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

net/publication/381836178

A Novel Hybrid Deep Neural Network Classifier for EEG Emotional Brain Signals

Article in International Journal of Advanced Computer Science and Applications · June 2024
DOI: 10.14569/IJACSA.2024.01506107

CITATIONS READS
0 3

3 authors, including:

Abdelrahman Tarek Elgohr

Horus University
4 PUBLICATIONS 9 CITATIONS

SEE PROFILE

All content following this page was uploaded by Abdelrahman Tarek Elgohr on 30 June 2024.

The user has requested enhancement of the downloaded file.

 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

A Novel Hybrid Deep Neural Network Classifier for

EEG Emotional Brain Signals
Mahmoud A. A. Mousa1, Abdelrahman T. Elgohr2, Hatem A. Khater3
Faculty of Engineering, Zagazig University, Zagazig, Egypt1, 2
Faculty of Engineering, Horus University, Damietta Egypt2, 3
School of Mathematical and Computer Sciences, Heriot Watt University, Dubai, UAE1

Abstract—The field of brain computer interface (BCI) is one of BCI technology can be used to control assistive devices such as
the most exciting areas in the field of scientific research, as it can wheelchairs, prostheses and communication systems, as well as
overlap with all fields that need intelligent control, especially the to monitor brain activity and diagnose neurological diseases.
field of the medical industry. In order to deal with the brain and Moreover, BCI technology can be used to provide a more
its different signals, there are many ways to collect a dataset of natural form of human-computer interaction, allowing users to
brain signals, the most important of which is the collection of control computers with thoughts [1]. BCI technology can be
signals using the non-invasive EEG method. This group of data divided into two main categories as shown in Fig.1: invasive
that has been collected must be classified, and the features and noninvasive. Invasive BCI requires the insertion of
affecting changes in it must be selected to become useful for use in
electrodes into the brain in order to capture brain signals, which
different control capabilities. Due to the need for some fields used
in BCI to have high accuracy and speed in order to comply with
is a risky and complicated process. On the other hand,
the environment's motion sequences, this paper explores the noninvasive BCI relies on measuring signals from the scalp or
classification of brain signals for their usage as control signals in other parts of the body to detect brain activities. Noninvasive
Brain Computer Interface research, with the aim of integrating BCI is more commonly used and includes
them into different control systems. The objective of the study is electroencephalography (EEG), magnetoencephalography
to investigate the EEG brain signal classification using different (MEG), and functional near-infrared spectroscopy (fNIRS).
techniques such as Long Short-Term Memory (LSTM), EEG is the most widely used BCI technique and is based on
Convolutional Neural Networks (CNN), as well as the machine electrical signals generated by the brain [2].
learning approach represented by the Support Vector Machine
(SVM). We also present a novel hybrid classification technique
EEG signals are a type of electrical activity that can be
called CNN-LSTM which combines CNNs with LSTM networks. measured from the brain. They are used in a variety of
This proposed model processes the input data through one or more engineering fields, including medical, robotics, and computer
of the CNN’s convolutional layers to identify spatial patterns and engineering. In medical engineering, EEG signals are used to
the output is fed into the LSTM layers to capture temporal diagnose and monitor neurological conditions. EEGs can be
dependencies and sequential patterns. This proposed combination used to detect seizures, diagnose sleep disorders, and monitor
uses CNNs’ spatial feature extraction and LSTMs’ temporal brain activity during surgery. EEGs can also be used to measure
modelling to achieve high efficacy across domains. A test was done brain activity during cognitive tasks, such as memory tests. This
to determine the most effective approach for classifying emotional can help doctors better understand how the brain works and
brain signals that indicate the user's emotional state. The dataset how to treat neurological conditions [3]. In robotics
used in this research was generated from a widely available MUSE engineering, EEG signals are used to control robotic devices.
EEG headgear with four dry extra-cranial electrodes. The By measuring the electrical activity of the brain, robots can be
comparison came in favor of the proposed hybrid model (CNN- programmed to respond to certain commands. This can be used
LSTM) in first place with an accuracy of 98.5% and a step speed to create robots that can interact with humans in a more natural
of 244 milliseconds/step; the CNN model came in the second place way. For example, robots can be programmed to respond to
with an accuracy of 98.03% and a step speed of 58 facial expressions or voice commands [4]. In computer
milliseconds/step; and in the third place, the LSTM model engineering, EEG signals are used to create brain-computer
recorded an accuracy of 97.35% and a step speed of 2 sec/step;
interfaces. These allow users to control computers with their
finally, in last place, SVM came with 87.5% accuracy and 39
milliseconds/step running speed.
thoughts. This technology is still in its early stages, but it has
the potential to revolutionize the way we interact with
Keywords—BCI; EEG; Brain Signals Classification; SVM; computers [5].
LSTM, CNN; CNN-LSTM The most common EEG signal classification methods as
I. INTRODUCTION shown in Fig. 2 are supervised learning algorithms, such as
Support Vector Machines (SVMs), Artificial Neural Networks
Brain Computer Interface (BCI) is a technology that enables (ANNs), and decision trees. These algorithms are used to
direct communication between the brain and an external device identify patterns in EEG signals that can be used to diagnose
using signals generated from the brain. It has been proposed as and monitor neurological conditions [6]. For example, SVMs
a potential therapeutic treatment for various neurological can be used to classify EEG signals into different categories,
disorders and a tool for efficient human-computer interaction. such as normal or abnormal, or to detect changes in EEG signals

