Indonesian Journal of Electrical Engineering and Computer Science

Vol. 34, No. 3, June 2024, pp. 2045~2054

ISSN: 2502-4752, DOI: 10.11591/ijeecs.v34.i3.pp2045-2054  2045

Enhancing EEG-based brain-computer interface systems

through efficient machine learning classification techniques

Ferdi Ahmed Yassine1,2, Ghazli Abdelkader1

1
Department of Mathematics and Computer Sciences, Faculty of Exact Sciences, Tahri Mohamed University, Bechar, Algeria
2
Laboratory of LTIT, Tahri Mohamed University, Bechar, Algeria

Article Info ABSTRACT

Article history: Advances in the fields of neuroscience and computer science have greatly
enhanced the human brain’s ability to communicate and interact with the
Received Jan 21, 2024 surrounding environment. In addition, recent steps in machine learning (ML)
Revised Feb 24, 2024 have increased the use of electroencephalography (EEG)-based BCIs for
Accepted Mar 10, 2024 artificial intelligence (AI) applications. The prevailing challenge in
recording EEG sensor data is that the captured signals are mixed with noise,
which makes their effective use difficult. Therefore, strengthening the
Keywords: classification stage becomes extremely important and plays a major role in
addressing this problem. In this study, we chose five most widely used
Attention classification models that obtained the best results in this field and tested
Brain-computer interface them on two open-source databases. We also focused on improving the
Classification hyperparameters of each algorithm to obtain best results. Our results indicate
Electroencephalography excellent results on the first dataset and acceptable for most models on the
Focus second, while RF showed superior performance on both with an accuracy of
Machine learning 100% on the first dataset and 86.47% on the second. This was achieved with
the lowest training costs, and better performance compared to previous
works we evaluated that used the same databases. These results provide
valuable insights and advance the development of brain-computer interface
(BCI) technology and design.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Ferdi Ahmed Yassine
Department of Mathematics and Computer Sciences, Faculty of Exact Sciences, Tahri Mohamed University
Kenadsa Street, Bechar, Algeria
Email: ferdi.ahmed@univ-bechar.dz

1. INTRODUCTION
A brain-computer interface (BCI) is a technology that captures and interprets an individual’s brain
signals to perform a desired action. Among the various techniques employed in BCI applications, one of the
most widely used is electroencephalography (EEG) [1], [2]. BCI provides a unique opportunity to develop
innovative forms of communication technology controlled by the brain, offering significant advantages to
individuals with motor impairments [3]. Brain-computer interfaces can be employed to create a range of
applications such as brain-controlled prosthetic limbs, adaptive chairs, speech systems, emotion detection,
states of focus and attention and more. For instance, interfacing a humanoid robot with this communication
system opens up numerous possibilities for replicating human movements, both in terms of physical
appearance and the range of motions it can achieve [4], [5]. Several methods can be used to obtain brain
signals, including electrocorticography (ECoG), near-infrared spectroscopy (NIRS) and electroencephalo-
graphy (EEG). The detection and analysis of EEG is referred to as electroencephalography where
electroencephalogram (electro=electrical, encephalo=brain, gram=record). An EEG captures the electrical
signals generated by brain cells. These signals, also known as local field potentials, are recorded using

Journal homepage: http://ijeecs.iaescore.com

2046  ISSN: 2502-4752

electrodes either placed on the scalp or inserted directly into the cortex, referred to as an electrocorticogram.
The monitoring of EEG can occur in various contexts, such as in response to stimuli (event-related potential
or ERP) or in the absence of any specific stimulus, termed spontaneous EEG [6]. EEG has been a
fundamental technique in clinical neurology for many years. Bioelectric potentials are created by the
electrochemical activity of excitable cells found in neural, muscular, or glandular tissue [7]. The first
observations of bioelectric potentials in the brains of rabbits and monkeys date back to the 1870s, thanks to
the work of English physiologist Richard Caton. Meanwhile, the human EEG was first identified in 1924 by
German psychiatrist Hans Berger, who even believed he had experienced mental telepathy with his sister
during a serious accident hundreds of kilometers away [8]. These voltages are generated by the brain’s
neuronal activity in response to various external circumstances, events, or stimuli. The examination of EEG
rhythms allows for the assessment of shifts in neural activities for clinical diagnosis. EEGs typically exhibit
frequencies ranging from 0.5 to 40 Hz and amplitudes between 10 and 200 V [9]. Five distinct EEG rhythms
have been identified: delta (0.5-4) Hz, theta (4-8) Hz, alpha (8-13) Hz, beta (13-30) Hz and gamma
(over 30 Hz), as shown in Table 1.
Nowadays, many researchers are exploring the integration of deep learning techniques in the world
of brain-computer interfaces. However, when it comes to real-world applications, using deep learning
requires complex calculations and a deeper understanding of tuning various parameters, including
architectural setup and hyperparameters [10]. Therefore, machine learning techniques have become widely
used because their results are superior to deep learning techniques. In this study, five machine learning
techniques were applied, which are considered common, most widely used, and obtain the best results in this
field. support vector machine (SVM), linear discriminant analysis (LDA), k-nearest neighbor (KNN),
decision tree (DT) and random forest (RF) were selected for use in this experiment. By comparing the
accuracy rates, these five methods are evaluated and compared, with the aim of determining the most
effective classifier. Two open-source databases are used. The first contains electroencephalography (EEG)
data to detect the state of mental attention, and the second contains brainwave data for the state of
concentration for students as shown in Figure 1. In contrast, in most current works, the accuracy is very low,
which makes this data not recommended or its use is limited. This is what causes very slow progress in this
field. The new results in this work are better than the current works in terms of the accuracy obtained, thanks
to the methods of data processing and improving the machine learning algorithms used in this study. The
remainder of the paper is structured as follows: the “Previous works” part highlights earlier research in this
field, and the “Material and methods” section describes the data and machine learning methods utilized in
this study for categorization. After then, the numerical evaluation, a thorough description of the techniques
and instruments employed in the inquiry, and the stages of the applied aspect are explained in the
“Experimental setup” section The EEG classification, discussion and conclusion are presented in the “Results
and Discussion” section. In the “Conclusion” section, we offer a summary at the conclusion.

