See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/319409826

Quantification of Human Stress Using Commercially Available Single Channel

EEG Headset

Article in IEICE Transactions on Information and Systems · September 2017

DOI: 10.1587/transinf.2016EDL8248

CITATIONS READS

30 2,011

3 authors:

Sanay Muhammad Umar Saeed Syed Anwar

University of Engineering and Technology, Taxila University of Engineering and Technology, Taxila
14 PUBLICATIONS 237 CITATIONS 141 PUBLICATIONS 4,246 CITATIONS

SEE PROFILE SEE PROFILE

Muhammad Majid
University of Engineering and Technology, Taxila
108 PUBLICATIONS 3,520 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Stress Quantification System View project

Speech Quality Assessment View project

All content following this page was uploaded by Syed Anwar on 04 September 2017.

The user has requested enhancement of the downloaded file.

IEICE TRANS. INF. & SYST., VOL.E100–D, NO.9 SEPTEMBER 2017
2241

LETTER
Quantification of Human Stress Using Commercially Available
Single Channel EEG Headset
Sanay MUHAMMAD UMAR SAEED†a) , Syed MUHAMMAD ANWAR††b) ,
and Muhammad MAJID†c) , Nonmembers

SUMMARY A study on quantification of human stress using low beta Neurophysiologically, the presence of stress has been
waves of electroencephalography (EEG) is presented. For the very first shown to effect the prefrontal cortex and frontal areas of
time the importance of low beta waves as a feature for quantification of hu-
brain. Structural and functional imaging studies have iden-
man stress is highlighted. In this study, there were twenty-eight participants
who filled the Perceived Stress Scale (PSS) questionnaire and recorded tified varying brain regions that are effected due to stress,
their EEG in closed eye condition by using a commercially available sin- but prefrontal cortex appears to be a common area in most
gle channel EEG headset placed at frontal site. On the regression analysis subjects [8]. The glucose metabolism and blood flow both
of beta waves extracted from recorded EEG, it has been observed that low are effected in the frontal brain regions during stress. The
beta waves can predict PSS scores with a confidence level of 94%. Con-
sequently, when low beta wave is used as a feature with the Naive Bayes
brain activity has been shown to vary with the behaviour
algorithm for classification of stress level, it not only reduces the computa- of normal and depressed subjects in the prefrontal cortex
tional cost by 7 folds but also improves the accuracy to 71.4%. area. Antidepressant treatment medications also effect the
key words: wearable headset, electroencephalography, chronic stress, re- prefrontal cortical activities [8], [9]. Among different neu-
gression, classification
ral oscillations, higher level of beta waves act as a marker
denoting that a person is not in a calm state [10]. The pres-
1. Introduction ence of stress has been shown to increase the spectral power
in EEG beta band [11], [12]. Prefrontal relative gamma has
Psychological stress is a serious health issue and can reduce also been used for human stress assessment using Montreal
the working ability of an individual irrespective of age [1]. imaging stress task (MIST). The experimental design relies
Quantification of stress using different sensors has been at- on the stressor and by using the stressed and calm states as-
tracting a number of researchers [2]. Among other sensors, sess human stress [9].
electroencephalography (EEG) sensors provide good tem- In literature, beta waves are considered significant to
poral resolution and mobility. An increasing level of stress predict stress [10]–[12], and in this letter, for the very first
or fatigue could change four rhythms of EEG [3]. In most time we highlight the importance of low beta waves as a
cases, any reliable measure of stress using EEG requires a feature for quantification of human stress by using a single
dense placement of EEG electrodes [4]. In order to observe channel headset placed at frontal site; a preliminary experi-
stress in daily life activities, wearable EEG is required that is mental work for quantification of human stress using brain
unobtrusive, mobile and could be used for longer durations. waves was reported in [11]. In particular, an experiment was
Clinical dense EEG systems with 128−, 64− or 32− chan- designed to gather data of twenty-eight participants. PSS
nels do not fulfil these requirements. On the other hand, questionnaire was filled by each participant. PSS scores
single channel EEG headset has been used effectively for were calculated on the scale of 0 to 40. Participants were
emotion recognition [5], and following this a single channel labelled as stressed and not stressed on the basis of PSS
EEG dry sensor will be used to detect psychological stress. scores. EEG was recorded in closed eye condition to obtain
This will have lesser computational cost and provide a wear- baseline EEG signal. These signals were then transformed
able solution. The effect on the frontal site of brain caused into frequency subbands and we applied regression analy-
by a neuro-chemical change in response to stress has been sis over beta waves to predict the PSS scores. P value of
established [6]. Perceived Stress Scale (PSS) [7], is a widely low beta waves clearly indicated that it is significant at the
used and validated measure for chronic mental stress level. confidence level of 94% to predict stress level in humans.
Consequently, when low beta wave is used as a sole clas-
Manuscript received December 20, 2016.
Manuscript revised April 18, 2017. sification feature, it gives an accuracy of 71.43% with the
Manuscript publicized June 2, 2017. Naive Bayes algorithm. Moreover, the feature vector is re-
† duced by seven folds without compromising the classifica-
The authors are with Department of Computer Engineering,
University of Engineering and Technology, Taxila Pakistan. tion accuracy [11].
††
The author is with Department of Software Engineering, Uni-
versity of Engineering and Technology, Taxila Pakistan. 2. Experimental Arrangement
a) E-mail: [email protected]
b) E-mail: [email protected]
c) E-mail: [email protected] A block diagram of the major steps followed in the study is
DOI: 10.1587/transinf.2016EDL8248 shown in Fig. 1. A total of twenty-eight participants (66%

