Hindawi Publishing Corporation

ISRN Neuroscience
Volume 2014, Article ID 730218, 7 pages
http://dx.doi.org/10.1155/2014/730218

Review Article
Methods of EEG Signal Features Extraction Using Linear
Analysis in Frequency and Time-Frequency Domains

Amjed S. Al-Fahoum1 and Ausilah A. Al-Fraihat2

1
Biomedical Systems and Informatics Engineering Department, Hijjawi Faculty for Engineering Technology,
Yarmouk University, Irbid 21163, Jordan
2
Biomedical Engineering Department, Faculty of Engineering, Hashemite University, Zarqa 13115, Jordan

Correspondence should be addressed to Amjed S. Al-Fahoum; [email protected]

Received 20 October 2013; Accepted 9 January 2014; Published 13 February 2014

Academic Editors: A. Grant, J. A. Hinojosa, and M. S. Oliveira

Copyright © 2014 A. S. Al-Fahoum and A. A. Al-Fraihat. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.

Technically, a feature represents a distinguishing property, a recognizable measurement, and a functional component obtained
from a section of a pattern. Extracted features are meant to minimize the loss of important information embedded in the signal. In
addition, they also simplify the amount of resources needed to describe a huge set of data accurately. This is necessary to minimize
the complexity of implementation, to reduce the cost of information processing, and to cancel the potential need to compress
the information. More recently, a variety of methods have been widely used to extract the features from EEG signals, among
these methods are time frequency distributions (TFD), fast fourier transform (FFT), eigenvector methods (EM), wavelet transform
(WT), and auto regressive method (ARM), and so on. In general, the analysis of EEG signal has been the subject of several studies,
because of its ability to yield an objective mode of recording brain stimulation which is widely used in brain-computer interface
researches with application in medical diagnosis and rehabilitation engineering. The purposes of this paper, therefore, shall be
discussing some conventional methods of EEG feature extraction methods, comparing their performances for specific task, and
finally, recommending the most suitable method for feature extraction based on performance.

1. Introduction [13]. Many methods have been introduced to eliminate these

unwanted signals. Each of them has its advantages and
In recent years, brain computer interface and intelligent disadvantages. Nevertheless, there is a common path for EEG
signal segmentation have attracted a great interest ranging signal processing (Figure 2). The first part is preprocessing
from medicine to military objectives [1–6]. To facilitate brain- which includes acquisition of signal, removal of artifacts,
computer interface assembly, a professional method of feature signal averaging, thresholding of the output, enhancement of
extraction from EEG signal is desired. the resulting signal, and finally, edge detection. The second
The brain electrical activity is represented by the elec- step in the operation is the feature extraction scheme which
troencephalogram (EEG) signals. Many neurological diseases is meant to determine a feature vector from a regular vector.
(i.e., epilepsy) can be diagnosed by studying the EEG signals A feature is a distinctive or characteristic measurement,
[7–9]. The recoding of the EEG signals is performed by fixing transform, structural component extracted from a segment of
an electrode on the subject scalp using the standardized a pattern [14]. Statistical characteristics and syntactic descrip-
electrode placement scheme (Figure 1) [10–12]. However, tions are the two major subdivisions of the conventional
there are many sources of artifacts. The signal noise which feature extraction modalities. Feature extraction scheme is
can set in when signal is being captured will adversely affect meant to choose the features or information which is the most
the useful feature in the original signal. The major sources important for classification exercise [15–17]. The final stage is
of the artifact are muscular activities, blinking of eyes during signal classification which can be solved by linear analysis,
signal acquisition procedure, and power line electrical noise nonlinear analysis, adaptive algorithms, clustering and fuzzy
 23567872, 2014, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2014/730218, Wiley Online Library on [19/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 ISRN Neuroscience