1045 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

over time. ANNs, which include Convolutional Neural basis function kernel that outperforms prior methods on the
Network (CNN), can be used to identify patterns in EEG signals same dataset achieves an accuracy of 95.70%. This article
that can be used to diagnose and monitor neurological presents a nonparametric method for decomposing EEG signals
conditions. Decision trees can be used to identify patterns in to improve efficiency. This approach can enhance the progress
EEG signals that can be used to diagnose and monitor of BCI system development by utilizing machine learning
neurological conditions [7]. techniques [11].
Chowdary MK, et al., (2022) aim to classify emotions from
electroencephalogram signals by utilizing different recurrent
neural network structures. Three architectures employed in this
study for emotion recognition using EEG signals are RNN
(recurrent neural network), LSTM (long short-term memory
network), and GRU (gated recurrent unit). Experimental data
confirmed the efficiency of these networks in terms of
performance measures. The study utilized the EEG Brain Wave
Dataset: Feeling Emotions and obtained an average accuracy of
95% for RNN, 97% for LSTM, and 96% for GRU in detecting
emotions [12].
Fig. 1. Brain computer interface technology [1]. EEG capture and emotion categorization in a simulated
driving environment is suggested by Chen J. et al. (2024) to
study panic emotion and accident-avoidance skills. The
program models obstacle avoidance at different risk levels
using vehicle speed. The system models the brain's
physiological structure for data processing using graph neural
networks (GNN) with functional connection and attention
mechanisms. Various research compared entropy and power
properties. The top single-label F1 score was 76.7%, and the
three-class classification was 75.26 % accurate. Binary
classification had 91.5% accuracy and the highest F1 score for
Fig. 2. Dataset classification techniques [8]. a single label was 91.86%. Deep learning algorithms can
accurately mimic hazardous events, record the driver's EEG
A. Related Work data, and quickly track emotional states, according to
Z. -T. Liu et al. (2019) tested a proposed approach on DEAP experiments [13].
dataset, classifying Valance and Arousal emotional states using This research investigates classifying brain signals for use
K-nearest neighbor and support vector machine. The as control signals in Brain-Computer Interface (BCI) systems
experiments compare temporal windows of different lengths designed for various robotic applications. The aim is to
and three EEG signal rhythms. The results show that the EEG compare four methods for multi-class classification: Long
signal with one temporal window has the highest recognition Short-Term Memory (LSTM) and Convolutional Neural
accuracy of 86.46%. A multimodal emotional communication- Networks (CNN) from deep learning, a proposed hybrid CNN-
based humans-robots interaction system would use the LSTM approach, and Support Vector Machine (SVM) from
suggested approach for real-time emotion identification [9]. machine learning. Ultimately, this research seeks to determine
T. Song et al. (2020) to recognize emotions in multichannel the most effective method for classifying emotional brain
EEG data used a dynamical graph convolutional neural network signals that reflect the user's emotional state.
(DGCNN). Our EEG emotion recognition method uses a graph The rest of this paper is organized as follows: Section II
to describe multichannel EEG data and classify emotions using demonstrates the main concepts for signals classification
this model. EEG emotion recognition is improved by learning overview; Section III presents the classification models;
new features from the adjacency matrix. Emotion EEG datasets Section IV describes the dataset; Sections V and VI elaborate
SEED and DREAMER were extensively studied. The proposed the classification results and a discussion of the results
recognition method is more accurate than current methods. On generated from the tests; Section VII mentions the applications
SEED, it averaged 90.4% in subject-dependent experiments that can benefit from this research topic; Section VIII concludes
and 79.95% in subject-independent cross-validation [10]. the paper and presents the future work.
In 2020, S. K. Khare and colleagues introduced an adaptive II. CLASSIFICATION OVERVIEW
tunable Q wavelet transform for selecting tuning parameters
automatically. Grey wolf optimization identifies the best tuning Dataset classification is a process of organizing data into
parameters. GWO tuning parameters divide EEG signals into categories based on certain characteristics. It is a way of
sub bands. Time-domain properties of SB are inputted into a organizing data into meaningful groups so that it can be more
multiclass least-squares support vector machine. Evaluating the easily analyzed and understood. Dataset classification is used
classification accuracy of four main emotions - happiness, fear, in a variety of fields, including data mining, machine learning,
sadness, and relaxation - compared to current methods. A radial and artificial intelligence.

1046 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

reduce the complexity of the model and prevent overfitting.

SVMs are also very efficient in terms of both time and memory.
This is because they only need to store a subset of the training
data, which makes them very efficient in terms of memory
usage. In addition, SVMs are very versatile and can be used for
a variety of tasks such as classification, regression, and outlier
detection [17].
Building a Support Vector Machine (SVM) algorithm with
Python as shown in Algorithm 1 [18], is a relatively
straightforward process. The first step is to import the necessary
libraries. The most common libraries used for SVM in Python
are Scikit-learn, Numpy, and Matplotlib. Once the libraries are
imported, the next step is to prepare the data. This involves
loading the data into a Pandas Data Frame, cleaning the data,
and splitting it into training and testing sets. It is important to
ensure that the data is properly scaled and normalized before
training the model. The next step is to create the SVM model.
This is done by instantiating an SVM classifier object from the
Scikit-learn library. The classifier object can then be fitted to
the training data using the fit() method. Once the model is
trained, it can be used to make predictions on the test data. This
is done by calling the predict() method on the classifier object.
The predictions can then be evaluated using a variety of metrics
Fig. 3. Dataset classification process overview [14]. such as accuracy, precision, recall, and F1 score [19].

The process of dataset classification, as shown in Fig. 3 Algorithm 1: SVM model

begins with the identification of the data that needs to be Input: X (array of input data (features)), Y (array of output data
classified. This data can come from a variety of sources, such (classes - labels))
as databases, text documents, images, and audio files. Once the Output: performance of model (accuracy – precision – confusion
data has been identified, it is then divided into categories based matrix)
on certain characteristics. These characteristics can include 1. Function:
size, type, content, and other attributes.
Training _ SVM
Once the data has been divided into categories, it is then clf = svm.SVC(kernel='kernal type')
analyzed to determine the relationships between the different clf.fit(X_train, y_train)
categories. This analysis can be done using a variety of
techniques, such as clustering, decision trees, and neural
2. Initialize:
networks. The goal of this analysis is to identify patterns and
trends in the data that can be used to make predictions or Learning rate – Number of runs (epoch)
decisions. Once the data has been classified and analyzed, it can for i in X array
then be used for a variety of purposes. For example, it can be if (Y(i) x X(i) x q) > 1
used to create predictive models, to identify customer segments, then
or to detect anomalies in the data. It can also be used to create update: q = q + learning rate x ((X(i)*Y(i))*(-2*(1/epoch)*q)
visualizations of the data, such as charts and graphs, which can else
be used to better understand the data [15]. update: q = q + learning rate x (-2*(1/epoch)*q)
III. CLASSIFICATION MODELS end if
end
A. Support Vector Machine (SVM)
Support Vector Machines (SVMs) are a powerful and In the context of multi-class classification, SVMs can be
versatile machine learning algorithm used for classification and used to construct a maximum-margin hyperplane that divides
regression tasks. SVMs are a supervised learning algorithm that the feature space into regions, each corresponding to a
can be used to classify data into two or more classes. They are particular class. The algorithm then searches for the optimal
based on the concept of finding a hyperplane that best divides a hyperplane that maximizes the margin between the classes.
dataset into two classes. The main advantage of SVMs is that This hyperplane is then used to classify new data points. The
they are very effective in high dimensional spaces. This is advantage of SVMs is that they can be used to classify data with
because they use a kernel trick to map the data into a higher a large number of features and classes, as well as data with non-
dimensional space, where it can be separated by a hyperplane. linear boundaries. Furthermore, SVMs are robust to outliers and
This allows them to capture complex relationships between the can be used to classify data with a high degree of accuracy [19],
data points [16]. SVMs are also very robust to overfitting. This [20].
is because they use a regularization parameter which helps to