Copyright 
c 2017 The Institute of Electronics, Information and Communication Engineers
 IEICE TRANS. INF. & SYST., VOL.E100–D, NO.9 SEPTEMBER 2017
2242

Fig. 1 Steps involved in this study for human stress quantification using Fig. 3 Neurosky single channel EEG headset.
a commercially available single channel Neurosky Mindwave headset.

brain, referenced to an electrode placed at the ear lobe. The

device uses Thinkgear application specific integrated circuit
module (TGAM) dry electrode technology that operates at
a minimum of 2.7V and covers a bandwidth of 3 − 100Hz.
The TGAM electrodes are made of silver as shown in Fig. 3
and are suitable for non hairy areas [13]. The device was
configured at a sampling rate of 512 Hz to record discrete
EEG data. The subjects were instructed to close their eyes
and to keep their mind free of distracting thoughts. The
wearable headset was individually setup on all subjects and
data was recorded for three minutes in closed eye condi-
tion. The recorded data from the device was transmitted via
Bluetooth. All data recordings were done in a room with
similar lighting conditions and peaceful environment, that
Fig. 2 PSS Scores of participants involved in the study (the red bar shows
participants labelled with higher stress). would not induce any additional stress. Frequency domain
analysis of the EEG data was done by applying fast Fourier
transform. The data was filtered into low beta (13–17 Hz)
male and 33% female) were involved, with age in the range and high beta (18–30 Hz) bands. Multiple linear regression
22–33 years and a mean of 27.5 years. The experimental is used to explain the relationship between a dependent vari-
design does not rely on an induced stressor like MIST [9], able from two or more independent variables. In our case,
and is proposed to monitor chronic stress. All participants the independent variables in this analysis are continuous. In
selected for the study were involved with education either the first step, multiple linear regression is applied over fre-
as a student or as instructor and no mental illness was re- quency subbands of low beta (13–17 Hz) and high beta (18–
ported. The selected questionnaire for this particular study 30 Hz) to predict the PSS scores of a participant. Mathemat-
was PSS-10, that comprised of ten questions. It records ically, multiple linear regression can be expressed as:
how frequent a stress related event occurred during the past
PS S = b0 + b1 βL + b2 βH , (1)
thirty days. The questions can be answered in a range of
0 to 4, where 0 represents that the event never occurred where, PSS score is a dependent variable, b0 , b1 and b2 are
and 4 represents a very frequent occurrence. The total re- regression weights, βL is the value of low beta wave and βH
sponse from PSS questionnaire directly correlates with the is the value of high beta wave. The Naive Bayes algorithm
perceived stress level, where a higher score means a higher is used for classification of stress levels as it takes less com-
level of mental stress. This questionnaire can be used with putational time compared to support vector machine and
subjects that have a minimum education level equivalent to multi-layer Perceptron [11]. It is a nonlinear classifier and
junior high school. The participants involved in the study gives good results in real-world problems. In addition, it
were asked to fill in this questionnaire and all of them ful- only requires a small amount of training data to approximate
filled the basic aptness requirements. The analysis of the the statistical parameters like mean and variance, which is
responses recorded showed that a mean and standard de- essential for classification. Because independent variables
viation of the PSS of total responses was μ = 17.6 and are assumed in each class, only the variances of the vari-
σ = 4.04, respectively. For this study, all participants with ables are required instead of the whole covariance matrix of
PSS scores below threshold T = μ + σ2 = 19.6, were con- model [14].
sidered to be relatively stress free and those above T were
considered to be relatively stressed. Figure 2 shows the PSS 3. Experimental Results
scores for all participants. The EEG data of the subjects
was recorded with commercially available Neurosky Mind- Results of multiple linear regression applied over frequency
set device, that provides a single channel of EEG recording subbands of beta waves in terms of coefficients, standard er-
from a dry electrode placed at the frontal location (FP1) of ror, t-stat and P-value are given in Table 1. From Table 1, it
LETTER
2243