Nasion domain. Different common methods of interest were com-

pared and the general advantages and disadvantages of these
modalities were discussed.
Pp1 Pp2
2.1. Fast Fourier Transform (FFT) Method. This method emp-
F7 F8
loys mathematical means or tools to EEG data analysis.
Fz
F3 F4 Characteristics of the acquired EEG signal to be analyzed
are computed by power spectral density (PSD) estimation in
A2 A3
order to selectively represent the EEG samples signal. How-
T3 C3 C2 C4 T4 ever, four frequency bands contain the major characteristic
waveforms of EEG spectrum [18].
The PSD is calculated by Fourier transforming the esti-
L P3 Pz P4 R
T6 mated autocorrelation sequence which is found by nonpara-
T5
metric methods. One of these methods is Welch’s method.
The data sequence is applied to data windowing, producing
O1 O2
modified periodograms [19]. The information sequence 𝑥𝑖 (𝑛)
is expressed as
Anion
𝑥𝑖 (𝑛) = 𝑥 (𝑛 + 𝑖𝐷) , 𝑛 = 0, 1, 2, . . . , 𝑀 − 1
Figure 1: Standardized electrode placement scheme [11]. (1)
while 𝑖 = 0, 1, 2, . . . , 𝐿 − 1;

Raw EEG
take 𝑖𝐷 to be the point of start of the 𝑖th sequence. Then 𝐿
of length 2𝑀 represents data segments that are formed. The
resulting output periodograms give
󵄨 󵄨󵄨2
1 󵄨󵄨󵄨󵄨𝑀−1
Segment
detection ≈ (𝑖) −𝑗2𝜋𝑓𝑛 󵄨󵄨󵄨
𝑃𝑥𝑥 (𝑓) = 󵄨 ∑ 𝑥 (𝑛)𝑤(𝑛)𝑒 󵄨󵄨 . (2)
𝑀𝑈 󵄨󵄨󵄨󵄨 𝑛=0 𝑖 󵄨󵄨
󵄨
Feature Here, in the window function, 𝑈 gives normalization factor
extraction of the power and is chosen such that

1 𝑀−1 2
Classification 𝑈= ∑ 𝑤 (𝑛) , (3)
𝑀 𝑛=0

where 𝑤(𝑛) is the window function. The average of these

Output
modified periodograms gives Welch’s power spectrum as
follows:

Figure 2: Stages of EEG signal processing. 𝑊 1 𝐿−1 ≈ (𝑖)

𝑃𝑥𝑥 = ∑ 𝑃 (𝑓) . (4)
𝐿 𝑖=0 𝑥𝑥

techniques, and neural networks. This is done by exploiting 2.2. Wavelet Transform (WT) Method. WT plays an impor-
the algorithmic characteristics of the feature vector of the data tant role in the recognition and diagnostic field: it compresses
input and thus gives rise to a hypothesis [10, 15]. the time-varying biomedical signal, which comprises many
This paper presents a short review of mathematical data points, into a small few parameters that represents the
methods for extracting features from EEG signals. The review signal [14].
considers five different methods for EEG signal extracting. As the EEG signal is nonstationary [7], the most suitable
The adopted approach is such that a full literature review way for feature extraction from the raw data is the use of
is introduced for the five different techniques, summarizing the time-frequency domain methods like wavelet transform
their strengths and weaknesses. (WT) which is a spectral estimation technique in which
any general function can be expressed as an infinite series
2. Methods of wavelets [20–22]. Since WT allows the use of variable
sized windows, it gives a more flexible way of time-frequency
Different articles were used to extract advantages and disad- representation of a signal. In order to get a finer low-
vantages of selected methods by thoroughly reviewing chosen frequency resolution, WT long time windows are used; in
articles including the main methods for linear analysis of contrast in order to get high-frequency information, short
one-dimensional signals in the frequency or time-frequency time windows are used [13].
 23567872, 2014, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2014/730218, Wiley Online Library on [19/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ISRN Neuroscience 3

D1 The relationship between WTs and filter ℎ, that is, low

g[n] ↓2 pass, can be represented as follows:
D2
g[n] ↓2
X[n] D3
A1
g[n] ↓2 𝐻 (𝑧) 𝐻 (𝑧−1 ) + 𝐻 (−𝑧) 𝐻 (−𝑧−1 ) = 1. (7)
h[n] ↓ 22 A2
h[n] ↓2 A3 Here, 𝐻(𝑧) represents filter’s ℎ 𝑧-transform. The high-pass
h[n] ···
↓2 filter’s complementary 𝑧-transform is expressed as
Figure 3: Implementation of decomposition of DWT [14].
𝐺 (𝑧) = 𝑧𝐻 (−𝑧−1 ) . (8)

