1

Classification of EEG Signals Utilizing DWT for

Feature Extraction and Evolutionary Algorithms for
Feature Selection
Mayur Akewar
Alumni, Department of Computer Science & Engineering
Shri Ramdeobaba College of Engineering & Management
Email: mayurakewar87@[Link]

Abstract—The electroencephalogram (EEG) signals, which nologies bring sophistication to the analysis, enabling a more
capture the electrical activities of the brain, serve as a valuable nuanced interpretation of intricate EEG patterns.
tool for understanding neuronal function. The analysis of EEG The methodology adopted in this paper intricately involves
signals is pivotal for detecting and diagnosing various brain
disorders, with a specific focus on conditions such as epilepsy discrete wavelet transform (DWT) for extracting features from
and seizures. This paper introduces an EEG signal classifi- EEG signal datasets. DWT provides a robust framework for
cation approach, leveraging machine learning algorithms. The unveiling essential patterns within the signals, contributing
methodology involves the extraction of features from EEG signal to a more comprehensive analysis. Evolutionary algorithms,
datasets through discrete wavelet transform (DWT). Optimal notably Genetic Algorithm (GA) and Particle Swarm Opti-
feature selection is then accomplished using evolutionary algo-
rithms, specifically Genetic Algorithm (GA) and Particle Swarm mization (PSO), take center stage for optimal feature selection.
Optimization (PSO). To identify the most effective classification This strategic approach ensures that the features selected for
method, various machine learning algorithms, including Support classification enhance the model’s discriminative power. The
Vector Machines (SVM), Naive Bayes, Decision Trees, and Ran- arsenal of machine learning algorithms deployed in this study
dom Forest, are systematically compared. This comprehensive encompasses Support Vector Machines (SVM), Naive Bayes,
evaluation aims to enhance the accuracy and efficiency of EEG
signal classification for improved diagnosis and understanding Decision Trees, and Random Forest. These algorithms are
of neurological conditions. The code is published at the URL systematically compared and evaluated to identify the most
[[Link] effective classification method. This rigorous analysis seeks
Index Terms—EEG, Discrete Wavelet Transform, Genetic Al- not only to enhance the accuracy and efficiency of EEG signal
gorithm, Particle Swarm Optimization, Machine Learning. classification but also to foster a deeper understanding of
neurological conditions. The journey outlined in this paper
represents a significant stride towards advancing the field, with
I. I NTRODUCTION implications for more precise diagnoses and improved insights
into the intricate landscape of brain disorders.
T He human brain’s intricate electrical activities are en-
capsulated through electroencephalogram (EEG) signals,
providing an invaluable gateway into understanding neuronal
This paper represents a significant stride in exploring EEG
signals, merging sophisticated analytical techniques, machine
learning, and evolutionary algorithms. The ultimate goal is to
function. This paper delves into the realm of EEG signal anal-
not only enhance the accuracy of EEG signal classification but
ysis, an essential endeavor for the detection and diagnosis of
also to deepen our understanding of neurological conditions,
various brain disorders, with a specific emphasis on conditions
paving the way for more effective diagnoses and targeted
such as epilepsy and seizures. The complexity and richness of
interventions.
EEG signals necessitate advanced analytical tools, prompting
the introduction of a novel EEG signal classification approach
that harnesses the power of machine learning algorithms. EEG II. L ITERATURE R EVIEW
signals, as snapshots of brain activity, present a wealth of In the vast landscape of neuroscience, the classification
information crucial for unraveling the mysteries of neuro- of EEG signals stands as a crucial frontier, with machine
logical conditions. The benefits of EEG signal classification learning methods emerging as powerful allies in unraveling
are profound, serving as a diagnostic beacon that illuminates the intricate patterns within these brainwave recordings. This
the path toward accurate identification and understanding of literature review delves into the diverse methodologies and
disorders affecting the brain. The integration of machine advancements employed in EEG signal classification. In [1],
learning (ML) and artificial intelligence (AI) in EEG signal the authors address the critical problem of EEG signal analysis
classification emerges as a pivotal advancement. These tech- for detecting epileptic seizures in health informatics. Tradi-
tionally, EEG signal classification involves two key steps:
November 20, 2023. The author is associated with the Department of Com-
puter Science & Engineering, Shri Ramdeobaba College of Engineering & signal preprocessing and classification. While preprocessing
Management, Nagpur, Maharashtra 440013, India. Phone: +91-712-2580011. aims to reduce noise and signal dimensionality, it inevitably
 2