can be observed that the P-value of high beta wave is con- over, the Kappa statistics has a non-negative value of 0.2
siderably higher than that of low beta wave. A low P value when low beta wave is used as a feature; the F-measure
of low beta wave indicates that a change in the value of is improved from 0.531 to 0.662 that clearly indicates that
beta wave will change the PSS score considerably. It can the precision and recall of stress classification system is im-
be seen that at the confidence level of 94%, the P-value of proved. Kappa statistic is also improved to a positive value
low beta wave is acceptable for prediction of PSS scores. of 0.2 from negative value of 0.6. Values of error parameters
R Squared has a low but non-negative value of 0.134. Even like Root Mean Squared Error (RMSE) and Root Absolute
when R Squared is low, smaller P value of low beta wave in- Error (RAE) are reduced, while computational time is re-
dicates its relationship with PSS score. Another reason for duced by a factor of 7 in case of low beta waves. In compar-
low R Squared is less number of predictors in our regression ison, human stress has been assessed using prefrontal rela-
model. The value of significance F is 0.16, which indicates tive gamma by inducing stress [9], whereas, in the proposed
with 84% of confidence level that overall regression results method the baseline stress is identified without using a well
are not by chance. defined stressor such as MIST.
The residual scatter plot for regression analysis is
shown in Fig. 4. It is evident from Fig. 4 that residuals have 4. Conclusion
no pattern, which shows the linear marginal relationship be-
tween dependent and independent variables. The regression It has been shown that low beta waves can be used as a fea-
plot for normalised low beta power and PSS score is shown ture for effective quantification of human stress by using a
in Fig. 5. The trend shows that the normalised low beta single channel headset placed at the frontal site. It not only
power increases with the PSS score. A persistence of high reduces the computational time but also improves the accu-
PSS scores increases the risk of depression. These results racy of the Naive Bayes algorithm by 7.14%. On the re-
follow that low beta wave can be used as an effective fea- gression analysis of beta waves, it has been observed that
ture for classification of stress level in humans. Our result low beta waves can predict PSS scores at a confidence level
suggests that variability in the normalized beta power can of 94%, and the significance F shows with 84% confidence
differentiate normal from prodromal stages of stress. that these results are not by chance. The Naive Bayes al-
Table 2 gives a comparison of performance parameters gorithm has shown considerable improvement in terms of
for the Naive Bayes algorithm. It can be observed that the performance parameters, when low beta wave is used as a
Naive Bayes algorithm when used with only low beta waves feature vector. F-measure shows the precision and recall of
as a feature gives 7.14% higher accuracy as compared to the the system has significant values for the quantification of the
case when all brain waves used as a feature vector. More- stress level. The advantage of using low beta wave as a sole
feature with the Naive Bayes algorithm is that it can monitor
Table 1 Parameters related to multiple linear regression.
stress of a person in daily life activities with lesser computa-
tional time and is intended to identify the stressors that can
Coefficients Standard Error t-stat P-value
Intercept −2.35 13.30 −0.17 0.86 affect the beta activity of a person. With the availability of
Low beta 2.29 × 10−6 1.20 × 10−6 1.89 0.06 commercial single channel EEG headset, it will lead to po-
High beta 1.20 × 10−7 1.33 × 10−6 0.09 0.92 tential real time applications in both monitoring and iden-

Fig. 4 Residual scatter plot for regression analysis. Fig. 5 Regression plot for normalised low beta power and PSS score.