By precisely describing the features of the signal segment

Furthermore, WT only involves multiscale structure and within a specified frequency domain and localized time dom-
not single scale. This method is just the continuation of ain properties, there are a lot of advantages that overshadow
the orthodox Fourier transform method [23]. Moreover, it the high computational and memory requirement of the
is meant to resolve issues of nonstationary signals such as conventional convolution based implementation of the DWT
EEG [14]. In the WT method, the original EEG signal is [14, 23].
represented by secured and simple building blocks known as
wavelets. The mother wavelet gives rise to these wavelets as
part of derived functions through translation and dilation, 2.3. Eigenvectors. These methods are employed to calculate
that is, (shifting) and (compression and stretching) opera- signals’ frequency and power from artifact dominated mea-
tions along the time axis, respectively [24]. There are two surements. The essence of these methods is the potential of
categories for the WT; the first one is continuous while the the Eigen decomposition to correlate even artifact corrupted
other one is discrete [14]. signal. There are a few available eigenvector methods, among
them are Pisarenko’s method, MUSIC method, and mini-
mum-norm method [25, 26].
2.2.1. Continuous Wavelet Transform (CWT) Method. This
can be expressed as
2.3.1. Pisarenko’s Method. Pisarenko’s method is among the
∞ available eigenvector approaches used to evaluate power
∗
CWT (𝑎, 𝑏) = ∫ 𝑥 (𝑡) 𝜓𝑎,𝑏 (𝑡) 𝑑𝑡, (5) spectral density (PSD). To calculate the PSD, the mathemati-
−∞
cal expression 𝐴(𝑓) is employed and given as [27, 28]
𝑥(𝑡) stands for the unprocessed EEG, where 𝑎 stands for
𝑚
dilation, and 𝑏 represents translation factor. The 𝜓𝑎,𝑏 (𝑡) den-
𝐴 (𝑓) = ∑ 𝑎𝑘 𝑒−𝑗2𝜋𝑓𝑘 . (9)
otes the complex conjugate and can be calculated by
𝑘=0

1 𝑡−𝑏 In the equation above, 𝑎𝑘 stands for coefficients of the defined

𝜓𝑎,𝑏 (𝑡) = 𝜓( ), (6)
√|𝑎| 𝑎 equation and 𝑚 defines eigenfilter’s 𝐴(𝑓) order [25, 26].
Pisarenko method uses signal desired equation to estimate
where 𝜓(𝑡) means wavelet. However, its major weakness is the signal’s PSD from eigenvector equivalent to the minimum
that scaling parameter 𝑎 and translation parameter 𝑏 of CWT eigenvalue as follows:
change continuously. Thus, the coefficients of the wavelet for
all available scales after calculation will consume a lot of effort 1
𝑃PISARENKO = 󵄨 .
and yield a lot of unused information [14]. 󵄨󵄨𝐴 (𝑓)󵄨󵄨󵄨2 (10)
󵄨 󵄨
2.2.2. Discrete Wavelet Transform (DWT). In order to address
2.3.2. MUSIC Method. This method eradicates issues related
the weakness of the CWT, discrete wavelet transform (DWT)
to false zeros by the help of the spectra’s average equivalent
has been defined on the base of multiscale feature rep-
to artifact subspace of the whole eigenvectors [28]. Resulting
resentation. Every scale under consideration represents a
power spectral density is therefore obtained as
unique thickness of the EEG signal [23]. The multiresolution
decomposition of the raw EEG data 𝑥(𝑛) is shown in Figure 3. 1
Each step contains two digital filters, 𝑔(𝑛) and ℎ(𝑛), and two 𝑃MUSIC (𝑓) = . (11)
󵄨󵄨 󵄨2
󵄨󵄨𝐴 (𝑓)󵄨󵄨󵄨
𝐾−1
downsamplers by 2. The discrete mother wavelet 𝑔(𝑛) is a high (1/𝐾) ∑𝑖=0
pass in nature, while its mirror image is ℎ(𝑛) is a low-pass in
nature. 2.3.3. Minimum Norm Method. This method makes false
As shown in Figure 3, each stage output provides a detail zeros in the unit circle to separate them from real zeros to be
of the signal 𝐷 and an approximation of the signal 𝐴, where able to calculate a demanded noise subspace vector 𝑎 from
the latest becomes an input for the next step. The number of either the noise or signal subspace eigenvectors. However,
levels to which the wavelet decomposes is chosen depending while the Pisarenko technique form application of only the
on the component of the EEG data with dominant frequency noise subspace eigenvector corresponding to the minimum
[14]. eigenvalue, the minimum norm technique picks a linear
 23567872, 2014, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2014/730218, Wiley Online Library on [19/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 ISRN Neuroscience