1047 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

B. Long Short Term Memory (LSTM) the optimizer, and the metrics to be used for evaluating the
The Long Short-Term Memory (LSTM) classifier is a model. Finally, the model can be trained. This involves
powerful deep learning algorithm that can be used to classify specifying the number of epochs, the batch size, and the
data. It is a type of recurrent neural network (RNN) that is validation split. It is important to monitor the training process
capable of learning long-term dependencies in data. The LSTM to ensure that the model is not overfitting or underfitting the
classifier is a powerful tool for predicting and classifying data, data [22].
and it has been used in a variety of applications, such as natural C. Convolutional Neural Network (CNN)
language processing, speech recognition, time series
forecasting, and classifying sequences of data, such as text, A convolutional neural network (CNN) is a type of artificial
audio, and video. As shown in Fig. 4, the LSTM classifier is neural network used in deep learning that is specifically
composed of a series of memory cells, each of which contains designed to process data that has a grid-like structure, such as
a set of weights and biases. The weights and biases are adjusted tabular and images datasets. CNNs are composed of multiple
during the training process to learn the patterns in the data [21]. layers of neurons that each perform a specific task as shown in
The memory cells are connected in a chain, and each cell is Fig. 5. The first layer of neurons is responsible for detecting
connected to the next cell in the chain. This allows the network edges and other basic features in the input image. The second
to remember information from previous cells and use it to make layer of neurons is responsible for detecting more complex
predictions [22]. The LSTM classifier is trained using a features, such as shapes and patterns. The third layer of neurons
supervised learning algorithm. During the training process, the is responsible for recognizing objects in the image. The fourth
network is presented with a set of input data and the desired layer of neurons is responsible for recognizing more complex
output. The network then adjusts the weights and biases of the objects, such as faces or animals [24].
memory cells to learn the patterns in the data. Once the training CNNs are particularly useful in robotics because they are
is complete, the network can be used to make predictions on able to process large amounts of data quickly and accurately.
new data [23]. For example, a CNN can be used to identify objects in an image
or video feed. It can also be used to analyze cognitive data
represented in a database that enables robots to understand the
surrounding environment and also understand the commands
stored within it and classify them according to the event the
robot is exposed to. This is useful for robots that need to identify
objects in their environment in order to navigate or interact with
them. CNNs can also be used to classify objects in a scene,
which is useful for robots that need to recognize and interact
with objects in their environment [25].

Fig. 4. LSTM classifier model flowchart [21].

Fig. 5. Convolutional Neural Network (CNN) architecture.
Building an LSTM model with Python is a great way to get
Building a Convolutional Neural Network (CNN) model is
started with deep learning. The first step in building an LSTM
a complex process that requires a lot of knowledge and
model with Python is to import the necessary libraries. The
experience. However, with the right guidance, it can be done
most popular library for deep learning in Python is TensorFlow,
relatively easily. The following steps outline the process of
which provides a high-level API for building and training
building a CNN model as described in Algorithm 2 [26]:
neural networks. Other popular libraries include Keras,
PyTorch, and Theano. Once the libraries are imported, the next  Data Preparation: The first step in building a CNN
step is to prepare the data. This involves loading the data, model is to prepare the data. This includes gathering the
preprocessing it, and splitting it into training and test sets. It is data, cleaning it, and formatting it into a suitable format
important to ensure that the data is properly normalized and for the model. This step is important as it ensures that
scaled before training the model. The next step is to define the the model is trained on the most accurate and up-to-date
model architecture. This involves specifying the number of data.
layers, the number of neurons in each layer, the type of
activation functions, and the type of optimizer. It is also  Model Architecture: The next step is to decide on the
important to specify the input and output shapes of the model. model architecture. This includes deciding on the
Once the model architecture is defined, the next step is to number of layers, the type of layers, and the number of
compile the model. This involves specifying the loss function, neurons in each layer. This step is important as it

1048 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

determines the complexity of the model and how well it spatial and temporal data. This powerful combination leverages
will perform. CNNs' ability to extract spatial features and LSTMs' strength in
temporal modeling, leading to high effectiveness across various
 Training: Once the model architecture is decided, the domains.
next step is to train the model. This involves feeding the
data into the model and adjusting the weights and biases This hybrid model finds applications in tasks involving
of the neurons in order to minimize the error. This step complex sequential data. It utilizes CNNs for spatial analysis
is important as it ensures that the model is able to and LSTMs for understanding temporal sequences. The CNN-
accurately predict the output given the input. LSTM model processes input data through one or more
convolutional layers to identify spatial patterns. The output
 Evaluation: After the model is trained, the next step is to from these layers then feeds into LSTM layers to capture
evaluate the model. This involves testing the model on temporal dependencies and sequential patterns. Finally, dense
unseen data and measuring its performance. This step is layers are often used for classification or regression tasks.
important as it allows us to determine how well the Algorithm 3 lists the whole process of proposed model.
model is performing and if it needs to be improved.
The model's strength lies in the specialized functions of its
 Deployment: The final step is to deploy the model. This layers. CNNs excel at extracting features from spatial data,
involves making the model available to users so that while LSTMs represent complex temporal connections. This
they can use it to make predictions. This step is combination allows the model to learn both spatial and
important as it allows the model to be used in real-world temporal characteristics simultaneously, enabling a
applications. comprehensive interpretation of the data. However, achieving
These are the basic steps for building a CNN model. optimal performance requires careful hyperparameter tuning
However, there are many other steps that can be taken to for both CNN and LSTM components, and ensuring
improve the model, such as hyperparameter tuning, compatibility between the input data shape and both layer types.
regularization, and data augmentation. With the right guidance A small code example using the Keras library shows how to
and experience, building a CNN model can be a relatively sequentially add CNN and LSTM layers for spatiotemporal
straightforward process. modelling [27].