leads to a loss of information before classification. In this and loss models conducted using Python in Keras. In [5],
study, the authors propose an innovative approach consid- the authors introduce a privacy-preserving deep learning (DL)
ering the loss of information in the classification step. The architecture called Federated Transfer Learning (FTL) for elec-
novelty lies in incorporating a fuzzy classifier to account troencephalographic (EEG) classification in Brain-Computer
for the uncertainty caused by the loss of information during Interfaces (BCI). Due to privacy concerns, constructing large
dimensionality reduction. To integrate the fuzzy classifier, EEG-BCI datasets is challenging. FTL operates within the fed-
the preprocessing step is modified to include a fuzzification erated learning framework, utilizing the single-trial covariance
procedure. The Fuzzy Decision Tree (FDT) is employed as matrix to extract discriminative information from multi-subject
the fuzzy classifier for epileptic seizure detection, achieving an EEG data through domain adaptation techniques. Evaluating
impressive 99.5% classification accuracy. Comparative analy- the architecture on the PhysioNet dataset for 2-class motor
sis with other studies underscores the effectiveness of FDT imagery classification, FTL achieves a 2% higher classifica-
in detecting epileptic seizures, demonstrating its potential as a tion accuracy in subject-adaptive analysis without actual data
valuable tool in EEG signal classification. The current focus in sharing. In the absence of multi-subject data, FTL outperforms
brain-computer interface (BCI) research is on motor imagery other state-of-the-art DL architectures by 6% in accuracy. In
(MI) electroencephalogram (EEG) signal classification, with [6], the authors tackle the challenging task of epileptic seizure
deep learning emerging as a promising method for automatic detection by classifying electroencephalography (EEG) signals
feature extraction and classification. However, challenges arise into ictal and interictal classes. Given the complex and non-
in practical applications due to the need for a large amount linear nature of EEG signals, extracting discriminative features
of labeled data and the computational expense of training is crucial for accurate classification. This paper introduces an
deep learning models from scratch. To address these issues, approach that combines fuzzy-based and traditional machine
the paper [2] introduces a deep transfer convolutional neural learning algorithms to address this issue. The proposed frame-
network (CNN) framework based on VGG-16 for EEG signal work successfully classifies unknown EEG signal segments
classification. This framework utilizes a pre-trained VGG-16 into ictal and interictal classes, with empirical evaluation on
CNN model on ImageNet as a base and transfers its parameters benchmark datasets (Bonn and CHB-MIT) showing that K-
to a target CNN model for MI EEG signal classification. The Nearest Neighbor (KNN) and Fuzzy Rough Nearest Neighbor
target model’s front-layer parameters are frozen, while later- (FRNN) achieve the highest classification accuracy, along
layer parameters are fine-tuned using the target MI dataset. with improved sensitivity and specificity percentages. In [7],
The proposed framework, validated on the BCI competition the authors propose a multilevel machine learning approach
IV benchmark dataset 2b, demonstrates improved accuracy and for diagnosing epilepsy using electroencephalography (EEG)
efficiency in EEG signal classification compared to traditional signals. The method comprises pre-processing, feature extrac-
methods like support vector machine (SVM), artificial neural tion, feature concatenation, feature selection, and classification
network (ANN), and standard CNN. In [3], the authors present phases. Tunable-Q wavelet transform (TQWT) is employed
a novel approach, termed epileptic EEG signal classification for pre-processing, generating 25 frequency coefficient sub-
(EESC), for accurate categorization of different epileptic states bands. Feature extraction utilizes quadruple symmetric pattern
using electroencephalogram (EEG) signals. The methodology (QSP), extracting 256 features from the raw EEG signal and
involves transforming epileptic EEG signals into power spec- the 25 sub-bands. Neighborhood component analysis (NCA)
trum density energy diagrams (PSDEDs), followed by the is applied for feature selection, choosing the most significant
application of deep convolutional neural networks (DCNNs) features. The classification phase employs the k nearest neigh-
and transfer learning to automatically extract features from bors (kNN) classifier. Tested on the Bonn EEG dataset, the
these PSDEDs. The proposed approach exhibits superior per- proposed method achieves a 98.4% success rate for a case
formance in comparison to existing methods, demonstrating with five classes, demonstrating its potential for application in
high accuracy and efficiency. In a case study utilizing CHB- larger datasets for further validation.
MIT epileptic EEG data, the EESC method achieves an In [8], the authors present a study utilizing Wavelet Trans-
average classification accuracy exceeding 90%, highlighting form (WT) for processing Electroencephalogram (EEG) sig-
its effectiveness in epilepsy analysis. In [4], the authors nals. They apply linear discriminant analysis (LDA) for feature
explore the application of neural networks (NN) in EEG signal selection and dimensionality reduction, utilizing the resulting
classification, comparing the performance of SVM, logistic re- two-dimensional features as a benchmark for classification
gression, and NN machine learning algorithms. They introduce via a Multi-Layer Perceptron (MLP) neural network. The
two-layer LSTM and four-layer improved NN deep learning proposed model achieves a high sensitivity, specificity, and
algorithms, featuring novel one-dimensional gradient descent accuracy of 100% for five classification problems. Compar-
activation functions with radial basis operations in the initial ative analysis with previous literature methods demonstrates
layers for enhanced performance. Statistical features such as the superiority of their approach in EEG signal classification
mean, standard deviation, kurtosis, and skewness are extracted and automated diagnosis. In [9], the authors introduce a novel
from EEG data sourced from the Bonn database and applied to EEG classification network, Separable EEGNet (S-EEGNet),
different classification techniques. Evaluation metrics include leveraging Hilbert–Huang Transform (HHT) and a separable
accuracy, precision, recall, and F1 score. The improved NN Convolutional Neural Network (CNN) with bilinear interpola-
and LSTM architectures outperform others, with extensive tion. HHT transforms EEG signals into a time-frequency repre-
simulations involving various activation functions, optimizers, sentation, enhancing their description in the frequency domain.
 3