Table 2 Comparison of performance parameters of the Naive Bayes algorithm when all brain waves
and only low beta wave is used as features.
Brain Accuracy Kappa F-Measure Time RMSE RAE
waves
All 64.29% −0.68 0.531 0.032 0.59 80.76
Low beta 71.43% 0.2 0.662 0.004 0.52 65.29
 IEICE TRANS. INF. & SYST., VOL.E100–D, NO.9 SEPTEMBER 2017
2244

tification of stressed individuals. Although, the wearable [6] B.S. McEwen and J.H. Morrison, “The brain on stress: vulnerability
system should be unobtrusive and the real time data record- and plasticity of the prefrontal cortex over the life course,” Neuron,
ings that are in an uncontrolled (out of lab) environment are vol.79, no.1, pp.16–29, 2013.
[7] S. Cohen, T. Kamarck, and R. Mermelstein, “A global measure of
effected by artifacts reducing the quality of observed EEG. perceived stress,” Journal of health and social behavior, vol.24, no.4,
These limitations need to be addressed before the system pp.385–396, 1983.
can be adopted for daily life use. [8] H.S. Mayberg, “Targeted electrode-based modulation of neural cir-
cuits for depression,” The Journal of clinical investigation, vol.119,
References no.4, pp.717–725, 2009.
[9] J. Minguillon, M.A. Lopez-Gordo, and F. Pelayo, “Stress assessment
[1] S.J. Lupien, B.S. McEwen, M.R. Gunnar, and C. Heim, “Effects of by prefrontal relative gamma,” Frontiers in Computational Neuro-
stress throughout the lifespan on the brain, behaviour and cognition,” science, vol.10, 2016.
Nature Reviews Neuroscience, vol.10, no.6, pp.434–445, 2009. [10] S. Sanei and J.A. Chambers, EEG signal processing, John Wiley &
[2] N. Sharma and T. Gedeon, “Objective measures, sensors and com- Sons, 2013.
putational techniques for stress recognition and classification: A [11] S.M.U. Saeed, S.M. Anwar, M. Majid, and A.M. Bhatti, “Psycho-
survey,” Computer methods and programs in biomedicine, vol.108, logical stress measurement using low cost single channel eeg head-
no.3, pp.1287–1301, 2012. set,” 2015 IEEE International Symposium on Signal Processing and
[3] S.A. Hosseini, M.A. Khalilzadeh, M.B. Naghibi-Sistani, and V. Information Technology (ISSPIT), pp.581–585, IEEE, 2015.
Niazmand, “Higher order spectra analysis of eeg signals in emo- [12] N.H.A. Hamid, N. Sulaiman, Z.H. Murat, and M.N. Taib, “Brain-
tional stress states,” Information Technology and Computer Science waves stress pattern based on perceived stress scale test,” 2015 IEEE
(ITCS), 2010 Second International Conference on, pp.60–63, IEEE, 6th Control and System Graduate Research Colloquium (ICSGRC),
2010. pp.135–140, IEEE, 2015.
[4] R.S. Lewis, N.Y. Weekes, and T.H. Wang, “The effect of a naturalis- [13] M.A. Lopez-Gordo, D. Sanchez-Morillo, and F.P. Valle, “Dry eeg
tic stressor on frontal eeg asymmetry, stress, and health,” Biological electrodes,” Sensors, vol.14, no.7, pp.12847–12870, 2014.
psychology, vol.75, no.3, pp.239–247, 2007. [14] I.H. Witten and E. Frank, Data Mining: Practical machine learning
[5] A.M. Bhatti, M. Majid, S.M. Anwar, and B. Khan, “Human emo- tools and techniques, Morgan Kaufmann, 2005.
tion recognition and analysis in response to audio music using brain
signals,” Computers in Human Behavior, vol.65, pp.267–275, 2016.

View publication stats