International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656

Feature Extraction Techniques and Classification

Algorithms for EEG Signals to detect Human Stress - A
Review
Chetan Umale

Amit Vaidya

Shubham Shirude

Akshay Raut

MIT College of
Engineering,

MIT College of
Engineering,

MIT College of
Engineering,

MIT College of
Engineering,

Pune, India

Pune, India

Pune, India

Pune, India

Abstract: EEG (Electroencephalogram) signal is a neuro signal which is generated due the different electrical activities in the brain.
Different types of electrical activities correspond to different states of the brain. Every physical activity of a person is due to some
activity in the brain which in turn generates an electrical signal. These signals can be captured and processed to get the useful information
that can be used in early detection of some mental diseases. This paper focus on the usefulness of EGG signal in detecting the human
stress levels. It also includes the comparison of various preprocessing algorithms ( DCT and DWT.) and various classification algorithms
(LDA, Naive Bayes and ANN.). The paper proposes a system which will process the EEG signal and by applying the combination of
classifiers, will detect the human stress levels.
Keywords: Stress, DWT, KNN, LDA, Nave Bayes, EEG signal, NeuroSky Mindwave.

1.

INTRODUCTION

Stress is generally defined as a response of a person

to the environmental demands or pressures. It results from
interaction between a person and his/her environment that are
perceived as straining or exceeding their adaptive capacities
and threatening their well-being. It can be understood from
above definition that stress is part and parcel of todays life
style. If ignored, stress can lead to chronic diseases. The risk
factors for stress related diseases are a mixture of personal,
interpersonal and social variables. Thus, it can affect various
phases of our life. So it is necessary to detect stress at an early
stage and take appropriate measures.
Now, the question is, Is it possible to detect stress at
early stages? Yes, it is. This is done by many psychologists or
counselors. But it requires active participation from the person
seeking counseling. This might not be possible in some cases
when a stressed person is unable to express himself frankly. It
makes the job of a counselor difficult . This problem can be
solved, if the brain signals are recorded and analyzed to detect
stress.
Brain signals are neuron signals. The electrical
activity of the neurons inside the brain cause electric potential
to be generated across different parts of the brain. The
difference between these electric potential levels can be
captured and used for various applications including stress
detection. These brain signals are called as EEG signals Electroencephalogram signal. Different types of states of the
brain are due to different types of electrical activities of brain
neurons. So, different signal values correspond to different
mental states.

www.ijcat.com

These signals can be captured using various available

equipments which generally consists of electrodes which are
placed on the scalp with a conductive gel between the
electrodes and the scalp. Electrodes are placed at different
positions on the scalp which capture the signals from different
parts of the brain. Raw EEG signals cannot be used directly
for stress detection. Pre-processing is required to extract useful
features which can further used with various machine learning
algorithms.
The aim of the paper is to review various feature
extraction techniques and classification algorithms which can
be used for detection of stress levels. Based on the review, a
system is proposed which will use a single electrode EEG
headset(Neurosky MindWave) to record raw EEG signals
which will be pre-processed using Discrete Wavelet
Transform(DWT) and classified using a combination of
classifiers approach to detect stress levels. Section 1 gives an
introduction on how EEG signals can be used in detection of
stress. Section 2 contains literature survey. Section 3 contains
the proposed system for stress detection. Section 4 is
conclusion and Section 5 includes future scope.

2.
LITERATURE SURVEY
2.1 EEG
Electroencephalography (EEG) is nothing but
recorded electrical activity generated by brain [2].The first
report on electrical brain activity in humans was published in
1929 which allowed doctors and scientists to observe the brain
in action in a meaningful way [5]. There are millions of neurons
in our brain. These activities generate millions of small electric
voltage fields. The aggregate of these voltage fields can be
detected by electrodes placed on the scalp. Thus we can say
that, EEG is the superposition of many smaller signals. The

International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656
amplitude of these signals ranges from 1 V to 100 V in a
normal person.
The different electrical frequencies in EEG can be
associated with different physical actions and mental states [3].
So EEG shows a wide variations in amplitude depending on
external stimulation and different internal mental states. The
different frequency bands are Delta, theta, alpha, beta and
gamma. Frequency bands associated with the different mental
states are given in the Table 1.
Table 1: EEG frequency bands
Brainwave
type

Frequency
range(Hz)

Delta

0.1 to 3

Theta

4 to 7

Alpha

8 to 12

Low Beta

12 to 15

Midrange
Beta

16 to 20

High Beta

21 to 30

Alertness, agitation

30 to 100

Motor functions, higher mental

activity

Gamma

[7]. A mapping of these electrode positions according to the 1020 international system is shown in the Figure 1.