Algorithm 2: CNN model Algorithm 3: CNN-LSTM model

Input: tabular EEG emotional brain signal dataset Input: tabular EEG emotional brain signal dataset
Output: confusion matrix and model testing accuracy Output: confusion matrix and model testing accuracy
1. Import necessary libraries 1. Import necessary libraries
(Numpy as np, pandas as pd, tensorflow as tf, Sequential, Dense, (Numpy as np, pandas as pd, tensorflow as tf, Sequential, Dense,
Conv1D, MaxPooling1D, and Flatten) Conv1D, MaxPooling1D, and Flatten)
2. Load the emotional dataset 2. Load the emotional dataset
dataset = pd read _ datatype ('dataset . datatype') dataset = pd read _ datatype ('dataset . datatype')
3. Analysis the dataset 3. Analysis the dataset
Input signals = dataset drop (columns = ['target columns']) Input signals = dataset drop (columns = ['target columns'])
Labels = dataset ['last column'] Labels = dataset ['last column']
4. Split the dataset into training and testing sets 4. Split the dataset into training and testing sets
data train, data test, labels train, labels test = train test split (data, data train, data test, labels train, labels test = train test split (data,
labels, test size = test ratio to complete dataset, random state = no. of labels, test size = test ratio to complete dataset, random state = no. of
states) states)
5. Build the CNN model 5. Build the CNN-LSTM model
model = Sequential ([ model = Sequential ([
Conv1D parameter definition (filters, kernel size, activation  Conv1D parameters definition (filters, kernel size, activation
functions) functions)
Input shape = X train shape. Input shape = X train shape.
MaxPooling1D size. MaxPooling1D size.
Dense (output layer count, output activation function  LSTM parameters definition (units’ size, return sequences)
6. Train the model  Dense (output layer count, output activation function)
Training history = model fit (data train, labels train, epochs number, 6. Train the model
batch size, validation data (data test, labels test)) Training history = model fit (data train, labels train, epochs number,
7. Evaluate the model batch size, validation data (data test, labels test))
loss accuracy = model evaluate (data test, labels test) 7. Evaluate the model
print Test Loss loss accuracy = model evaluate (data test, labels test)
print Test Accuracy print Test Loss
print confusion matrix print Test Accuracy
print confusion matrix
D. CNN-LSTM Hybrid Model
The CNN-LSTM model, which combines Convolutional IV. DATASET DESCRIPTION
Neural Networks (CNNs) with Long Short-Term Memory Datasets can be used to analyze trends, identify patterns,
(LSTM) networks, excels at modeling the interdependence of and make predictions. They can also be used to compare

1049 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

different groups of people or different types of data, store identify patterns and relationships in data. Classification results
information about people, places, events, and other topics, and are used to make predictions about future data points, and can
create visualizations like charts, graphs, and maps. They can be used to make decisions about how to best utilize resources.
even be used to generate reports and presentations. Classification results are typically presented in the form of a
confusion matrix, which is a table that shows the number of true
Datasets can be used to analyze trends, identify patterns, positives, false positives, true negatives, and false negatives
and make predictions. They can also be used to compare [30]. The confusion matrix is used to evaluate the accuracy of
different groups of people or to compare different types of data. the classification model, and can be used to identify areas where
They can also be used to store information about people, places, the model is performing well or poorly. Classification results
events, and other topics. Datasets can be used to create can also be used to identify important features in the data that
visualizations, such as charts, graphs, and maps. They can also are driving the model’s predictions. This can be done by
be used to create reports and presentations. They can also be looking at the feature importance scores, which are calculated
used to store information about people, places, events, and other by the model and indicate how important each feature is in
topics [28]. making the prediction. This can be used to identify which
The dataset used in this research is a mental emotional features are most important for making accurate predictions,
sentiment dataset that was collected by other researchers using and can be used to inform decisions about which features to
a commercial MUSE EEG headband which was used with a focus on when building a model [31], [32].
resolution of four (TP9, AF7, AF8, TP10) electrodes. To collect
the data, researchers used a widely available MUSE EEG
headgear with four dry extra-cranial electrodes. As can be seen
in Fig. 6, micro voltage readings are taken from electrodes TP9,
AF7, AF8, and TP10. Two individuals (1 male, 1 female, aged
20-22) each provided 60 seconds of data for each of the 6 film
segments, for a total of 12 minutes (720 seconds) of brain
activity data (6 minutes for each emotional state). A total of 36
minutes of EEG data was obtained from each individual,
including six minutes of "neutral brainwave" data. The brain's
waves were captured at a variable frequency and then
resampled to 150Hz, yielding a collection of 324,000 data Fig. 7. Classes appearance analysis.
points. The positive and negative valence descriptors were
evaluated instead of the emotions themselves to determine Finally, classification results can be used to compare
which activities were most likely to elicit. For a third category, different models and determine which one is the best for a given
representing the subject's baseline emotional state, neutral data task. This can be done by looking at the accuracy scores of each
were also obtained before any data on emotions were gathered model, as well as other metrics such as precision, recall, and F1
(to prevent contamination from the latter). We only gathered score. Comparing the results of different models can help
data from each participant for three minutes every day to identify which model is best suited for a given task, and can
minimize the influence of a baseline emotional state [29]. help inform decisions about which model to use [32].
To classify the dataset, it must first understand its details,
the resultant classes from each row of input, and the number of
instances of each class over the whole dataset. As the
information from the dataset were analyzed, it was discovered
that there are three separate classes as a consequence of all the
input rows, which are positive, negative, and neutral, as they
represent an indicator of the subject's emotional state. After
each class was counted, it was discovered that the positive case
occurred 708 times, the negative case appeared 708 times, and
the neutral case appeared 716 times as shown in Fig. 7. These
statistics reveal the dataset's balance, from which the difference
in findings may be precisely calculated. As the last stage in
studying the dataset, a sample may be obtained for each class
using a variety of inputs, as illustrated in Fig. 8.
A. SVM Results
When implementing the SVM algorithm, the tensorflow
library was used for deep learning in Python, on the Kaggle
Fig. 6. Position of used EEG electrodes on human skull [29]. coding website. And by classifying the studied dataset, and
specifying each of the training data percentage as 50%, the
V. CLASSIFICATION RESULTS testing data percentage as 25%, the validation data percentage
Classification results are the outcomes of a classification as 25%, and 100 epochs to test the algorithm and conclude the
process, which is a type of data mining technique used to best classification result.

1050 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

because the dataset contains more than two expected results (3

classes), the Softmax Activation Function is used.

Fig. 8. Classes sample.