The network combines depthwise and pointwise elements to ing techniques, highlighting the exceptional capabilities of
extract feature maps, incorporating a bilinear interpolation Electroencephalogram (EEG) signals, including high tempo-
method for flexible grid deformation based on local, dense, ral resolution, affordability, portability, and non-invasiveness.
and adaptive EEG data characteristics. S-EEGNet conducts They explore different frequency bands associated with brain
end-to-end learning of time and space dimensions, improving signals and provide a literature survey covering diverse appli-
EEG classification accuracy. Experimental results on two EEG cations of EEG signals in computer-aided technologies. These
datasets demonstrate accuracy improvements of 3.6%, 1.15%, applications range from diagnosing neurological disorders
and 1.33% for motor imagery classification, and 89.91%, like epilepsy and major depressive disorder to monitoring
and 88.31% for emotion classification. In [10], the authors motor imagery, identity authentication, emotion recognition,
present a machine learning-based classifier designed to differ- sleep stage classification, eye state detection, and drowsiness
entiate between Schizophrenia patients and healthy controls monitoring. The survey includes a comparative analysis of
using features extracted from electroencephalograph (EEG) EEG datasets, data acquisition methods, preprocessing and
signals based on event-related potential (ERP). Utilizing an feature extraction techniques, classification models, and sta-
online EEG dataset with 81 subjects (32 healthy controls tistical tests. The paper also identifies research gaps and
and 49 Schizophrenia patients), various features are extracted, outlines future directions, encouraging further exploration in
preprocessed, and subjected to the random forest machine the field. The survey concludes with a brief overview of
learning algorithm due to the dataset’s limited size. The study studies investigating the fusion of brain signals from multiple
concludes that the classification accuracy can be enhanced by modalities. In [15], the authors introduce a novel approach,
incorporating specific features extracted from EEG signals. In Pin-UTSVM, for the classification of Electroencephalogram
[11], the authors introduce an EEG classification approach for (EEG) signals. They address the computational complexity of
imagined speech with high accuracy and efficiency. Utilizing Support Vector Machines (SVMs) and twin support vector ma-
correntropy spectral density (CSD) matrices from EEG signals chines (TWSVM) by incorporating prior information through
across different channels, the distances between these matri- universum data in universum SVM (USVM) and universum
ces serve as measures for imagined speech recognition. The twin SVM (UTSVM) models. The proposed Pin-UTSVM,
simplicity and accuracy of Riemannian distance, a statistical utilizing the pinball loss function, enhances stability and noise
method, are leveraged in the evaluation of channel selection resilience, maintaining computational efficiency. Universum
and frequency band detection during imagined speech. The data, specifically interictal EEG signals, is employed in the
study, conducted on the ”Kara One” database featuring EEG approach. Experimental results across various noise levels
signals during imagined speech, demonstrates an average demonstrate the robustness of Pin-UTSVM, outperforming
classification accuracy of 90.25% for all subjects, showcasing existing models in multiple feature extraction techniques. Sta-
the efficiency and accuracy of the proposed method compared tistical tests affirm the significant superiority of Pin-UTSVM
to other imagined speech classification approaches. In [12], the over baseline models in EEG signal classification.
authors present a novel approach for the automatic detection The literature review underscores the multifaceted landscape
of seizure activity in epileptic EEG signals. They employ a of EEG signal classification, revealing the dynamic interplay
Long Short-Term Memory (LSTM) network for classification, between advancements in machine learning and the intricate
utilizing Discrete Wavelet Transform (DWT) for noise removal realm of neuroscience. Our approach uses DWT and statisti-
and feature extraction, resulting in 20 eigenvalue features. cal features and selects the best features using evolutionary
Feature selection is optimized through correlation and P-value algorithms for EEG signal classification.
analysis, aiming to minimize LSTM trainable parameters while
maintaining high accuracy. Comparative analysis demonstrates
III. M ETHODOLOGY
the superiority of the proposed method over other frameworks,
including logistic regression (LR), support vector machine In our proposed methodology as shown in Figure 1, we
(SVM), K-nearest neighbor (K-NN), and decision tree (DT). In initiate the process by taking EEG features as input, rec-
[13], the authors introduce a novel feature selection technique ognizing the significance of these features in capturing the
based on Rényi min-entropy to enhance the efficiency of intricate electrical activities of the brain. To enhance efficiency
brain–computer interface (BCI). In the context of multiple- and eliminate redundancy, we employ Principal Component
class problems, where classification accuracy relies on ef- Analysis (PCA) to reduce the dimensionality of the feature
fective feature selection from wavelet packet transformation space. This reduction not only streamlines the subsequent
(WPT) features extracted from electroencephalogram (EEG) processes but also retains the most critical information for clas-
signals, conventional methods use Shannon entropy and mu- sification. Subsequently, we employ Discrete Wavelet Trans-
tual information for feature selection. The proposed Rényi form (DWT) and statistical feature extraction to delve deeper
min-entropy-based approach is demonstrated to outperform into the essence of the EEG signals, capturing both temporal
these traditional methods in the classification of multiple and statistical characteristics. Recognizing the importance of
EEG signals. The BCI competition-IV dataset, containing 4- selecting the most pertinent features, we incorporate evolu-
class motor imagery EEG signals, is utilized for experimenta- tionary algorithms, specifically Genetic Algorithm (GA) and
tion, showcasing the superiority of the proposed approach in Particle Swarm Optimization (PSO), to iteratively identify
multiple-class BCI scenarios. In [14], the authors conduct a the optimal subset of features. Finally, employing a machine
comprehensive comparison of various functional neuroimag- learning algorithm, such as Support Vector Machines (SVM)
 4

Fig. 1. EEG Signal classification workflow.