combination of the whole of noise subspace eigenvectors Table 1: Comparison between FFT and AR [8].
[25, 26]. This technique is depicted by
Method Frequency resolution Spectral leakage
1 FFT Low High
𝑃MIN (𝑓, 𝐾) = 󵄨 .
󵄨󵄨𝐴 (𝑓)󵄨󵄨󵄨2 (12)
AR High Low
󵄨 󵄨
WT High Low
All the aforementioned eigenvector methods can best
address the signal that is composed of many distinctive
sinusoids embedded in noise. Consequently, they are prone
to yield false zeros and thus resulting in a relatively poor 2.5. Autoregressive Method. Autoregressive (AR) methods
statistical accuracy [26]. estimate the power spectrum density (PSD) of the EEG using
a parametric approach. Therefore, AR methods do not have
problem of spectral leakage and thus yield better frequency
2.4. Time-Frequency Distributions. These methods require resolution unlike nonparametric approach. Estimation of
noiseless signals to provide good performance. Therefore, PSD is achieved by calculating the coefficients, that is, the
very restricted preprocessing stage is necessary to get rid of parameters of the linear system under consideration. Two
all sorts of artifacts. Being time-frequency methods they deal methods used to estimate AR models are briefly described
with the stationary principle; windowing process is therefore below [18, 19].
required in the preprocessing module [29]. The definition of
TFD for a signal 𝑥(𝑛) was generalized by Cohen as [30]
2.5.1. Yule-Walker Method. In this method, AR parameters or
1 ∞ ∞ coefficients are estimated by exploiting the resulting biased
𝑃 (𝑡, 𝑤) = ∫ ∫ 𝐴 (𝜃, 𝜏) Φ (𝜃, 𝜏) 𝑒−𝑗𝜃𝑡−𝑗𝜔𝜏 𝑑𝜃 𝑑𝜏,
2𝜋 −∞ −∞ approximate of the autocorrelation data function. This is done
(13) by subsequently finding the minimization of the least squares
of the forward prediction error as given below [31]:
where
1 ∞ 𝜏 𝜏 𝑟 (0)𝑥𝑥 𝑟 (−1)𝑥𝑥 ⋅ ⋅ ⋅ 𝑟(−𝑝 + 1)𝑥𝑥 𝑎 (1)
𝐴 (𝜃, 𝜏) = ∫ 𝑥 (𝑢 + ) 𝑥∗ (𝑢 − ) 𝑒𝑗𝜃𝑢 𝑑𝑢. (14)
2𝜋 −∞ 2 2 [ 𝑟 (1)𝑥𝑥 𝑟 (0)𝑥𝑥 . . . 𝑟(−𝑝 + 2)𝑥𝑥 ] [ ]
[ ] [ 𝑎 (1) ]
[ .. .. .. ] × [ .. ] ,
𝐴(𝜃, 𝜏) is popularly known as ambiguity Function, and [ . . d . ] [ . ]
Φ(𝜃, 𝜏) refers to kernel of the distribution, while 𝑟 and 𝑤 are 𝑟 (𝑝 − 1) 𝑟 (𝑝 − 2)𝑥𝑥 ⋅⋅⋅ 𝑟 (0)𝑥𝑥
[ 𝑥𝑥 ] [𝑎 (𝑝)]
time and frequency dummy variables, respectively. (18)
Smooth pseudo-Wigner-Ville (SPWV) distribution is a
variant method which incorporates smoothing by indepen- where 𝑟𝑥𝑥 can be defined by
dent windows in time and frequency, namely, 𝑊𝑤 (𝜏) and
𝑊𝑡 (𝑡) [29]:
1 𝑁−𝑚−1 ∗
𝑟𝑥𝑥 (𝑚) = ∑ 𝑥 (𝑛) 𝑥 (𝑛 + 𝑚) , 𝑚 ≥ 0. (19)
SPWV (𝑡, 𝑤) 𝑁 𝑁=0
∞ ∞
𝜏
=∫ 𝑊𝑤 (𝜏) [∫ 𝑊𝑡 (𝑢 − 𝑡) 𝑥 (𝑢 + ) Calculating the above set of (𝑝 + 1) linear equations, the AR
−∞ −∞ 2 (15) coefficients can be obtained:
𝜏 2
× 𝑥∗ (𝑢 − ) 𝑑𝑢] 𝑒−𝑗𝜔𝜏 𝑑𝜏. 𝐵𝑈
𝜎𝑤𝑝
2 𝑃𝑥𝑥 (𝑓) = 󵄨
󵄨󵄨 󵄨󵄨2 , (20)
󵄨󵄨1 + ∑𝑘=1 𝑎 𝑝 (𝑘)𝑒−𝑗2𝜋𝑓𝑘 󵄨󵄨󵄨
𝑃 _
The feature extraction using this method is based on the 󵄨 󵄨
energy, frequency, and the length of the principal track. Each
segment gives the values 𝐸𝑘 , 𝐹𝑘 , and 𝐿 𝑘 . The EEG signal is while 𝜎̂𝑤𝑝 gives the approximated lowest mean square error
firstly divided into 𝑘 segments; then, the construction of a of the 𝑝th-order predictor given as follows:
three-dimensional feature vector for each segment will take
𝑝
place. Energy of each segment 𝑘 can be calculated as follows: 󵄨 󵄨2
2
𝜎𝑤𝑝 = 𝐸𝑝𝑓 = 𝑟𝑥𝑥 (0) ∏ [1 − 󵄨󵄨󵄨𝑎𝑘 (𝑘)󵄨󵄨󵄨 ] . (21)
∞ ∞
𝑘=1
𝐸𝑘 = ∫ ∫ 𝜗𝑘 (𝑡, 𝑓) 𝑑𝑡 𝑑𝑓, (16)
−∞ −∞

