IJCSNS International Journal of Computer Science and Network Security, VOL.24 No.

11, November 2024 113

Voting Classifier based Model for Mental Stress Detection and

Classification Using EEG Signal
Navdeep Shakya, Rahul Dubey, Laxmi Shrivastava
[email protected]
Department of Electronics Engineering, Madhav Institute of Technology & Science, Gwalior, India

brain neurons are used to calculate the voltage variation

Summary caused by the ionic current. EEG is used for identifying a
According to World Health Origination (WHO), stress is a major stroke, a focal brain disorder, and tumors as a first-line
problem of human beings which have a large effect on physical method [6]. EEG will track the long-term sleep stage or
and mental health. The state of emotional tension from adverse or epilepsy near the bed of the patient's bio signals [6]. During
demanding circumstances is called stress. It can be experienced by various phases of life, EEG may detect brain changes
each person in regular lifestyle due to job, some family problems,
and other personal issues. Some kind of stress is important for
without distracting a patient, e.g., EEG Sleep Analysis [1].
completing the task but a lot of stress causes harm to human health. EEG signal is classified into two types; healthy EEG signal
Hence, nowadays, identification of stress levels is important. This and unhealthy EEG signal. Abnormal activity of the brain
paper proposes one of the simplest methodology for the detection can be determine using unhealthy EEG signal.
of stress by the analysis of EEG signal. Fast fourier transform Classification of EEG signals into respective classes of
(FFT) is used in the proposed method to generate the power healthy and unhealthy EEG has already been explored by
spectral density (PSD) vector. PSD vector is used as an input to various researchers in the literature [7].
the voting classifier. Voting classifier is a combination of K-
nearest neighbour (KNN), and random forest (RF). The proposed Sub-band power ratio is used to classify the EEG
method achieves the highest classification accuracy of 88%.
signals into relaxed and stress classes of EEG signals [8].
Keywords:
Mental Stress, EEG, KNN, Random Forest, Voting classifier.
Artificial neural network based approaches has been used in
next stage to classify the EEG signals into two respective
classes. 10-fold cross-validation method was utilized for the
1. Introduction performance evaluation of the proposed features using
support vector machine (SVM) and K-nearest neighbor. In
The human body's response to unusual mental [9]; author proposed that beta brain waves are highly
conditions is considered as stress [1]. In healthy people, recommended to detect the stress using EEG signals.
there is a balance between parasympathetic and sympathetic Authors consider various stress level to detect. Average
arms of the autonomous nervous system. When an abnormal intra-subject classification accuracy of 85.6% on stress vs.
condition is experienced, a flight response is invoked. Stress rest, 71.2% on two stress and rest levels, and 58.4% on three
leads to a persistent feeling of energy efficiency and stress and rest levels were obtained using Linear
depression. discriminant classifier with 10-fold cross-validation.

Stress is classified into two categories namely, routine In [10]; the researchers examine the use of EEG for
stress and chronic stress. Routine stress induces inefficiency predicting people's stress levels without manual
and makes it a social cause while chronic stress adversely intervention and recording through BCI (Brain-Computer
affect the anxiety and bipolar diseases. It is also well proved Interface). The efficiency of SVM and KNN Machine
fact that physical health is also affected by mental stress [2]. learning algorithms is evaluated. A training set, target class
At the present, the main cause of most medical conditions is allocated by a stress value derived from the PSS-14 (Full
is mental stress. These comprise strokes, heart attacks, form of PSS) questionnaire response. Using the KNN
depression, nervousness, PTSD (Post-Traumatic Stress algorithm, the maximum average classification accuracy is
Disorders) and immunological disorders [3-4]. Counselling 74.43%. Author confirmed that band power ratios of
is a conventional method to detect stress in human but this different bands are correlated to the stress level of subjects
suffers from a disadvantage that it requires a person to be from measured EEG signals in the front part of the brain.
able to speak frankly. Stress can also affect the activity and
structure of the brain. The study of brain activity is In [11]; Signals coming from the human brain frontal
necessary for the early detection of stress to avoid disease, lobe are used for stress measurement. The brain waves of
and to decrease the risk of brain damage [5]. the thirty subjects are recorded with five different levels of
stress: relaxed, stressed less, stressed moderately, stressed
EEG is one of the most accurate measurement of higher, and stressed alarmingly. Preprocessing of the
electrical activity in the brain. The voltage changes in the recorded EEG signal is followed by the feature extraction.
Manuscript received November 5, 2024
Manuscript revised November 20, 2024
https://doi.org/10.22937/IJCSNS.2024.24.11.12
114 IJCSNS International Journal of Computer Science and Network Security, VOL.24 No.11, November 2024