Fig. 2. EEG Signals Captured Data.

or Decision Trees, we classify the EEG signals based on the essential information. By identifying and eliminating redun-
refined and selected features, resulting in a comprehensive and dant features, PCA helps streamline the dataset, enhancing
effective approach for EEG signal classification. computational efficiency and simplifying subsequent analysis.
This technique aids in focusing on the most significant patterns
A. Data and variations in the EEG signals, contributing to a more
An EEG (Electroencephalogram) signal dataset typically refined and informative representation of the neural activity
consists of a series of numerical values corresponding to differ- recorded by the EEG.
ent channels or electrodes capturing electrical brain activity. Principal Component Analysis (PCA) [16] is a statistical
In the provided snippet, X1 through X178 likely represent technique widely used in data analysis and dimensionality
data points or readings from ten different channels or features reduction. It serves as a valuable tool for transforming high-
within the EEG dataset. Each variable (X1, X2, ..., X178) dimensional datasets into a lower-dimensional space while
corresponds to a specific measurement or attribute of the EEG retaining essential information. The importance of PCA lies
signal, providing information about the electrical activity at in its ability to simplify complex datasets, eliminate redun-
different points on the scalp or brain regions. These values dancies, and identify the most significant patterns or features.
could represent voltage measurements, frequency components, Mathematically, PCA [16] involves finding the eigenvectors
or other relevant features extracted from the EEG signals and eigenvalues of the covariance matrix of the original
during a specific time period, allowing for the analysis and dataset. The covariance matrix captures the relationships be-
classification of brain activity patterns. tween different variables. Let X be the data matrix with each
row representing an observation and each column representing
a variable. The covariance matrix
B. Feature Reduction
1
The utilization of Principal Component Analysis (PCA) C= (X − X − )T (X − X − )
[16] in processing EEG signals serves as a strategic approach n−1
to mitigate redundant features within the dataset. PCA is , where n is the number of observations and X − is the
employed to transform the original high-dimensional EEG mean vector. The next step is to find the eigenvectors
signal dataset into a reduced-dimensional space while retaining (v1, v2, ....., vp) and eigen values λ1, λ2, ..., λp of C These
 5

eigenvectors represent the directions of maximum variance in (e) Coefficient of Variation: Measures the relative variability
the original data, and the corresponding eigenvalues quantify or dispersion of the EEG signal. A higher coefficient
the amount of variance along those directions. of variation indicates greater relative variability in the
The principal components are then obtained by selecting the signal.
top k eigenvectors corresponding to the k largest eigenvalues. (f) Skewness: Indicates the asymmetry of the distribution of
The transformed dataset Y is given by Y = [Link] , where Vk the EEG signal. Positive skewness suggests a longer tail
is the matrix containing the selected eigenvectors. In summary, on the right, while negative skewness suggests a longer
PCA [15] simplifies datasets by transforming them into a tail on the left.
space defined by their principal components, preserving the (g) 1st Difference Mean: Represents the mean of the first
most critical information while reducing dimensionality. This difference of the EEG signal. It offers insights into the
process facilitates better visualization, noise reduction, and average rate of change in the signal.
improved efficiency in subsequent data analysis tasks. (h) 1st Difference Max: Denotes the maximum value of
the first difference of the EEG signal. Highlights the
C. Feature Extraction maximum rate of change in the signal.
(i) Variance: Measures the spread or dispersion of the EEG
In the process of feature extraction, our approach considers
signal. Higher variance indicates greater variability in
a combination of statistical features and features obtained
the signal values.
through Discrete Wavelet Transform (DWT), all derived from
(j) Mean of Vertex to Vertex Slope: Represents the average
the reduced dimensionality features.
slope between successive maxima and minima in the
1) Statistical Features:
EEG signal. It characterizes the average steepness of
(a) Hjorth Features: Hjorth parameters are statistical mea- the signal between consecutive extrema.
sures used to analyze the time-domain characteristics of In summary, these statistical features provide diverse
a signal. They are commonly employed in the analysis of insights into different aspects of the EEG signal’s dis-
electroencephalogram (EEG) signals. The three Hjorth tribution, variability, and rate of change, contributing to
parameters are: a comprehensive analysis of the signal’s characteristics.
• Activity: Activity represents the total energy or the
2) Discrete Wavelet Transform: The Discrete Wavelet
variance of the signal. It is computed as the variance
Transform (DWT) [17] is a mathematical technique used for
of the signal. Higher activity values indicate a higher
analyzing signals and images. It decomposes a signal into
level of overall signal energy.
different frequency components, revealing both low and high-
• Mobility: Mobility measures the mean frequency
frequency information. DWT is widely employed in various
or the average frequency at which the energy is
applications, including signal processing, image compression,
distributed in the signal. It is computed as the ratio
and denoising. Mathematically, the one-dimensional DWT for
of the standard deviation of the first derivative of the
a discrete signal x[n] is expressed as follows:
signal to the standard deviation of the signal. Higher
mobility values suggest a higher mean frequency X
and greater frequency dispersion in the signal. W (a, b) = nX[n].ψa ,b [n]
• Complexity: Complexity is an indicator of the sig-
nal’s irregularity or waveform complexity. It is Here, W (a, b) represents the wavelet coefficients after trans-
computed as the ratio of the mobility of the first formation. a and b are scaling and translation parameters.
derivative to the mobility of the signal. Higher ψa ,b [n] is the wavelet function.
complexity values indicate a more irregular and The two parameters, a and b, control the scaling and
complex signal waveform. translation of the wavelet function. The DWT is applied
These Hjorth parameters offer insights into different by successively dividing the signal into approximation (low-
aspects of the signal, with Activity representing overall frequency) and detail (high-frequency) components through
energy, Mobility capturing frequency characteristics, and a process known as ”downsampling” or ”decimation.” For
Complexity describing the irregularity or complexity of a two-dimensional signal (e.g., an image), the DWT can be
the signal waveform. extended similarly. The process involves applying the one-
(b) Kurtosis: Kurtosis is a statistical measure that assesses dimensional DWT separately to rows and columns. The key
the tailedness and presence of outliers in the distribution steps of the DWT algorithm involve a series of convolutions
of the EEG signal. Higher kurtosis values indicate and downsamplings using the wavelet function. The result is a
heavier tails and a higher likelihood of outliers in the set of coefficients that represent different levels of detail and
signal distribution. approximation of the input signal or image. DWT is known
(c) 2nd Diff Mean: This feature represents the mean of the for its ability to capture both localized and global features of
second derivative of the EEG signal. It provides insights signals and images efficiently, making it a powerful tool in
into the average rate of change or curvature of the signal. various signal processing applications.
(d) 2nd Diff Max: Denotes the maximum value of the 3) Wavelet transform features: These features are com-
second derivative of the EEG signal. It highlights the puted based on the Discrete Wavelet Transform (DWT) [17]
maximum rate of change or curvature in the signal. and provide statistical and informational measures for both
 6