The classification results of the algorithm came with an

accuracy of 87.5%, after only 15 epochs (early stop), radial
basis function (RBF) kernel [33], and 39 ms/step running speed.
When viewing the confusion matrix as shown in Fig. 9, which
describe matching between actual label and predicted label. It
can be seen that the results are not mixed up, or in another sense,
the algorithm is not confused between dataset classes when
Fig. 9. Confusion Matrix of SVM model result.
determining the result significantly.
B. LSTM Results
The Long Short-Term Memory (LSTM) model was
developed on the Kaggle coding platform using the Python
tensorflow deep learning framework. By categorising the
examined dataset, designating the percentage of each training
data as 50%, the percentage of test data as 25%, and the
percentage of validation data as 25%, 100 epochs for model
testing, and selecting the best classification result. This model
was built to contain the input layer, and the last layer is
responsible for the output, and because the dataset contains
more than two expected results (3 classes), the Softmax
Activation Function is used [34].
The algorithm's classification results showed an accuracy of
97.35% after just 38 epochs (early stopping) and a running
speed of 2 s/step, which is an excellent result. This result was
reached by setting the learning rate to 0.001 and using Adam as
the model's optimization library. In addition, through Fig. 10, it
can be noted that the classification model was very sharp in
showing the results, as it was not confused with the actual result
Fig. 10. Confusion matrix for LSTM model results.
of implementing the classification except in very simple cases
that do not exceed 0.06 for each class confused with other The algorithm's classification results came with an accuracy
classes. of 98.03 % after only 38 epochs (early stopping) and 58 ms/step
C. CNN Results running speed, which is a great result. This result was obtained
as a result of setting the learning rate to 0.001 and using
On the Kaggle coding platform, the tensorflow deep
Adamax as the optimization library on the model. Moreover, it
learning library in Python was utilized to create the
can be shown in Fig. 11 that the classification model was
convolutional neural network model. By classifying the
extremely crisp in displaying the results, as it was not confused
investigated data set, defining the percentage of each training
with the real result of implementing the classification except in
data as 50%, the proportion of test data as 25%, and the
very basic examples where 0.03 for each class confused with
proportion of validation data as 25%, 100 epochs for model
other classes was not exceeded.
testing, and identifying the best classification result.
This model was built to contain the input layer, five hidden D. CNN-LSTM Results
layers all of which contain an activation function of the type The CNN-LSTM model was implemented on the Kaggle
Rectified Linear Unit (ReLU) [35], and this function is coding platform using the Python tensorflow deep learning
considered one of the best choices in the selection of the framework. By classifying the dataset, allocating 50% of the
activation. The last layer is responsible for the output, and data for training, 25% for testing, and 25% for validation,

1051 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

conducting 50 epochs for model evaluation, and choosing the VI. RESULT AND DISCUSSION
optimal classification outcome. This model was constructed When comparing the results of different models to classify
with an input layer, three CNN layers utilizing the ReLU the studied dataset (SVM, LSTM, CNN, and CNN-LSTM),
activation function and progressively increasing filter sizes more than one aspect can be relied on for comparison, the first
starting from 64 bits. It also includes LSTM layers and a final and most important of which is the accuracy of the
output layer. Since the dataset consists of three distinct classes, classification when implementing the model, the second is the
the Softmax Activation Function is employed. confusion matrix, which is related in one way or another to the
first factor in the comparison, and the third factor that was taken
into account when comparing is speed of implementation of the
model, as measured by the speed of the test steps and also the
speed of early stopping when testing the model. Table I
compiles these features for all models utilized in the paper.

TABLE I. COMPARISION FACTOR CONCLUSION

Comparison factors
Model
Accuracy Test speed Early stop
SVM 87.5 % 39 ms/step 15 epochs
LSTM 97.35 % 2 sec/step 15 epochs
CNN 98.03 % 58 ms/step 38 epochs
CNN-LSTM 98.50 % 244 ms/step No

With regard to the first factor in the comparison, which is

the accuracy of the model in implementation, it came in the
foreground, and it is considered one of the best classification
Fig. 11. Confusion matrix for CNN model results. results applied to the studied data set. It is the result of
classification using CNN-LSTM with an accuracy of 98.50%.
Then it comes in second place, and not by a large difference, is
the result of classification using CNN, with an accuracy of 98
%, while in third place was the LSTM model with 97.35 %
accuracy, finally SVM where the classification accuracy was
not good enough compared to the previous three models with
an accuracy of 87.5 %.
As mentioned previously, the confusion matrix is linked to
the accuracy of the classification, or in other words, this matrix
is a breakdown of the characteristics of the classification result
that lead to its accuracy. The order of the models when
comparing the results based on the quality of the matrix came
in the same order as the models in terms of accuracy.
As for the third factor in the comparison, it is actually
divided into two different factors, which are the speed of
implementation by step and the speed of implementation in
early stopping when testing the model. In view of the speed of
execution by step, the SVM model came in first place with a
speed of 39 ms/step, and in second place came the CNN model
Fig. 12. CNN-LSTM confusion matrix. with 58 ms/step, and the CNN-LSTM model came in third place
with a large difference from its predecessors with 244 ms/step.
The algorithm achieved a classification accuracy of 98.50% and in last place LSTM with extreme test step speed time with
after completing 100 epochs without early stopping. 2 sec/step However, when looking at the speed of early
Additionally, it demonstrated a running speed of 244 ms/step, stopping, the order can differ relatively, as the SVM model
which is considered an outstanding outcome. The attainment of comes in first place equally with the LSTM model by stopping
this outcome was accomplished by configuring the learning rate after only 15 epochs out of 100 epochs that were specified for
to 0.001 and employing Adam as the optimization library for implementation, and the CNN-LSTM model remains in last
the model. Furthermore, Fig.12 demonstrates that the place with no early stop out of 100 epochs also for
classification model exhibited a high level of accuracy in implementation.
presenting the findings, as it only encountered confusion with As a result of this comparison, it can be concluded that
the actual outcome of the classification in rare instances that did although the classification of the CNN-LSTM is the best as a
not surpass 0.06 for each misclassified class. model for classification, it must be taken into account that the