This system is used for multichannel electrode system which

are generally used for research purpose and are complex and
not portable. Alternative to such systems is single electrode
system which uses single channel for recording signal and are
simpler and portable. An example of such system is NeuroSky
MindWave headset. The comparison between the signals
captured by single channel NeuroSky MindWave and
multichannel Biopac system is. The red line is Biopac and a

Mental states and conditions

Deep, dreamless sleep,

unconscious
Intuitive, creative, recall,
fantasy
Relaxed but not drowsy,
tranquil
Formerly SMR, relaxed yet
focused
Thinking, aware of self &
surrounding
Figure 2: Electrode positions in 10-20 international
system

2.2 SIGNAL ACQUIRING METHODS

Various methods are used for acquiring EEG signals.
They differ in the way the electrodes are placed. The methods
can be categorized as follows:
a) Invasive: Invasive EEG recordings are those recordings that
are captured with electrodes that are placed on the surface or
within the depth of the brain[6]. These type of methods are
generally used in medical surgeries or implants. Again the two
types of electrodes used in this method are

Figure 1: Signals captured by Mindwave (blue) and Biopac

(red)
blue line is NeuroSky.

2.3 MATLAB
i) Subdural EEG electrodes: Subdural EEG electrodes are
the electrodes which sit over the surface of the brain. The
placement of these electrodes is often confirmed with coregistration on an MRI scan image.
ii) Depth EEG electrodes: Depth EEG electrodes are those
which are placed within the substance of the brain.
b) Non-invasive: Non-invasive EEG recordings are those that
are captured with electrodes that are placed on the scalp rather
than placing it on the surface or within the depth of the brain.
Electrodes used here are small metal discs which are made of
stainless steel, tin, gold or silver covered with a silver chloride
coating. They are placed on the scalp in special positions. These
positions are specified using the International 10/20 system.
Each electrode site is labeled with a letter and a number. The
letter refers to the area of brain underlying the electrode e.g. FFrontal lobe and T - Temporal lobe. Even numbers denote the
right side of the head and odd numbers the left side of the head

www.ijcat.com

MATLAB (Matrix Laboratory) is a multi-paradigm

numerical computing environment and high level programming
language developed by Math works. MATLAB allows matrix
manipulations, plotting of functions and data, implementation
of algorithms, creation of user interfaces, and interfacing with
programs written in other languages, including C, C++, Java,
Fortran and Python[9].EEG signals being electrical signals are
vulnerable to outside interference and artifacts. Using signal
processing capabilities of MATLAB, these problems can be
resolved effectively. MATLAB provides an interactive toolbox
called EEGLAB, which is effective for continuous and eventrelated EEG signals analysis using independent component
analysis (ICA), time/frequency analysis (TFA), as well as
standard averaging methods [10]. EEGLAB supports loading
of existing EEG datasets as well as real time EEG datasets
through software like Neuroscan.

International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656
transient nature. This transform when applied to EEG signal
will reveal features that are transient in nature [8].

2.4 Pre-processing
EEG signals are non-stationary and non-linear. EEG
signals are susceptible to noise and interference caused by eye
movement and muscle movement. The electronic devices in the
vicinity can also cause interference. Also, the amount of raw
data required for classification is impractical for most machine
learning algorithms. Thus feature extraction is necessary for
successful
classification.
Pre-processing
includes
transformation of EEG signals from time domain to frequency
domain and removal of noise and artifacts.
A variety of feature extraction methods exist for BCI
applications, such as Discrete Cosine Transform (DCT),
Discrete Wavelet Transform (DWT).