Feed forward neural network is used to identify the stress in

the brain. FFT was utilized in the proposed method in [12]
for feature extraction. KNN was used for classifying and
detecting stress. The proposed features were most
appropriate for stress detection and mediation applications.
Theta, delta, beta, and alpha waves were used with the best
value for K=3 in KNN with 80% classification accuracy
[12].

The application of IoT technologies in the healthcare

sector with the cloud to assess the physical and mental
activity of the brain using an EEG headset is proposed in
[13-14]. Traditional EEG headsets depend upon multi-
electrode invasive EEG, which stresses and concerns
patients with the use of such headsets. The proposed
architecture model will gather data through car racing
competition using a new, convenient, easy-wear EEG
headset. Such data would be forwarded to Cloud to analyze
and display brainwave levels to control robotic cars to reach
a certain direction of difficulty. This concept makes it
simple to use and flexible for many changes and
developments in the complexity of the traditional mental
assessment using the Neuro-sky Mind-wave EEG headset.
The main objective of the proposed method in [15-16] is to
incorporate the concept of a brain race for the diagnosis and
assessment of mental status as a new transition from an
evaluation of entertainment and brain activity in various
scenarios.

2. Methodology
The purpose of this paper is to use the auxiliary
classifier, a more efficient voting classifier to classify the
EEG signals into healthy and unhealthy EEG signal. The
used auxiliary classifier is based on the KNN and RF
classifier to increase the accuracy of the current model for a
greater number of datasets. A computer-aided diagnostic Fig 1. Flow chart of the proposed method
tool is highly required to help people to diagnose stress and
to suggest proactive action to minimize stress. In the 2.1 Fast Fourier Transform (FFT)
proposed method; FFT is used to generate the periodogram FFT is used to transform the signal from time domain
of input EEG signal. PSD vector value from periodogram to frequency domain [15]. The FFT is called the
has been calculated and used as an input to the proposed computationally efficient implementation of the DFT
voting classifier. The block diagram of the proposed method, described in a set of N samples 𝑥 as follows:
methodology is shown in Fig. 1.
∑ /
𝑋 𝑥 ∗𝑒 ,𝑘 0,1,2, … . . , 𝑁 1
(1)

Where 𝑋 is a complex number known as an absolute

value (frequency magnitude or modulus). 𝑋 is resulting
sequence defined as follows: positive frequencies 0 𝑓
𝐹 /2 correspond to values 0 𝑛 𝑁 1 /2 , while
negative frequencies 𝐹 /2 𝑓 0 correspond to 𝑁
1 /2 𝑛 𝑁 1 . 𝐹 denotes sampling rate. Results
IJCSNS International Journal of Computer Science and Network Security, VOL.24 No.11, November 2024 115

achieved are considered voice signal frequency spectrum

[17]. 2.5 Voting classifier
A voting classifier also called auxiliary classifier is a
2.2 Power Spectral Density (PSD) machine learning based approach. This classifier is based
The power spectral density (PSD) demonstrates the on the maximum likelihood of the output from the chosen
intensity of the energy variations as a frequency function. class, and trains on an ensemble of various examples.
In other words, it demonstrates the strong frequency Rather than creating diverse special models for every one of
variations and the weaknesses of frequencies. The PSD unit them, we construct a solitary model which prepares and
consists of energy by frequency (width) which you can get predicts execution dependent on the aggregate vote larger
through the integration of PSD in a particular frequency part of the exhibition gatherings. Random forest and KNN
range. PSD is computed directly using the FFT or (typically varying types) and basic statistics (such as the
calculating autocorrelation function and then transformed. average) are introduced in our proposed projects. The
In this work, the PSD vector generation is done using the architecture of the proposed voting classifier is shown in Fig.
Welch periodogram method for group 1 and group 2. The 2.
Welch algorithm [18] is a non-parametric way of estimating
PSD which making the frequency spectrum smoother than
the raw FFT flow.