1052 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

previous results are dependent on the input factors of each VII. APPLICATIONS
model, which were fixed in all cases, so that comparison can be Brain Computer Interfacing (BCI) is a technology, which
made based on the equality of classification characteristics. enables communication between humans and machines through
The CNN-LSTM was the best of them, as it has the highest the direct interpretation of brain activities. This technology has
accuracy, which is the most important factor in the comparison. a wide range of potential applications as shown in Fig. 13 and
In addition, the execution speed (step speed) was not bad has been explored by researchers in fields such as medical
(between the other models), but despite that, it was the most in diagnostics, prosthetics, human-machine interaction, and
the number of epochs that the model needed to infer the best communication aids [36].
classification result (training the model) but this did not
significantly affect the outcome of the total execution time of
the model.
And if we exclude the factor of execution speed, then CNN-
LSTM, CNN and LSTM are very close in the result of
classification accuracy, then these models can be equally
reliable on the classification of the data set. In contrast, if the
accuracy factor is excluded, the SVM model is the fastest in
step speed and the least in the number of epochs required to
train the model equally with the LSTM model, so SVM can be
said that it is the best in the speed factor.
Considering that the CNN-LSTM result is the best model
studied in this paper in terms of accuracy, which is the most
reliable factor in the comparison. In the end, this model can also
be compared with the similar models previously published on a
similar database and with the different models. Through the Fig. 13. BCI applications.
previous literature study during previous years, published
research showed limited accuracy of the techniques used, as K- After classifying the data set, the classification model may
nearest neighbor recorded an accuracy of 86.46% in Z-T. Liu. be used to connect the model's inputs, which are brain signals,
et al.'s 2019 research. T. Song et al. also recorded in their and any application that can be controlled by the dataset's
research published in 2020, an accuracy of 90.4% using distinct classifications. When looking at the field of medical
DGCNN. Also, it was followed by the accuracy of the research industries, it is possible to coordinate between commands to
of S.K. Khare et al. 2020, which reached 95.7% using the control prosthetic limbs through brain signals directly, through
LSSVM, and then the RNN, LSTM, and GRU models obtained three different commands linked to the three groups of the data
an accuracy of 95%, 97%, and 96% respectively in the research set, for patients with paraplegia or total paralysis who are
of Chowdary MK et al. 2022. Finally, Chen J. et al. 2024 are unable to move their natural organs, or move the muscles to
achieved 75.26% multi-classification accuracy by using GNN control the Industrial limb [37]–[39]. Among the applications
model in their published study. On the other hand, expectations is Brain control in industrial robots in smart industries. It is also
were raised for an impressive result using the studied model possible to link the results of classification (one of the classes
(CNN-LSTM), where a classification accuracy was obtained for the data set) and a set of successive commands that include
that achieved a gorgeous mark in prediction. A summary of the a path for a complete industrial process, so that the controller
results of the reviewed researches appears in Table II, where has the ability to control the brain in three separate industrial
each row indicates the research person, the model used in it, and processes [40]. In addition, brain control technology can be
the model’s accuracy result during testing. used to control laboratory robots, and this can be used in
research and scientific projects that allow the formation of
TABLE II. LITERATURE RESULTS SUMMARY COMPARED WITH PROPOSED innovative systems of intelligent control [41].
MODEL
One of the most exciting applications of BCI in gaming and
Comparison entertainment is the ability to control game characters and
Author
Model Accuracy objects with your thoughts. This could allow players to control
Z-T. Liu et al. (2019) [9] K-NN 86.46 % their characters in a more natural and intuitive way, as well as
allowing for more complex interactions with the game world.
T. Song et al. (2020) [10] DGCNN 90.40 %
For example, a player could use their thoughts to control a
S.K. Khare et al. (2020) [11] LSSVM 95.70 % character’s movements, or to manipulate objects in the game
RNN 95.00 % world. This could open up a completely new level of immersion
and interaction with games [42].
Chowdary MK et al. (2022) [12] LSTM 97.00 %
GRU 96.00 % Brain-Computer Interface (BCI) technology might change
how individuals see themselves and their surroundings. BCI
Chen J. et al. (2024) [13] GNN 75.26 % technology raises ethical and legal issues. BCI technology may
Proposed model CNN-LSTM 98.50 % enhance quality of life and give therapeutic advantages, but also