Discrete wavelet transform depends on low pass

filter g and high pass filter h. The working is based on two
important functions namely wavelet function i,l(k) and scale
function i,l(k) which can be defined as :
i,l(k)=2i/2 gi(k-2il)
i,l(k)=2i/2 hi(k-2il)
where the factor 2i/2 is an inner product normalization, i and l
are the scale parameter and the translation parameter,
respectively. The DWT decomposition is described as:
a(i)(l)=x(k)*i,l(k)

2.4.1 Discrete Cosine Transform (DCT)

d(i)(l)=x(k)*i,l(k)
Discrete Cosine Transform is a method to convert
time series signals into frequency components. In context of
BCI, DCT is used to calculate maximum, minimum and mean
value of EEG signal. The one-dimensional DCT for a list of N
real numbers is expressed by the following formula:
2

(2x+1).
)
2

() = () 1
x=0 ()cos (
where
1
() = 2 if j=0

() = 1 if 0
The input is a set of N data values (EEG samples) and
the output is a set of N DCT transform coefficients Y(u). The
first coefficient Y(0) is called the DC coefficient and it holds
average signal value.. The rest coefficients are referred to as the
AC coefficients [11]. DCT produces concentrated signals,
where the energy is concentrated into few coefficients. Thus
DCT is effective for data compression which leads to reduced
size of input vector for machine learning algorithms and also
reduces time required for learning.

2.4.2

where a(i) (l) is the approximation coefficient and d(i)(l) is the

detail coefficient at resolution i.
The DWT decomposition of the input signal into different
frequency bands is obtained by consecutive high-pass and lowpass filtering of the time domain signal. This decomposition is
shown in the Figure 3.
In the given diagram, x[n] is the mother wavelet, h[n] is the
high pass filter and g[n] is the low pass filter. EEG signals do
not have any useful frequency components above 30 Hz [8].So
the decomposition levels can be selected as 5. So the final
relevant wavelet decomposition will be obtained at level
A5.Sample EEG signal decomposition is shown in figure 4.

Discrete Wavelet Transform (DWT)

Wavelet transform is the process of expressing any

general function as an infinite series of wavelets. The main idea
behind wavelet analysis is of expressing a signal as a linear
combination of the particular set of functions by shifting and
expanding the original wavelet (mother wavelet). This
decomposition gives a set of coefficients called as wavelet
coefficients, due to this the signal can be reconstructed as a
linear combination of the wavelet functions weighted by the
wavelet coefficients. The main feature of the wavelets is that
most of their energy is restricted to a finite time interval .This
is called as time-frequency localization. Frequency localization
means that the Fourier transform is band limited. This timefrequency localization provides good frequency localization at
low frequencies and good time localization at high frequencies.
This produces segmentation of the time-frequency plane that is
appropriate for most physical signals, especially those of a

www.ijcat.com

10

International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656

Figure

4: Decomposition of EEG signal using DWT

A commonly used distance metric is the Euclidean

distance. The Euclidean distance of two points or tuples, say,
X1 = (x11, x12, , x1n) and X2 = (x21, x22, , x2n) , is

Figure 3: Decomposition of signal using DWT

2.5

CLASSIFICATION

2.5.1

K-Nearest Neighbor (KNN)

K-nearest neighbor is an instance-based, lazy and supervised

learning algorithm. KNN is a simple algorithm that stores all
available cases and classifies new cases based on a similarity
measure. KNN is a non-parametric method that classifies the
data by comparing the training data and testing data based on
estimating the feature values[12]. These feature values are
calculated by using the distance function such as Euclidean
distance which is not difficult if the given parameter values are
numeric. An object is then classified by the majority vote of its
neighbors, with the object being assigned to the class most
common among its k nearest neighbors. The value of the k
refers to how many nearest values should be considered before
the output class is decided.

www.ijcat.com

where, x1i and x2i represents the training and testing data
respectively[13]. Different attributes are measured on different
scales, so if the Euclidean distance formula is used directly, the
effect of some attributes might be completely dwarfed by others
that have larger scales of measurement.
After feature extraction process the EEG training
data and test data is passed to the classification process. Then
Euclidean distance is calculated between each EEG training
sample and testing sample. The class for first K neighbors is
considered and the majority vote is the classified class. The
accuracy for the KNN is high as compared to the other
classifiers.