the approximate and detailed coefficients obtained through the features are integrated to form a set of final features. This
wavelet transform. integration enables a comprehensive representation of the EEG
signal, combining information about its statistical properties
(a) Wavelet Approximate Mean: It calculates the mean (av-
and frequency characteristics.
erage) of the approximate coefficients obtained through
The ultimate goal of this feature extraction process is
wavelet transform. Represents the central tendency of
to create a more enhanced and representative feature set
the approximate coefficients, providing insight into the
that encapsulates crucial aspects of the EEG signals. These
overall energy distribution at the approximate level.
extracted features are often utilized as inputs for machine
(b) Wavelet Approximate Std Deviation: Computes the stan-
learning algorithms, contributing to tasks like EEG signal
dard deviation of the approximate coefficients. Measures
classification for various neurological applications. The combi-
the dispersion or spread of the approximate coefficients,
nation of statistical features and wavelet transform coefficients
indicating how much they deviate from the mean. Higher
ensures a thorough analysis of EEG signals, capturing both
values suggest greater variability.
their statistical properties and frequency characteristics. This
(c) Wavelet Detailed Mean: Calculates the mean of the
process lays the groundwork for subsequent tasks such as
detailed coefficients obtained through wavelet transform.
classification and interpretation in the realm of EEG signal
Reflects the central tendency of the detailed coefficients,
analysis.
giving information about the overall energy distribution
at the detailed level.
(d) Wavelet Detailed Std Deviation: Computes the standard D. Feature Selection
deviation of the detailed coefficients. Indicates the vari- The process involves the integration of features derived
ability or spread of the detailed coefficients around their from both Discrete Wavelet Transform (DWT) and statistical
mean. Higher values imply greater fluctuations. analyses. These combined features serve as the input to sophis-
(e) Wavelet Approximate Energy: Measures the energy of ticated feature selection algorithms, namely Particle Swarm
the approximate coefficients. Provides an indication of Optimization (PSO) and Genetic Algorithm (GA). This inte-
the overall strength or magnitude of the approximate co- gration represents a comprehensive approach, leveraging the
efficients, revealing the contribution of each coefficient strengths of both DWT, which captures frequency information
to the signal. in the signal, and statistical features, which provide insights
(f) Wavelet Detailed Energy: Calculates the energy of the into its distribution and characteristics. By subjecting this
detailed coefficients. Represents the strength or magni- amalgamated feature set to the optimization power of PSO and
tude of the detailed coefficients, indicating their contri- GA, the selection algorithms work synergistically to identify
bution to capturing fine details in the signal. the most relevant and discriminative features, optimizing the
(g) Wavelet Approximate Entropy: Computes the entropy input for subsequent stages, such as machine learning-based
of the approximate coefficients. Quantifies the unpre- EEG signal classification.
dictability or disorder in the approximate coefficients, 1) Genetic Algorithm: Genetic Algorithms (GAs) [18] are
offering insights into the complexity of the signal. optimization algorithms inspired by the process of natural
(h) Wavelet Detailed Entropy: Measures the entropy of the selection and genetics. They are used to find approximate
detailed coefficients. Quantifies the unpredictability or solutions to optimization and search problems. The algorithm
disorder in the detailed coefficients, providing informa- works by evolving a population of potential solutions over
tion about the complexity of fine details in the signal. multiple generations to improve their fitness for solving a
These features offer a comprehensive analysis of the statis- particular problem.
tical properties, energy distribution, and informational content (a) Basic Concepts:
at both the approximate and detailed levels of the wavelet- • Population: A set of potential solutions to the prob-
transformed signal. lem, represented as individuals or chromosomes.
The process involves the extraction of features from EEG • Chromosome: A representation of a potential solu-
signals, employing both statistical and wavelet transform tion, often as a binary string or a set of parameters.
coefficient methods. Statistical features encompass measures • Fitness Function: A function that evaluates the per-
like mean, standard deviation, skewness, kurtosis, and various formance or quality of a solution. It assigns a fitness
others, offering insights into the distribution and character- score to each individual in the population.
istics of the EEG signal. These features capture essential • Selection: The process of choosing individuals from
statistical properties, aiding in understanding the central ten- the population to form a new generation based on
dency, variability, and shape of the EEG signal’s distribution. their fitness scores. Higher fitness individuals have
Utilizing wavelet transform involves decomposing the EEG a higher chance of being selected.
signal into different frequency components using wavelets. • Crossover (Recombination): The process of com-
Features are then derived from the coefficients obtained during bining genetic material of two parents to produce
this transformation. Wavelet transform captures both time and offspring. It mimics the crossover of genetic mate-
frequency information, making it suitable for analyzing signals rial in biological reproduction.
with non-stationary characteristics, such as EEG signals. The • Mutation: A random change applied to some indi-
calculated statistical features and wavelet transform coefficient viduals in the population to introduce diversity.
 7