1053 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

presents privacy, autonomy, and informed consent problems REFERENCES

[43]. [1] S. N. Abdulkader, A. Atia, and M. S. M. Mostafa, “Brain computer
interfacing: Applications and challenges,” Egyptian Informatics Journal,
VIII. CONCLUSION AND FUTURE WORK vol. 16, no. 2. Elsevier B.V., pp. 213–230, Jul. 01, 2015. doi:
10.1016/j.eij.2015.06.002.
The domain of Brain-Computer Interface (BCI) stands as an
exceptionally captivating realm of scientific inquiry due to its [2] Anupama H S, N. K. Cauvery, and Lingaraju G M, “Brain Computer
Interface and Its Types-A Study,” Int J Adv Eng Technol, vol. 3, no. 2,
potential intersections with diverse industries, particularly pp. 739–745, 2012, Accessed: Feb. 22, 2024. [Online]. Available:
those requiring intelligent control, such as industry and https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=15cd
medicine. Various methodologies are employed to assemble 5fc6fd521f60cc35a2f4079b14360601117a.
datasets of cerebral signals for the comprehensive [3] M.-P. Hosseini, A. Hosseini, and K. Ahi, “A Review on Machine
understanding of the intricate signals emanating from the brain. Learning for EEG Signal Processing in Bioengineering,” IEEE Rev
Among these methodologies, the non-invasive Biomed Eng, vol. 14, pp. 204–218, 2021, doi:
10.1109/RBME.2020.2969915.
Electroencephalogram (EEG) method holds particular
[4] L. Bi, X.-A. Fan, and Y. Liu, “EEG-Based Brain-Controlled Mobile
significance. The acquired dataset necessitates meticulous Robots: A Survey,” IEEE Trans Hum Mach Syst, vol. 43, no. 2, pp. 161–
categorization, wherein the identification of influential 176, 2013, doi: 10.1109/TSMCC.2012.2219046.
characteristics responsible for inducing changes becomes [5] R. Bhavsar, Y. Sun, N. Helian, N. Davey, D. Mayor, and T. Steffert, “The
imperative for its applicability across diverse control correlation between EEG signals as measured in different positions on
modalities. scalp varying with distance,” in Procedia Computer Science, Elsevier
B.V., 2018, pp. 92–97. doi: 10.1016/j.procs.2018.01.015.
Furthermore, the demand for precision and expeditious [6] J. das C. Rodrigues, P. P. R. Filho, E. Peixoto, A. K. N, and V. H. C. de
processing in BCI applications, especially in alignment with Albuquerque, “Classification of EEG signals to detect alcoholism using
dynamic environmental motion sequences, prompted a machine learning techniques,” Pattern Recognit Lett, vol. 125, pp. 140–
comparative evaluation of four alternative classification 149, 2019, doi: https://doi.org/10.1016/j.patrec.2019.04.019.
models, namely Support Vector Machine (SVM), Long Short- [7] H. Dose, J. S. Møller, H. K. Iversen, and S. Puthusserypady, “An end-to-
end deep learning approach to MI-EEG signal classification for BCIs,”
Term Memory (LSTM), Convolutional Neural Network Expert Syst Appl, vol. 114, pp. 532–542, 2018, doi:
(CNN), and the hybrid CNN-LSTM model. The findings of this https://doi.org/10.1016/j.eswa.2018.08.031.
comparative analysis underscore the notable efficacy of the [8] J. Abdillah, I. Asror, and Y. F. A. Wibowo, “Emotion Classification of
CNN-LSTM model, manifesting an accuracy of 98.5% Song Lyrics using Bidirectional LSTM Method with GloVe Word
alongside an operational speed of 244 milliseconds per step. Representation Weighting,” Jurnal RESTI (Rekayasa Sistem Dan
Following suit, the CNN model secured the second position, Teknologi Informasi), vol. 4, no. 4, pp. 723–729, 2020, Accessed: Feb.
22, 2024. [Online]. Available:
achieving an accuracy of 98% with a step speed of 58 http://jurnal.iaii.or.id/index.php/RESTI/article/download/2156/284.
milliseconds per step. Occupying the third position, the LSTM [9] Z.-T. Liu, Q. Xie, M. Wu, W.-H. Cao, D.-Y. Li, and S.-H. Li,
model demonstrated an accuracy of 97.35%, albeit at a “Electroencephalogram Emotion Recognition Based on Empirical Mode
comparatively slower step speed of 2 seconds per step. Decomposition and Optimal Feature Selection,” IEEE Trans Cogn Dev
Conclusively, the SVM model finalized the comparison, Syst, vol. 11, no. 4, pp. 517–526, Dec. 2019, doi:
registering an accuracy of 87.5% and a step speed of 39 10.1109/TCDS.2018.2868121.
milliseconds per step. These findings accentuate the CNN- [10] T. Song, W. Zheng, P. Song, and Z. Cui, “EEG Emotion Recognition
Using Dynamical Graph Convolutional Neural Networks,” IEEE Trans
LSTM model's prowess in BCI applications, positioning it as Affect Comput, vol. 11, no. 3, pp. 532–541, Jul. 2020, doi:
the preeminent choice for striking a commendable equilibrium 10.1109/TAFFC.2018.2817622.
between accuracy and processing speed within dynamic [11] S. K. Khare, V. Bajaj, and G. R. Sinha, “Adaptive Tunable Q Wavelet
environmental contexts. Transform-Based Emotion Identification,” IEEE Trans Instrum Meas,
vol. 69, no. 12, pp. 9609–9617, Dec. 2020, doi:
Future work can involve two avenues: exploring the 10.1109/TIM.2020.3006611.
accuracy of other classification models and developing entirely [12] M. K. Chowdary, J. Anitha, and D. J. Hemanth, “Emotion Recognition
new ones to advance the field. Additionally, we can investigate from EEG Signals Using Recurrent Neural Networks,” Electronics
the creation of an integrated system utilizing brain signals for (Basel), vol. 11, no. 15, Jul. 2022, doi: 10.3390/electronics11152387.
control and evaluate its overall performance in accurately [13] J. Chen, X. Lin, W. Ma, Y. Wang, and W. Tang, “EEG-based emotion
executing commands based on the underlying classification recognition for road accidents in a simulated driving environment,”
Biomed Signal Process Control, vol. 87, no. 8, Jan. 2024, doi:
accuracy. 10.1016/j.bspc.2023.105411.
ACKNOWLEDGMENT [14] B. Chakravarthi, S. C. Ng, M. R. Ezilarasan, and M. F. Leung, “EEG-
based emotion recognition using hybrid CNN and LSTM classification,”
We would like to thank Heriot-Watt University for its Front Comput Neurosci, vol. 16, Oct. 2022, doi:
support to publish this paper. We also thank Horus University 10.3389/fncom.2022.1019776.
(Egypt) for its explicit support throughout this research and for [15] M. Besserve, K. Jerbi, F. Laurent, S. Baillet, J. Martinerie, and L. Garnero,
providing access to high featured computer labs for models’ “Classification methods for ongoing EEG and MEG signals,” Biol Res,
vol. 40, no. 4, pp. 415–437, 2007, doi: http://dx.doi.org/10.4067/S0716-
implementation. 97602007000500005.
Kaggle platform was used as a workspace for modeling the [16] M. Yoshikawa, M. Mikawa, and K. Tanaka, “A myoelectric interface for
studied approaches and implementing the proposed models. robotic hand control using support vector machine,” in 2007 IEEE/RSJ
International Conference on Intelligent Robots and Systems, 2007, pp.
2723–2728. doi: 10.1109/IROS.2007.4399301.

1054 | P a g e
www.ijacsa.thesai.org
 (IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 15, No. 6, 2024