2.3 Random Forest

Leo Breiman from California University first
suggested the random forest in 2001. It consists of a large
number of basic classifiers (decision trees) that are fully
independent of each other. A test sample for this
classification will be determined depends upon voting
outcomes of every single classification and class label for
this sample [19-20].
The key steps to building up the random forest classifier
are as follows:

1. Specifies a no. of elements of every feature subset "M".

2. New feature subset ℎ is created from the entire
feature set at random based on the M value. ℎ is
independent by another subset in a sequence of
ℎ ,ℎ ,……,ℎ .
3. Training the data set for each training set group with a
feature sub-set to construct a decision tree. Every Fig 2. Architecture of a proposed voting classifier
single classifier can be h 𝑥, ℎ (where 𝑥 represents
inputs).
4. Select new ℎ also repeat the above process until
3. Results & Discussion
feature subsets have been traveled. A random forest This proposed methodology has been implemented
classifier is done. using Jupyter notebook in python programming. The
dataset used in the proposed method has EEG signal waves
2.4 k-Nearest Neighbour that include delta, beta, gamma, alpha, and theta waves. The
KNN method is also called a non-parametric method test results and efficiency of the proposed voting classifier
finds application in mathematical problem since the is discussed in this section. In the present study, the
beginning of the 1970s. The basic theory behind KNN is experiment has been conducted in two groups. There are 72
that set of k samples are found nearest to unknown samples files in these groups. These 72 files have been arranged in
in the calibration dataset (for example, depends upon healthy and unhealthy EEG signal with 36 signals of each
distance functions). The average response variables are class. We have applied FFT on both groups to generate PSD
calculated from these k samples by measuring the label of vector for delta, beta, gamma, alpha, and theta waves for
unknown samples. As a consequence, it can be concluded these groups. In the last stage, an auxiliary voting classifier
that the efficiency of KNN is dependent on the value of k. has been used in the experiment. The experiment performs
This means that k is the primary tuning parameter in KNN better computational efficiency when used with the value of
[19]. K=3.
116 IJCSNS International Journal of Computer Science and Network Security, VOL.24 No.11, November 2024

Theta, delta, alpha, gamma and beta are the used in the
experiment for analysis. The delta wave can calculate the
depth of the sleep, and its frequency (1-4Hz). Delta wave
can be identifying the slow-wave sleep using EEG, in slow-
wave sleep brain waves are very slow so this is called
dreamless sleep, and Dreams occurred very often.
Nightmares occur during this sleep but, we are not able to
recall the dreams. The following rates are decreasing during
this sleep BP, respiratory rate, and BMR. Theta wave can
be used to know the functions of the brain, which means
that the difficult task of the brain and it's associated with the
weakness level. The frequency is about (4-7Hz), Theta is
connected with all-around cerebral processing such as Fig 4. Plot of Cumulative sum of data frame of group 2 against
memory conceal and cognitive workload, it is also samples
calculating the tired level of humans. Alpha denotes our
mind-released state, and it records the relaxation of the Fig. 5 represents a scatter plot of cumulative sum of
brain whenever we closed our eyes, we turn into a calm state data frames for both groups after removing the gamma
at that time alpha wave take over, and it is related to shyness waves because it will not affect the accuracy of our model.
and attention, the frequency of alpha is (7-12Hz). Beta Fig. 6 and Fig. 7 shows the histogram plot of the cumulative
waves with frequencies of (12-30Hz). It can notice the body sum of data frames for the 1st and 2nd group respectively.
movements, such as limp movement fore limp (hand) hind
limp (leg), this increases in beta also perceptible as we
notice bodily movements of other peoples. Human brains
imitate the movements of their limbs and indicating the
mirror neuron system. The Gamma waves, typically the
gamma frequency is (>30Hz to 40Hz). Gamma waves can
give information about our sensory inputs. These waves are
similar to the REM (rapid eye movement). All the instances
were used to assess at a point during this process and the
remaining instances were utilized to the training of the
classifier [21].