2.5.2

Linear Discriminant Analysis (LDA)

Linear discriminant analysis (LDA) is one of the most popular

classification algorithms for Brain Computer Interface
applications, and has been used successfully in a large number
of systems. LDA linearly transforms data from high
dimensional space to low dimensional space. Finally, the
decision is made in the low dimensional space. Thus the
definition of the decision boundary plays an important role in
classification process. During this process, the class
distributions having some finite variance will be still kept in
the projected space. Hence, we assume that if the mean and
variance of the projected data is considered for the calculation
of the decision boundary, it may extend LDA method to deal

11

International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656
with the practical heteroscedastic distribution data, which
derives Z-LDA.
Consider the case of two classes (x11, x12, x13, ,
x1m) C1 and (x21, x22, x23, , x2n) C2 where m
and n being number of training samples.
X=(x11, x12, x13, , x1m, x21, x22, x23,, x2n)
be our input sample.
Calculate weight sum y(X) by, y(X)

~) =
~
~
(
where WT is weight vector.
Now, the parameters related to Gaussian distribution mean()
and standard deviation() are calculated as :-

where 1, 2 and 1, 2 and mean and standard deviations of

two training set samples CK (K=1,2).
During classification process, when any sample X is
input, first calculate weight sum y(x) and then perform
following normalization procedure:z1=(y(x) - 1)/ro1
z2=(y(x) - 2)/ro2
where z1 and z2 are z-scores to calculate how much weight
sums of given input sample is close to the training samples.
Finally, the larger value among these z scores gives the final
classification class for the input. That means, if z1>z2, the
sample is going to be classified in class c1; otherwise c2.

2.5.3

Figure 5: Proposed System Flow

When dealing with continuous data generated from
EEG signals, a typical assumption is that the continuous values
associated with each class are distributed according to a
Gaussian distribution. First the data is segment by the class, and
then compute the mean and variance of x in each class. Let c
be the mean of the values in x associated with class c, and let
c2 be the variance of the values in x associated with class c.
Then, the probability distribution of some value given a class,
p(x = v|c), can be computed by plugging v into the equation for
a Normal distribution parameterized by c and c2. That
equation is :

() =

Naive Bayes

1
2

()2

22

Probability can be interpreted from two views:

Objective and Subjective. The Subjective probability is called
as Bayesian Probability. Bayesian Probability is the process for
using the probability for predicting the likelihood of certain
events occurring in the future. Naive Bayes is a conditional
probability model where Bayes theorem is used to infer the
probability of hypothesis under the observed data or evidence
[14]. Bayes theorem states that

posterior=

www.ijcat.com

prior likelihood
evidence

12

International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656
The captured signal contains various EEG frequency bands.
The frequency range relevant for stress level detection is 4 to
40 Hz [2]. Thus, only these frequencies are extracted. This
process is called feature extraction. In the proposed system,
DWT is used to preprocess the data as it has the unique
advantage of time-frequency localization [8].
Once the preprocessing is done, next task is to
classify the signal into appropriate stress level. When a single
classifier is used, it is difficult to identify the misclassification
error. The solution to this problem is to develop a more
complex classifier with a little misclassification rate. This
approach may seem promising but it will make the system
complex. Another approach is to use combination of simple
classifiers rather than using one complex classifier. One more
advantage of this approach is that even if one of the classifier
misclassifies the data, the other classifiers can rectify this error.

Figure 6: NeuroSky Mindwave headset

3.
PROPOSED SYSTEM3.1 Data
collection equipment
Here the data required
is the EEG signal. In the proposed system, the equipment to be
used in the data collection is NeuroSky Mindwave headset
which uses single electrode for capturing the EEG signal. There
are various noise sources in this process such as muscle
movement or any electrical device in the vicinity. Primary
filtering of these noise signals is done by the ear clip which acts
as reference or ground.