[17] T. N. T. Thu and V. D. Xuan, “Using support vector machine in FoRex Machine Learning Section A-Research paper Eur,” Chem. Bull, vol. 2023,
predicting,” in 2018 IEEE International Conference on Innovative no. 8, pp. 7648–7656, doi: 10.48047/ecb/2023.12.8.619.
Research and Development (ICIRD), 2018, pp. 1–5. doi: [31] T. Kaur and T. K. Gandhi, “Deep convolutional neural networks with
10.1109/ICIRD.2018.8376303. transfer learning for automated brain image classification,” Mach Vis
[18] K. Harimoorthy and M. T, “Multi-disease prediction model using Appl, vol. 31, no. 3, p. 20, 2020, doi: 10.1007/s00138-020-01069-2.
improved SVM-radial bias technique in healthcare monitoring system,” J [32] F. Hemmatian and M. K. Sohrabi, “A survey on classification techniques
Ambient Intell Humaniz Comput, vol. 12, Feb. 2021, doi: for opinion mining and sentiment analysis,” Artif Intell Rev, vol. 52, no.
10.1007/s12652-019-01652-0. 3, pp. 1495–1545, 2019, doi: 10.1007/s10462-017-9599-6.
[19] M. A. Hearst, S. T. Dumais, E. Osuna, J. Platt, and B. Scholkopf, “Support [33] Y.-W. Chang, C.-J. Hsieh, K.-W. Chang, M. Ringgaard, and C.-J. Lin,
vector machines,” IEEE Intelligent Systems and their Applications, vol. “Training and Testing Low-degree Polynomial Data Mappings via Linear
13, no. 4, pp. 18–28, 1998, doi: 10.1109/5254.708428. SVM,” 2010.
[20] D. Tomar and S. Agarwal, “A comparison on multi-class classification [34] B. Gao and L. Pavel, “On the Properties of the Softmax Function with
methods based on least squares twin support vector machine,” Knowl Application in Game Theory and Reinforcement Learning,” Apr. 2017,
Based Syst, vol. 81, pp. 131–147, 2015, doi: [Online]. Available: http://arxiv.org/abs/1704.00805
https://doi.org/10.1016/j.knosys.2015.02.009.
[35] V. Nair and G. E. Hinton, “Rectified Linear Units Improve Restricted
[21] J. Hernandez, D. López, and N. Vera, “Primary user characterization for Boltzmann Machines,” in Proceedings of the 27th International
cognitive radio wireless networks using long short-term memory,” Int J Conference on International Conference on Machine Learning, in
Distrib Sens Netw, vol. 14, p. 155014771881182, Feb. 2018, doi: ICML’10. Madison, WI, USA: Omnipress, 2010, pp. 807–814.
10.1177/1550147718811828.
[36] S. N. Abdulkader, A. Atia, and M.-S. M. Mostafa, “Brain computer
[22] S. Tortora, S. Ghidoni, C. Chisari, S. Micera, and F. Artoni, “Deep interfacing: Applications and challenges,” Egyptian Informatics Journal,
learning-based BCI for gait decoding from EEG with LSTM recurrent vol. 16, no. 2, pp. 213–230, 2015, doi:
neural network,” J Neural Eng, vol. 17, no. 4, Aug. 2020, doi: https://doi.org/10.1016/j.eij.2015.06.002.
10.1088/1741-2552/ab9842.
[37] J. H. Jeong, K. H. Shim, D. J. Kim, and S. W. Lee, “Brain-Controlled
[23] H. Almutairi, G. M. Hassan, and A. Datta, “Classification of Obstructive Robotic Arm System Based on Multi-Directional CNN-BiLSTM
Sleep Apnoea from single-lead ECG signals using convolutional neural Network Using EEG Signals,” IEEE Transactions on Neural Systems and
and Long Short Term Memory networks,” Biomed Signal Process Rehabilitation Engineering, vol. 28, no. 5, pp. 1226–1238, May 2020, doi:
Control, vol. 69, p. 102906, 2021, doi: 10.1109/TNSRE.2020.2981659.
https://doi.org/10.1016/j.bspc.2021.102906.
[38] M. A. A. Mousa, A. Elgohr, and H. Khater, “Path Planning for a 6 DoF
[24] E. Pranav, S. Kamal, C. Satheesh Chandran, and M. H. Supriya, “Facial Robotic Arm Based on Whale Optimization Algorithm and Genetic
Emotion Recognition Using Deep Convolutional Neural Network,” in Algorithm,” Journal of Engineering Research, vol. 7, no. 5, pp. 160–168,
2020 6th International Conference on Advanced Computing and Nov. 2023, doi: 10.21608/erjeng.2023.237586.1256.
Communication Systems (ICACCS), 2020, pp. 317–320. doi:
10.1109/ICACCS48705.2020.9074302. [39] M. A. A. Mousa, A. T. Elgohr, and H. A. Khater, “Trajectory
Optimization for a 6 DOF Robotic Arm Based on Reachability Time,”
[25] K. Noda, H. Arie, Y. Suga, and T. Ogata, “Multimodal integration Annals of Emerging Technologies in Computing, vol. 8, no. 1, pp. 22–35,
learning of robot behavior using deep neural networks,” Rob Auton Syst, Jan. 2024, doi: 10.33166/AETiC.2024.01.003.
vol. 62, no. 6, pp. 721–736, 2014, doi: 10.1016/j.robot.2014.03.003.
[40] M. A. Elazab, hamouda Abueldahab, A. Elgohr, and M. S. Elhadidy, “A
[26] B. Cao, H. Niu, J. Hao, and G. Wang, “Building EEG-based CAD object Comprehensive Review on Hybridization in Sustainable Desalination
selection intention discrimination model using convolutional neural Systems,” Journal of Engineering Research, vol. 7, no. 5, pp. 89–99, Nov.
network (CNN),” Advanced Engineering Informatics, vol. 52, p. 101548, 2023, doi: 10.21608/erjeng.2023.235480.1238.
2022, doi: https://doi.org/10.1016/j.aei.2022.101548.
[41] B. Zhang, J. Wang, and T. Fuhlbrigge, “A review of the commercial brain-
[27] G. A. Marcoulides, “Discovering Knowledge in Data: an Introduction to computer interface technology from perspective of industrial robotics,” in
Data Mining,” J Am Stat Assoc, vol. 100, no. 472, pp. 1465–1465, Dec. 2010 IEEE International Conference on Automation and Logistics, 2010,
2005, doi: 10.1198/jasa.2005.s61. pp. 379–384. doi: 10.1109/ICAL.2010.5585311.
[28] Daniel T. Larose, Discovering Knowledge in Data : An Introduction to [42] S. Dutta, T. Banerjee, N. D. Roy, B. Chowdhury, and A. Biswas,
Data Mining, Second edition. Wiley-Interscience, Hoboken, N.J., ©2005, “Development of a BCI-based gaming application to enhance cognitive
2017. control in psychiatric disorders,” Innov Syst Softw Eng, vol. 17, no. 2, pp.
[29] J. Bird, A. Ekart, C. Buckingham, and D. Faria, “Mental Emotional 99–107, 2021, doi: 10.1007/s11334-020-00370-7.
Sentiment Classification with an EEG-based Brain-machine Interface,” [43] S. Burwell, M. Sample, and E. Racine, “Ethical aspects of brain computer
Feb. 2019. interfaces: a scoping review,” BMC Med Ethics, vol. 18, no. 1, p. 60,
[30] R. Ahmed Hegazii, E. Abdelhalim, and H. El-Din Mostafa, “A Proposed 2017, doi: 10.1186/s12910-017-0220-y.
Technique for Breast Cancer Prediction and Classification Based on

1055 | P a g e
www.ijacsa.thesai.org
View publication stats