saidatul A, Yaacob feature grouping. Neuroimage 59:3641–3651
S (2014) Classification of mental tasks using stockwell transform. 36. Ben-Hur A, Weston J (2010) A user’s guide to support vector
Comput Electr Eng 40:1741 machines. In: Data mining techniques for the life sciences,
17. Noshadi S, Abootalebi V, Sadeghi MT, Shahvazian MS (2014) Springer, New York, pp 223–239
Selection of an efficient feature space for EEG-based mental task 37. Pereira F, Mitchell T, Botvinick M (2009) Machine learning clas-
discrimination. Biocybern Biomed Eng 34:159–168 sifiers and fMRI: a tutorial overview. Neuroimage 45:S199–S209
18. Zhang L, He W, He C, Wang P (2010) Improving mental task 38. Subasi A, Gursoy MI (2010) EEG signal classification using
classification by adding high frequency band information. J Med PCA, ICA, LDA and support vector machines. Expert Syst Appl
Syst 34:51–60 37:8659–8666

123
Australas Phys Eng Sci Med (2015) 38:139–149 149

39. Witten IH, Frank E (2005) Data mining: practical machine 45. Amin H, Malik AS (2013) Human memory retention and recall
learning tools and techniques. Morgan Kaufmann, Burlington processes: a review of EEG and fMRI studies. Neurosciences
40. Daud M, Yunus J (2004) Classification of mental tasks using de- 18:330–344
noised EEG signals. In: 7th International conference on signal 46. Ergen M, Marbach S, Brand A, Başar-Eroğlu C, Demiralp T
processing, pp 2206–2209 (2008) P3 and delta band responses in visual oddball paradigm in
41. Guo L, Rivero D, Seoane JA, Pazos A (2009) Classification of schizophrenia. Neurosci Lett 440:304–308
EEG signals using relative wavelet energy and artificial neural 47. Demiralp T, Ademoglu A, Schürmann M, Basar-Eroglu C, Basar
networks. In: Proceedings of the first ACM/SIGEVO summit on E (1999) Detection of P300 waves in single trials by the wavelet
genetic and evolutionary computation, pp 177–184 transform (WT). Brain Lang 66:108–128
42. Diez PF, Mut V, Laciar E, Torres A, Avila E (2009) Application 48. Gennady GK (2012) EEG delta oscillations as a correlate of basic
of the empirical mode decomposition to the extraction of features homeostatic and motivational processes. Neurosci Biobehav Rev
from EEG signals for mental task classification. In: Annual 36:677–695
international conference of the IEEE engineering in medicine and 49. Baijal S, Srinivasan N (2010) Theta activity and meditative
biology society, 2009. EMBC 2009, pp 2579–2582 states: spectral changes during concentrative meditation. Cogn
43. Dimitriadis SI, Laskaris NA, Tsirka V, Vourkas M, Michelo- Process 11:31–38
yannis S (2010) What does delta band tell us about cognitive 50. Sakowitz O (2001) The selectively distributed theta system:
processes: a mental calculation study. Neurosci Lett 483:11–15 functions. Int J Psychophysiol 39:197–212
44. Harper J, Malone SM, Bernat EM (2014) Theta and delta band 51. Nigbur R, Ivanova G, Stürmer B (2011) Theta power as a marker
activity explain N2 and P3 ERP component activity in a go/no-go for cognitive interference. Clin Neurophysiol 122:2185–2194
task. Clin Neurophysiol 125:124–132

123