where 𝜗𝑘 (𝑡, 𝑓) stands for the time-frequency representation 2.5.2. Burg’s Method. It is an AR spectral estimation based
of the segment. However, to calculate the frequency of each on reducing the forward and backward prediction errors
segment 𝑘, we make use of the marginal frequency as follows: to satisfy Levinson-Durbin recursion [8]. Burg’s method
estimates the reflection coefficient directly without the need
∞
𝐹𝑘 = ∫ 𝜗𝑘 (𝑡, 𝑓) 𝑑𝑡. (17) to calculate the autocorrelation function. This method has the
−∞ following strength: Burg’s method can estimate PSD’s data
records to look exactly like the original data value. It can
Finally, for achieving good results, noiseless EEG signals yield intimately packed sinusoids in signals once it contains
or a well-denoised signal should be used for TFD [30]. minimal level of noise.
 23567872, 2014, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2014/730218, Wiley Online Library on [19/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ISRN Neuroscience 5

Table 2: Comparison between performances of EEG methods.

Analysis
Method name Advantages Disadvantages Suitability
method
(i) Weakness in analyzing nonstationary
signals such as EEG
(i) Good tool for stationary signal (ii) It does not have good spectral
processing estimation and cannot be employed for
(ii) It is more appropriate for narrowband analysis of short EEG signals Narrowband,
Fast fourier Frequency
signal, such as sine wave (iii) FFT cannot reveal the localized stationary
transform domain
(iii) It has an enhanced speed over spikes and complexes that are typical signals
virtually all other available methods in among epileptic seizures in EEG signals
real-time applications (iv) FFT suffers from large noise
sensitivity, and it does not have shorter
duration data record
(i) It has a varying window size, being
broad at low frequencies and narrow at
high frequencies
Both time and Transient and
Wavelet (ii) It is better suited for analysis of
Needs selecting a proper mother wavelet freq. domain, stationary
transform sudden and transient signal changes
and linear signal
(iii) Better poised to analyze irregular
data patterns, that is, impulses existing at
different time instances
Lowest eigenvalue may generate false
Provides suitable resolution to evaluate Frequency Signal buried
Eigenvector zeros when Pisarenko’s method is
the sinusoid from the data domain with noise
employed
(i) The time-frequency methods are
oriented to deal with the concept of
(i) It gives the feasibility of examining stationary; as a result, windowing process
Time Both time and
great continuous segments of EEG signal is needed in the preprocessing module Stationary
frequency frequency
(ii) TFD only analyses clean signal for (ii) It is quite slow (because of the signal
distribution domains
good results gradient ascent computation)
(iii) Extracted features can be dependent
on each other
(i) AR limits the loss of spectral problems
(i) The model order in AR spectral
and yields improved frequency resolution
estimation is difficult to select
(ii) Gives good frequency resolution
(ii) AR method will give poor spectral
(iii) Spectral analysis based on AR model Signal with
estimation once the estimated model is Frequency
Autoregressive is particularly advantageous when short sharp spectral
not appropriate, and model’s orders are domain
data segments are analyzed, since the features
incorrectly selected
frequency resolution of an analytically
(iii) It is readily susceptible to heavy
derived AR spectrum is infinite and does
biases and even large variability
not depend on the length of analyzed data

The difference between method of Yule-Walker and Burg’s 3. Performance of Methods

method is in the way of calculating the PSD. For Burg’s
method, the PSD is estimated as follows: The general aim of this review is to shed light on EEG
signal feature extraction and to show how fast the method
used for the signal extraction and how reliable it will be the
_
𝐸𝑝 extracted EEG signal features. Moreover, how these extracted
𝐵𝑈
𝑃𝑥𝑥 (𝑓) =󵄨 󵄨󵄨2 . (22) features would express the states of the brain for different
󵄨󵄨
󵄨󵄨1 + ∑𝑘=1 𝑎 𝑝 (𝑘)𝑒−𝑗2𝜋𝑓𝑘 󵄨󵄨󵄨
𝑃 _
mental tasks, and to be able to yield an exact classification
󵄨 󵄨
and translation of mental tasks. The speed and accuracy of
the feature extraction stage of EEG signal processing are
Parametric methods like autoregressive one reduce the therefore very crucial, in order not to lose vital information
spectral leakage issues and yield better frequency resolution. at a reasonable time. So far in the discussed literature, wavelet
However, selecting the proper model order is a very serious method is introduced as a solution for unstable signals; it
problem. Once the order is too high, the output will induce includes the representation by wavelets which are a group of
false peaks in the spectra. If the order is too low, the result will functions derived from the mother wavelet by dilation and
produce smooth spectra [32]. translation processes. The window with varying size is the
 23567872, 2014, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2014/730218, Wiley Online Library on [19/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 ISRN Neuroscience