(b) Algorithm: Initialization: Initialize particles with random posi-

•
• Initialization: Generate an initial population of in- tions and velocities.
dividuals. • Evaluation: Evaluate the fitness of each particle
• Evaluation: Evaluate the fitness of each individual using the objective function.
using the fitness function. • Update Personal Best (Pbest): Update the personal
• Selection: Select individuals based on their fitness best position for each particle based on its current
scores. Higher fitness individuals have a higher fitness.
probability of being chosen. • Update Global Best (Gbest): Update the global best
• Crossover (Recombination): Combine genetic mate- position based on the fitness of particles in the entire
rial of selected parents to create offspring. swarm.
• Mutation: Introduce random changes in some indi- • Update Velocity and Position: Update the velocity
viduals to maintain diversity. and position of each particle based on its previous
• Replacement: Create a new population from parents velocity, personal best, and global best positions.
and offspring. • Termination: Repeat steps 2-5 for multiple iterations
• Termination: Repeat the process for multiple gener- or until a stopping criterion is met.
ations or until a stopping criterion is met. (c) Mathematical Background:
(c) Mathematical Background: • Representation (Particle): A particle i is represented
• Representation (Chromosome): by a position vector
X = [x1 , x2 , ..., xn ] Xi = [xi1 , xi2 , . . . .., xin ]
, where xi represents a gene or parameter.
• Fitness Function: F (X), where F evaluates the in an n-dimensional search space.
quality of the solution represented by X. • Velocity Update: The velocity of particle i is up-
• Selection: Probability of selection for an individual dated using the formula:
F (i) Vi (t+1) = [Link] (t)+c1 .r1 .(P besti −Xi )+c2 .r2 .(Gbest−Xi )
i : P (i) = P
Fj F (j)
• Crossover (Recombination): Combine genetic mate- where w is the inertia weight, c1and c2 are acceler-
rial from two parents, e.g., single-point crossover: ation coefficients, and r1and r2 are random numbers
• Position Update: The position of particle i is up-
Xo f f spring = [x1 , ...., xk , ...., xn ] dated using the formula:
, where k is a crossover point. Xi (t + 1) = Xi (t) + Vi (t + 1)
• Mutation: Randomly change genes, e.g.,
xi = xi + mutationr ate ∗ randomnumber Inertia Weight (w): It controls the impact of the
previous velocity on the new velocity. A higher
Genetic algorithms are versatile and applicable to w emphasizes global exploration, while a lower w
various optimization problems, and their effective- emphasizes local exploitation. Acceleration Coeffi-
ness often lies in tuning parameters such as mutation cients (c1 and c2 ): They determine the impact of
rate, crossover rate, and population size. personal and global best positions on the particle’s
2) Particle Swarm Optimization: Particle Swarm Opti- movement.
mization (PSO) [19] is a population-based optimization algo-
rithm inspired by the social behavior of birds and fish. It is
used for solving optimization problems where the objective is E. Classification
to find the optimal solution in a search space. PSO simulates
The chosen features, derived through the combined pro-
the social interaction and cooperation among individuals in a
cess of feature extraction and selection, are then fed into
swarm to search for the best solution.
classification algorithms. In this stage, the primary objective
(a) Basic Concepts: is to employ machine learning techniques to categorize the
• Particle: An individual solution in the search space, EEG signals based on the discriminating features identified
represented by a position vector. earlier. The classification algorithms, such as Support Vector
• Swarm: A set of particles that move through the Machines (SVM), Naive Bayes, Decision Trees, and Random
search space to find the optimal solution. Forest, utilize the selected features to distinguish and assign
• Position: A vector representing a solution in the the EEG signals to different predefined classes. This final step
search space. completes the proposed methodology, where the integration
• Velocity: The rate of change of a particle’s position. of advanced feature extraction, selection, and classification
It guides the particle’s movement toward a better techniques aims to enhance the accuracy and efficiency of
solution. EEG signal classification for a more nuanced understanding
(b) Algorithm Steps: of neurological conditions.
 8