In the proposed system, an approach of combination

of classifiers is used. The classifiers which are to be uses are
KNN, Naive Bayes and LDA. The output class will be decided
by voting and the class which gets majority of votes will be the
output class.

4.

CONCLUSION

Early detection of stress can help in prevention of

chronic mental illness. Recording and analyzing EEG signals
can be an effective tool for this purpose. Different feature
extraction techniques and classification algorithms for EEG
signal analysis are discussed in this paper. This review suggests
use of single channel EEG headset which provides a portable
and affordable alternative to traditional multichannel
equipments.

5.

3.2 System flow

FUTURE WORK

The paper proposed the methodology for detecting

human stress level in real time. As mentioned, early detection
and treatment of stress is important. Music therapy is a good
option for stress treatment. Music with appropriate frequency
can be played automatically corresponding to detected stress
level. Another future application of EEG signal can be a
biometric authentication system, as pattern of EEG signal
captured from every person is unique.

Figure 6 shows the general flow of the proposed

system. EEG signal is captured using the NeuroSky Mindwave.
Processing is done in the MATLAB environment using the
EEGLAB interface. EEGLAB provides functions to capture
data in numerous scenarios such as at different sampling rate.

www.ijcat.com

13

International Journal of Computer Applications Technology and Research

Volume 5 Issue 1, 08 - 14, 2016, ISSN:- 23198656
[4] D. Puthankattil Subha, Paul K. Joseph, Rajendra Acharya
U, Choo Min Lim, EEG Signal Analysis: A Survey ,J
Med Syst (2010) 34:195212
[5] D. A. Kaiser. What is quantitative EEG.
[Online].Available:http://www.skiltopo.com/skil3/whatis-qeeg-by-kaiser.pdf
[6] https://my.clevelandclinic.org/services/neurological_insti
tute/epilepsy/diagnostics-testing/invasive-eeg-monitoring
[7] http://www.medicine.mcgill.ca/physio/vlab/biomed_sign
als/eeg_n.htm
[8] Abdulhamit Subasi, EEG signal classification using
wavelet feature extraction and a mixture of expert
model,Expert Systems with Applications 32 (2007)
10841093,2006 Elsevier Ltd.
[9] https://en.wikipedia.org/wiki/MATLAB
[10] http://sccn.ucsd.edu/eeglab/
[11] Darius Birvinskas,Vacius Jusas,Ignas Martiius, Robertas
Damaeviius, Data Compression of EEG Signals for
Artificial Neural Network Classification,ISSN 2335
884X (online) INFORMATION TECHNOLOGY AND
CONTROL, 2013, Vol.42, No.3

Figure 7: Combination of classifiers

6.

REFERENCES

[1] http://www.stress.org.uk/what-is-stress.aspx
[2]

NeuroSky Inc. Brain Wave Signal (EEG) of NeuroSky,

Inc.(December
15,
2009).
[Online].
Available:http://frontiernerds.com/files/neurosky-vsmedical-eeg.pdf

[3]

Erik Andreas Larsen, Classification of EEG Signals in a

Brain-Computer
Interface
System
,Norwegian
University of Science and Technology Department of
Computer and Information Science,June 2011

www.ijcat.com

[12] Tatiur Rahman, Apu Kumer Ghosh, Md. Maruf Hossain

Shuvo, Md. Mostafizur Rahman, Mental Stress
Recognition using K-Nearest Neighbor(KNN) Classifier
on EEG Signals, International Conference on Materials,
Electronics & Information Engineering, ICMEIE-2015.
[13] Chee-Keong Alfred Lim and Wai Chong Chia, Analysis
of Single-Electrode EEG Rhythms Using MATLAB to
Elicit Correlation with Cognitive Stress, International
Journal of Computer Theory and Engineering, Vol. 7, No.
2, April 2015
[14] Juliano Machado Alexandre Balbinot, Adalberto Schuck,
A study of the Naive Bayes classifier for analysing
imaginary movement EEG signals using the Periodogram
as spectral estimator, Bio-signals and Bio-robotics
Conference (BRC), 2013 ISSNIP, IEEE conference
publication.

14