most significant specification of this method since it ensures [2] X.-Y. Wang, J. Jin, Y. Zhang et al., “Brain control: human-
the suitable time frequency resolution in all frequency range computer integration control based on brain-computer inter-
[26]. Autoregression analysis suffers from speed and is not face approach,” Acta Automatica Sinica, vol. 39, no. 3, pp. 208–
always applicable in real-time analysis while FFT appears to 221, 2013.
be the least efficient of the discussed methods because of its [3] G. Al-Hudhud, “Affective command-based control system inte-
inability to examine nonstationary signals. The strength of grating brain signals in commands control systems,” Computers
AR method can be emphasized by further comparing its in Human Behavior, vol. 30, pp. 535–541, 2014.
performance with that of classical FFT as shown in Table 1. [4] J. L. S. Blasco, E. Iáñez, A. Úbeda, and J. M. Azorı́n, “Visual
evoked potential-based brain-machine interface applications to
It is highly recommend to use AR method in conjunction
assist disabled people,” Expert Systems with Applications, vol. 39,
with more conservative methods, such as periodograms, to no. 9, pp. 7908–7918, 2012.
help to choose the correct model order and to avoid getting
[5] H. Cecotti, “Spelling with non-invasive brain-computer inter-
fooled by spurious spectral features [32]. faces—current and future trends,” Journal of Physiology, vol. 105,
The most important application for eigenvectors is to eva- no. 1–3, pp. 106–114, 2011.
luate frequencies and powers of signals from noise corrupted [6] I. S. Kotchetkov, B. Y. Hwang, G. Appelboom, C. P. Kellner,
signal; the principle of this method is the decomposition of and E. S. Connolly Jr., “Brain-computer interfaces: military,
the correlation matrix of the noise corrupted. Three methods neurosurgical, and ethical perspective,” Neurosurgical Focus,
for eigenvectors module were discussed: Pisarenko, multiple vol. 28, no. 5, article E25, 2010.
signal classification (MUSIC), and minimum norm [27]. The [7] E. D. Übeyli, “Statistics over features: EEG signals analysis,”
good thing about the eigenvector method is that it produces Computers in Biology and Medicine, vol. 39, no. 8, pp. 733–741,
frequency spectra of high resolution even when the signal-to- 2009.
noise ratio (SNR) is low. However, this method may produce [8] R. Agarwal, J. Gotman, D. Flanagan, and B. Rosenblatt, “Auto-
spurious zeros leading to poor statistical accuracy [26]. matic EEG analysis during long-term monitoring in the ICU,”
The TFD method offers the possibility to analyze rela- Electroencephalography and Clinical Neurophysiology, vol. 107,
tively long continuous segments of EEG data even when the no. 1, pp. 44–58, 1998.
dynamics of the signal are rapidly changing. At the same time [9] A. Meyer-Lindenberg, “The evolution of complexity in human
a good resolution both in time and frequency is necessary, brain development: an EEG study,” Electroencephalography and
making this method not preferable to use in many cases [30]. Clinical Neurophysiology, vol. 99, no. 5, pp. 405–411, 1996.
Table 2 shows the summary of advantages and disad- [10] A. Subasi and E. Erçelebi, “Classification of EEG signals using