1) Support Vector Machine (SVM):: SVM is a super-

vised machine learning algorithm used for classification and
regression tasks. It finds a hyperplane in an N-dimensional
space that separates classes with a maximum margin. It is
effective in high-dimensional spaces and capable of handling
both linear and non-linear data. It is Robust against overfitting,
effective in high-dimensional spaces.
2) Naive Bayes:: Naive Bayes is a probabilistic clas-
sification algorithm based on Bayes’ theorem. It assumes
independence among features, hence the term ”naive.” It is
simple and computationally efficient. Performs well on text
classification and spam filtering. The algorithm is fast and
efficient, works well with categorical data.
3) k-Nearest Neighbors (KNN):: KNN is a simple,
instance-based learning algorithm. It classifies a new data
point based on the majority class of its k nearest neighbors.
No model training phase; it memorizes the entire dataset.
Effective for small to medium-sized datasets. The algorithm Fig. 3. Confusion Matrix for KNN and PSO.
is simple and intuitive, works well with locally clustered data.
Cons: Computationally expensive for large datasets, sensitive
to irrelevant features.
4) Random Forest:: Random Forest is an ensemble learn-
ing method that constructs a multitude of decision trees during
training. The final prediction is the mode of the predictions of
individual trees (classification) or the average (regression). It
combines multiple weak learners to create a strong learner.
Reduces overfitting and improves accuracy. It handles high-
dimensional data, robust to outliers.
We conducted a comprehensive comparison of all the algo-
rithms by evaluating their performance using selected features
obtained through evolutionary algorithms, namely Genetic
Algorithm (GA) and Particle Swarm Optimization (PSO). Ad-
ditionally, we compared their performance with the utilization
of all available features. This thorough evaluation allows us
to assess the effectiveness of the feature selection process
performed by GA and PSO in enhancing the classification
Fig. 4. Confusion Matrix for Naive Bayes and PSO.
accuracy of the algorithms compared to the scenario where
all features are considered. The goal is to identify the most
optimal combination of algorithms and features that yield
the highest precision and recall in the classification of EEG
signals.

IV. R ESULTS
The initial dataset, sourced from the reference, comprises 5
distinct folders, each containing 100 files representing individ-
ual subjects. Each file records brain activity over a duration of
23.6 seconds, resulting in a time-series of 4097 data points.
These data points represent EEG values This results in 11500
pieces of information, with 178 data points per second. The
data is further categorized into non-seizure and seizure activ-
ities. It undergoes division into training and testing datasets
using k-fold cross-validation. The standardized database is
then fed into a feature extraction algorithm based on the
aforementioned approach. The accuracy of the classification Fig. 5. Confusion Matrix for SVM and PSO.
model is calculated and compared.
PSO and Classification algorithms confusion matrix are
shown in Figure 3 to Figure 6.
 9

Fig. 6. Confusion Matrix for Random Forest and PSO.

Fig. 8. Accuracy for Naive Bayes and GA.

Fig. 7. Accuracy for KNN and GA.

GA and Classification algorithms test set accuracy are

shown in Figure 7 to Figure 10.
Fig. 9. Accuracy for SVM and GA.
V. C ONCLUSION AND F UTURE W ORK
In conclusion, our investigation into the classification of
EEG signals through the synergistic utilization of Discrete our approach aims to enhance diagnostic capabilities. The
Wavelet Transform (DWT) for feature extraction and evolu- successful implementation of genetic algorithms (GA) and
tionary algorithms for feature selection has yielded promis- Particle Swarm Optimization (PSO) in feature selection under-
ing results. The comprehensive evaluation of various ma- scores their potential for optimizing the discriminative power
chine learning algorithms, including Support Vector Machines of selected features, contributing to improved classification
(SVM), Naive Bayes, Decision Trees, and Random Forest, accuracy.
has provided insights into their efficacy for accurate EEG Building upon the foundations laid in this study, future work
signal classification. By addressing the intricate nature of in the realm of EEG signal classification holds substantial
neurological conditions, particularly epilepsy and seizures, potential for advancements. First and foremost, the exploration
 10