Fig. 3 and Fig. 4 are the representation of EEG bands

after applying the Fast Fourier Transform (FFT) for the 1st
and 2nd groups. It shows values of EEG bands that are
Fig 5. Scatter plot of cumulative sum of data frame of group 1 against
Theta, Delta, Beta, Alpha and Gamma. group 2

Fig 3. Plot of Cumulative sum of data frame of group 1 against Fig 6. Histogram plot of Cumulative sum of data frame of group 1
samples
IJCSNS International Journal of Computer Science and Network Security, VOL.24 No.11, November 2024 117

4. Conclusion
The work present in this paper proposed a method
to detect and classify stress using EEG signals. It is based
on the features extracted from the EEG waves Delta, Alpha,
Beta, Theta, and Gamma. A voting classifier model is used
in the proposed method to successfully detect stress at
different levels. EEG signal is subjected to FFT to generate
periodogram waves which used to generate the PSD vector.
PSD vector is used as an input to the proposed auxiliary
voting classifier. The voting classifier used in the
experiment is based on KNN and RF. Experimental results
Fig 7. Histogram plot of Cumulative sum of data frame of group 2
have found that the proposed voting classifier model is
better than the previous KNN classifier model in terms of
Comparative analysis of results of the proposed
accuracy and MSE. It has achieved 88.58% accuracy and
method with the existing method. Mean square error (MSE)
11.42% mean squared error. The main contribution of the
and accuracy is used as a performance parameter of the
paper is the use of voting classifier to detect and classify
proposed method. It is proved from the Table 1 that the
stress. In future work, we can use other biomedical signals
proposed method performs better in terms of performance
such as ECG, EMG etc. to detect the stress using the same
parameter with respect to existing methodology. The
methodology. Moreover, intervention and sensitivity
comparative study between the proposed methodology and
analysis methods on EEG signals can also be added to
existing methods is also shown graphically in Fig. 8 and Fig.
determine the stress level.
9 for MSE and accuracy respectively.

Table 1. Comparison of proposed method with the existing

methodologies References
Method MSE Accuracy [1] S. M. U. Saeed, S. M. Anwar, H. Khalid, M. Majid, U. Bagci,
(%) (%) EEG based classification of long-term stress using
SVM and KNN classifier-based method [9] 14.40 85.60 psychological labeling, Sensors 20 (2020) 1886.
Brain mapping based method [10] 25.77 74.33 [2] M. S. Kalas, B. Momin, Stress detection and reduction using
KNN classifier-based approach [12]] 19.95 80.04
eeg signals, in: 2016 International Conference on Electrical,
Proposed Method 11.41 88.58
Electronics, and Optimization Techniques (ICEEOT), IEEE,
2016, pp. 471-475.
[3] O. Attallah, An e_ective mental stress state detection and
evaluation system using minimum number of frontal brain
electrodes, Diagnostics 10 (2020) 292.
[4] A. R. Gaurav, V. Kumar, Eeg-metric based mental stress
detection, Netw Biol 8 (2018) 25-34.
[5] S. W. Lazar, C. E. Kerr, R. H. Wasserman, J. R. Gray, D. N.
Greve, M. T. Treadway, M. McGarvey, B. T. Quinn, J. A.
Dusek, H. Benson, et al., Meditation experience is
associated with increased cortical thickness, Neuroreport 16
(2005) 1893.
[6] M. N. B. Patil, R. Mirajkar, S. Patil, P. Patil, A method for
Fig 8. Comparison between the MSE of the existing methodologies
detection and reduction of stress using eeg, International
with the proposed method
Research Journal of Engineering and Technology (IRJET) 4
(2017) 1528-1604.
[7] P. Youse_ Zowj, Topics on Multiresolution Signal
Processing and Bayesian Modeling with Applications in
Bioinformatics, Ph.D. thesis, Georgia Institute of
Technology, 2021.
[8] T. H. Priya, P. Mahalakshmi, V. Naidu, M. Srinivas, Stress
detection from eeg using power ratio, in: 2020 International
Conference on Emerging Trends in Information Technology
and Engineering (ic-ETITE), IEEE, 2020, pp. 1-6.
[9] Y. Zhang, Q. Wang, Z. Y. Chin, K. K. Ang, Investigating
different stress-relief methods using electroencephalogram
(eeg), in: 2020 42nd Annual International Conference of the
Fig 9. Comparison between the accuract of the existing methodologies
with the proposed method
118 IJCSNS International Journal of Computer Science and Network Security, VOL.24 No.11, November 2024