vantages of the above-mentioned methods, their accuracies, neural network and logistic regression,” Computer Methods and
Programs in Biomedicine, vol. 78, no. 2, pp. 87–99, 2005.
speeds, and suitability to make it easier to compare their
[11] A. Subasi, “EEG signal classification using wavelet feature extra-
performances.
ction and a mixture of expert model,” Expert Systems with Appli-
cations, vol. 32, no. 4, pp. 1084–1093, 2007.
4. Conclusion [12] H. Jasper and L. D. Proctor, Reticular Formation of the Brain,
Five of the well-known methods for frequency domain and 1958.
time-frequency domain methods were discussed. Acclaim [13] M. R. N. Kousarrizi, A. A. Ghanbari, M. Teshnehlab, M. Ali-
about the definite priority of methods according to their yari, and A. Gharaviri, “Feature extraction and classification
capability is very hard. The findings indicate that each of EEG signals using wavelet transform, SVM and artificial
neural networks for brain computer interfaces,” in Proceedings
method has specific advantages and disadvantages which
of the International Joint Conference on Bioinformatics, Systems
make it appropriate for special type of signals. Frequency Biology and Intelligent Computing (IJCBS ’09), pp. 352–355,
domain methods may not provide high-quality performance August 2009.
for some EEG signals. In contrast, time-frequency methods, [14] D. Cvetkovic, E. D. Übeyli, and I. Cosic, “Wavelet transform fea-
for instance, may not provide detailed information on EEG ture extraction from human PPG, ECG, and EEG signal respo-
analysis as much as frequency domain methods. It is crucial to nses to ELF PEMF exposures: a pilot study,” Digital Signal Proce-
make clear the of the signal to be analyzed in the application ssing, vol. 18, no. 5, pp. 861–874, 2008.
of the method, whenever the performance of analyzing [15] H. Kordylewski, D. Graupe, and K. Liu, “A novel large-memory
method is discussed. Considering this, the optimum method neural network as an aid in medical diagnosis applications,”
for any application might be different. IEEE Transactions on Information Technology in Biomedicine,
vol. 5, no. 3, pp. 202–209, 2001.
Conflict of Interests [16] I. Güler and E. D. Übeyli, “Adaptive neuro-fuzzy inference sys-
tem for classification of EEG signals using wavelet coefficients,”
The authors declare that there is no conflict of interests Journal of Neuroscience Methods, vol. 148, no. 2, pp. 113–121,
regarding the publication of this paper. 2005.
[17] E. D. Übeyli, “Wavelet/mixture of experts network structure
References for EEG signals classification,” Expert Systems with Applications,
vol. 34, no. 3, pp. 1954–1962, 2008.
[1] L. Mayaud, M. Congedo, A. van Laghenhove et al., “A compar- [18] A. Subasi, M. K. Kiymik, A. Alkan, and E. Koklukaya, “Neural
ison of recording modalities of P300 Event Related Potentials network classification of EEG signals by using AR with MLE
(ERP) for Brain-Computer Interface (BCI) paradigm,” Clinical preprocessing for epileptic seizure detection,” Mathematical and
Neurophysiology, vol. 43, no. 4, pp. 217–227, 2013. Computational Applications, vol. 10, no. 1, pp. 57–70, 2005.
 23567872, 2014, 1, Downloaded from https://onlinelibrary.wiley.com/doi/10.1155/2014/730218, Wiley Online Library on [19/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
ISRN Neuroscience 7