[3] Yunyuan Gao, Bo Gao, Qiang Chen, Jia Liu, Yingchun Zhang Deep Con-
volutional Neural Network-Based Epileptic Electroencephalogram (EEG)
Signal Classification . METHODS article 2020, [Online]. Available:
[Link]
[4] P. Nagabushanam, S. Thomas George, S. Radha EEG sig-
nal classification using LSTM and improved neural network al-
gorithms . Springer Soft Computing 2020, [Online]. Available:
[Link]
[5] Ce Ju, Dashan Gao, Ravikiran Mane, Ben Tan, Yang Liu, Cun-
tai Guan Federated Transfer Learning for EEG Signal Classifica-
tion . 42nd Annual International Conference of the IEEE Engineering
in Medicine & Biology Society (EMBC) 2020, [Online]. Available:
[Link]
[6] Aayesha, Muhammad Bilal Qureshi, Muhammad Afzaal, Muham-
mad Shuaib Qureshi, Muhammad Fayaz Machine learning-based
EEG signals classification model for epileptic seizure detection .
Springer Multimedia Tools and Applications 2021, [Online]. Available:
[Link]
[7] Emrah Aydemir, Turker Tuncer, Sengul Dogan A Tunable-Q wavelet
transform and quadruple symmetric pattern based EEG signal
classification method . Medical Hypotheses 2020, [Online]. Available:
[Link]
[8] Romain Atangana, Daniel Tchiotsop, Godpromesse Kenne, Laurent
Chane, DjoufackNkengfac k EEG Signal Classification using LDA and
MLP Classifier . Health Informatics - An International Journal 2020.
[9] Wenkai Huang, Yihao Xue, Lingkai Hu, Hantang Liuli S-
EEGNet: Electroencephalogram Signal Classification Based
on a Separable Convolution Neural Network With Bilinear
Interpolation . IEEE Access 2020, [Online]. Available:
[Link]
Fig. 10. Accuracy for Random Forest and GA.
[10] Lei Zhang EEG Signals Classification Using Machine Learning
for The Identification and Diagnosis of Schizophrenia . 41st
Annual International Conference of the IEEE Engineering in
Medicine and Biology Society (EMBC) 2019, [Online]. Available:
of additional machine learning algorithms and their compara- [Link]
tive analysis could provide further refinement in classification [11] Mohamad Amin Bakhshali, Morteza Khademi, Abbas Ebrahimi-
accuracy. Further research may delve into the integration Moghadam, Sahar Moghimi EEG signal classification of imagined
speech based on Riemannian distance of correntropy spectral density
of deep learning techniques, such as convolutional neural . Biomedical Signal Processing and Control 2020, [Online]. Available:
networks (CNNs) or recurrent neural networks (RNNs), to un- [Link]
ravel more intricate patterns within EEG signals. Additionally, [12] Ibrahim Aliyu, Chang Gyoon Lim Selection of optimal wavelet
features for epileptic EEG signal classification with LSTM .
the consideration of diverse datasets encompassing a broader Springer Neural Computing and Applications 2021, [Online]. Available:
spectrum of neurological disorders would contribute to the [Link]
generalizability of the classification model. Incorporating real- [13] Md. Asadur Rahman, Farzana Khanam, Mohiuddin Ahmad, Mo-
hammad Shorif Uddin Multiclass EEG signal classification utiliz-
time applications and deploying the model in clinical settings ing Rényi min-entropy-based feature selection from wavelet packet
could validate its effectiveness and pave the way for practi- transformation . Springer Brain Informatics 2020, [Online]. Avail-
cal implementations in diagnostic procedures. Collaborations able: [Link]
020-00108-y
with medical practitioners and neuroscientists could provide [14] Ashima Khosla, Padmavati Khandnor, Trilok Chand A
valuable insights and ensure the translation of our findings comparative analysis of signal processing and classification
into impactful clinical solutions. The continual evolution of methods for different applications based on EEG signals .
Biocybernetics and Biomedical Engineering 2020, [Online]. Available:
machine learning and signal processing methodologies offers [Link]
an exciting trajectory for future endeavors in EEG signal [15] M. A. Ganaie, M. Tanveer, Jatin Jangir EEG signal classifica-
classification. By addressing the outlined avenues for future tion via pinball universum twin support vector machine . Springer
Annals of Operations Research volume 2022, [Online]. Available:
work, we aspire to contribute to the ongoing efforts to enhance [Link]
our understanding of neurological conditions and improve [16] SVANTE WOLD, KIM ESBENSEN, PAUL GELADI Principal Com-
diagnostic tools for the benefit of individuals affected by such ponent Analysis . Chemometrics and Intelligent Laboratory Systems 1987.
[17] Tim Edwards Discrete wavelet transforms: Theory and implementation
disorders. . 1991.
[18] Tom V. Mathew Genetic Algorithm . 2012, [Online]. Avail-
able: [Link]
[Tom-Mathew].pdf
R EFERENCES [19] Yuhui Shi Particle Swarm Optimization . IEEE Neural Networks Society
2004.
[1] Jan Rabcan, Vitaly Levashenko, Elena Zaitseva, Miroslav Kvassay Re-
view of Methods for EEG Signal Classification and Development of
New Fuzzy Classification-Based Approach . IEEE Access 2020, [Online].
Available: [Link]
[2] Gaowei Xu, Xiaoang Shen, Sirui Chen, Yongshuo Zong, Canyang
Zhang, Hongyang Yue, Min Liu, Fei Chen, Wenliang Che A
Deep Transfer Convolutional Neural Network Framework for EEG
Signal Classification . IEEE Access 2019, [Online]. Available:
[Link]
 11

Mayur Akewar is an AI researcher and Alumni

of Department of Computer Science & Engineer-
ing, Shri Ramdeobaba College of Engineering &
Management. He has total 12 years of academic
and industrial experience. His background in AI
spans around machine learning, deep learning and
algorithms to deal with the problems related to
computer vision including Hyperspectral Imaging
(HSI) and natural language processing, and he has
a proven track record of developing AI-based so-
lutions, conducting research, or collaborating with
organizations. He developed and designed many algorithms and trained and
fine tuned many ML and DL models for applications of object detection,
image, video, audio and text classification and generation.