IEEE Engineering in Medicine & Biology Society (EMBC),

IEEE, 2020, pp. 2999-3002.
Navdeep Shakya is the research scholar in the
[10] X. Gu, Z. Cao, A. Jolfaei, P. Xu, D. Wu, T.-P. Jung, C.-T. Department of Electronics Engineering, Madhav
Lin, EEG based brain-computer interfaces (bcis): A survey Institute of Technology and Science. He has done
of recent studies on signal sensing technologies and Bachelor of Engineering from <> in Electronics and
computational intelligence approaches and their applications, Communication Engineering. He is pursuing Master of
IEEE/ACM transactions on computational biology and Engineering in Communication, Control, and
bioinformatics (2021). Networking from Madhav Institute of Technology and
[11] P. Mukherjee, A. Roy, Detection of stress in human brain, Science, Gwalior. His research interest is mainly in the
field of Signal processing and Biomedical engineering.
2019 Second International Conference on Advanced
Computational and Communication Paradigms (ICACCP) Dr. Rahul Dubey is currently associated with Madhav
(2019) 1-6. Institute of Technology and Science, Gwalior. He is
[12] P. D. Purnamasari, A. Fernandya, Real time eeg-based stress working as an assistant Professor in the Department of
detection and meditation application with k-nearest neighbor, Electronics Engineering. He has more than 10 years of
in: 2019 IEEE R10 Humanitarian Technology Conference experience in the field of academics and research. His
research interest is in the field of Machine fault
(R10-HTC)(47129), 2019, pp. 49-54. doi:10.1109/R10-
diagnosis, Signals and System, Biomedical Signal
HTC47129.2019.9042488. processing.
[13] A. Mansour, H. T. Ouda, On the road to a comparative car
racing eeg-based signals for mental and physical brain Dr. Laxmi Shrivastava is currently associated with
activity evaluation, in: 2019 9th Annual Information Madhav Institute of Technology and Science, Gwalior.
Technology, Electromechanical Engineering and She is working as an Associate Professor in the
Microelectronics Conference (IEMECON), 2019, pp. 43-48. Department of Electronics Engineering. He has more
than 25 years of experience in the field of academics
doi:10.1109/IEMECONX.2019.8877037.
and research. His research interest is in the field of
[14] J. Mladenovi_c, Standardization of protocol design for user Wireless Sensor Networks, Biomedical Signal
training in eeg-based brain{computer interface, Journal of processing.
Neural Engineering 18 (2021) 011003.
[15] J. L. M. Kumar, M. Rashid, R. M. Musa, M. A. M. Razman,
N. Sulaiman, R. Jailani, A. P. A. Majeed, An evaluation of
different fast fourier transform transfer learning pipelines for
the classification of wink-based eeg signals,
MEKATRONIKA 2 (2020) 1-7.
[16] R. S. Alkhawaldeh, Arabic (indian) digit and written
recognition using recurrent transfer deep architecture, Soft
Computing 25 (2021) 3131-3141.
[17] M. H. Same, G. Gandubert, G. Gleeton, P. Ivanov, R. Landry,
Simplified welch algorithm for spectrum monitoring,
Applied Sciences 11 (2021)
[18] A. Parmar, R. Katariya, V. Patel, A review on random forest:
An ensemble classifier, in: J. Hemanth, X. Fernando, P.
Lafata, Z. Baig (Eds.), International Conference on
Intelligent Data Communication Technologies and Internet
of Things (ICICI) 2018, Springer International Publishing,
Cham, 2019, pp. 758-763.
[19] N. Zaabar, S. Niculescu, M. Mihoubi, Assessment of
combining convolutional neural networks and object based
image analysis to land cover classification using sentinel 2
satellite imagery (tenes region, algeria), International
Archives of the Photogrammetry, Remote Sensing and
Spatial Information Sciences 43 (2021) 383-389.
[20] D. Shon, K. Im, J.-H. Park, D.-S. Lim, B. Jang, J.-M. Kim,
Emotional stress state detection using genetic algorithm-
based feature selection on eeg signals, International Journal
of environmental research and public health 15 (2018) 2461.