[19] O. Faust, R. U. Acharya, A. R. Allen, and C. M. Lin, “Analysis

of EEG signals during epileptic and alcoholic states using AR
modeling techniques,” IRBM, vol. 29, no. 1, pp. 44–52, 2008.
[20] A. Procházka, J. Kukal, and O. Vyšata, “Wavelet transform use
for feature extraction and EEG signal segments classification,”
in Proceedings of the 3rd International Symposium on Commu-
nications, Control, and Signal Processing (ISCCSP ’08), pp. 719–
722, March 2008.
[21] I. Daubechies, “Wavelet transform, time-frequency localization
and signal analysis,” IEEE Transactions on Information Theory,
vol. 36, no. 5, pp. 961–1005, 1990.
[22] S. Soltani, “On the use of the wavelet decomposition for time
series prediction,” Neurocomputing, vol. 48, no. 1, pp. 267–277,
2002.
[23] N. Hazarika, J. Z. Chen, A. C. Tsoi, and A. Sergejew, “Classifi-
cation of EEG signals using the wavelet transform,” Signal Proce-
ssing, vol. 59, no. 1, pp. 61–72, 1997.
[24] M. Unser and A. Aldroubi, “A review of wavelets in biomedical
applications,” Proceedings of the IEEE, vol. 84, no. 4, pp. 626–
638, 1996.
[25] E. D. Übeyli, “Analysis of EEG signals by implementing
eigenvector methods/recurrent neural networks,” Digital Signal
Processing, vol. 19, no. 1, pp. 134–143, 2009.
[26] E. D. Übeyli, “Analysis of EEG signals by combining eigenvector
methods and multiclass support vector machines,” Computers in
Biology and Medicine, vol. 38, no. 1, pp. 14–22, 2008.
[27] E. D. Übeyli and I. Güler, “Features extracted by eigenvector
methods for detecting variability of EEG signals,” Pattern Reco-
gnition Letters, vol. 28, no. 5, pp. 592–603, 2007.
[28] S. A. Awang, M. Paulraj, and S. Yaacob, “Analysis of EEG signals
by eigenvector methods,” in Proceedings of the IEEE EMBS
Conference on Biomedical Engineering and Sciences (IECBES
’12), pp. 778–783, December 2012.
[29] C. Guerrero-Mosquera and A. N. Vazquez, “New approach in
features extraction for EEG signal detection,” in Proceedings of
the Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC ’09), pp. 13–16, September
2009.
[30] L. Cohen, Time-Frequency Analysis, Prentice Hall, Upper Saddle
River, NJ, USA, 1995.
[31] P. Stoica and R. L. Moses, Introduction to Spectral Analysis,
Prentice hall, Upper Saddle River, NJ, USA, 1997.
[32] M. Polak and A. Kostov, “Feature extraction in development
of brain-computer interface: a case study,” in Proceedings of the
20th Annual International Conference of the IEEE Engineering
in Medicine and Biology Society (EMBC ’98), pp. 2058–2061,
November 1998.