Table 1. Classification of frequency bands

Brain wave Frequency (Hz) Amplitude Brain states
delta (∆) 0.5-4 higher deep sleep, deepest meditation
theta (Θ) 4-8 high drowsiness, dreaming, deeply relaxed
alpha (α) 8-12 medium very relaxed, alert, positive attention
beta (β) 13-35 low active, attentive, judgment, relaxed
gamma (γ) >35 lower concentration, integrated thoughts

Figure 1. Flowchart of using ML techniques to classify EEG-based mental attention and confusion situations

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 3, June 2024: 2045-2054
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  2047

2. PREVIOUS WORKS
Rani et al. [11] claim that when using the same physiological data, support vector machines (SVM)
perform the best, featuring an 85.81% classification accuracy, closely followed by a Bayesian network at
74.03%, a regression tree at 83.5%, and a k-nearest neighbor (K-NN) at 75.16%. By utilizing informative
information, Bayesian network and K-NN algorithms can perform better. When it comes to physiological
signal databases obtained from ten to hundreds of users, SVM exhibits 25% and 33.3% accuracy for three
and four emotion categories, respectively. Using Marquardt forward propagation, K-NN, and discriminant
function analysis, Nasoz et al. [12] were able to discriminate between six emotions with a classification
accuracy ranging from 71% to 83%. Conati [13] proposed that probabilistic models may be created using a
process that takes into account the user’s personality, numerous body expressions, and the setting of the
interaction. Artificial neural network (ANN) has been used to assess mental fatigue, and the average
classification accuracies for the baseline, low task difficulty, and high difficult task states, respectively, were
85%, 82%, and 86% [14]. Fisher created an emotion-recognizer based on SVMs that had accuracy rates for
three, four, and five emotion categories of 78.4%, 61.8%, and 41.7%, respectively [15]. K-NN is one of the
most popular strategies for categorizing EEG data linked to certain affective/emotional states, according to a
thorough survey conducted by Rani et al. [11]. When analyzing EEG data to identify emotion sickness, Yu et
al. [16] discovered that K-NN was the most successful classifier. K-NN is said to be very effective for
classifying EEG data by Bhattacharyya et al. because it can handle discriminant analysis of challenging
probability densities [17]. In the medical industry, RT is frequently used to classify data like EEG, say
Wilson and Russell [14]. Additionally, Brown et al. indicate that RT is frequently used to categorize EEG
data [18]. Macas et al. [19] classified many emotional states using BN successfully. Rani et al. [11] fully
endorse SVM in their study and suggest using it to correctly identify EEG data. Chen and Hou [20] also lend
credence to this assertion. Yu et al. [16] and Huang et al. [21] experiments show that SVM can categorize
EEG data effectively and with promising results. Because ANN can handle noisy data effectively, Chen and
Hou [20] suggest that it is a useful approach for classifying EEG data. These five techniques (decision tree
(DT), random forest (RF), neural network (MLP-ANN), K-nearest neighbor (KNN) and support vector
machine (SVM)) have been found to be used in most experimental experiments.

3. MATERIAL AND METHODS

3.1. Datasets
3.1.1. EEG data for mental attention state detection (Dataset 1)
An original dataset gathered in [22] consisted of 25 hours of EEG recordings from 5 participants
engaged in 34 trials. These recordings were utilized to monitor attention states (focused, unfocused, and
drowsy) through passive EEG BCI. During a low-intensity control task, participants operated a computer-
simulated train using Microsoft Train Simulator. Throughout the trials, the experiment supervisor closely
observed participants, ensuring there were no significant disturbances like movement or speech, and recorded
the sessions on video. Each participant took part in 7 trials, with a maximum of one trial per day. The initial 2
trials served as habituation, while the subsequent 5 trials were designated for data collection. EEG data was
captured utilizing a modified EMOTIV Epoc EEG headset with its classic wet electrodes. This portable EEG
acquisition device offered 12 channels of real-time EEG data at a sampling rate of 128 Hz, a voltage
resolution of 0.51 V, and a bandwidth of 0.2-43 Hz. The device was connected to a computer via a wireless
Bluetooth link. Electrode positions followed the standard 10-20 system: C3, Cz, C4, F3, Fz, F4, T3, T4, T5,
T6, and Pz. Data extraction was accomplished through a customized Matlab script developed based on the
eeglogger.m sample program. Each MATLAB file contained the data object acquired from the EMOTIV
device during a single experiment.

3.1.2. Confused student EEG brainwave data (Dataset 2)

This dataset [23] was generated through a series of exercises involving 10 university students who
watched massive open online course (MOOC) videos. The selected videos aimed to be comprehensible to
college students, focusing on subjects like basic algebra and geometry. However, intentionally confusing
content was also prepared, featuring topics such as quantum mechanics and stem cell research. In total, 20
videos were prepared, with 10 falling into each category. Each video had a duration of approximately 2
minutes, deliberately ending in the middle of a topic to heighten confusion. The students wore a
single-channel wireless MindSet device, which measured activity in their frontal lobes. This device recorded
the voltage between an electrode on the forehead and two electrodes (one acting as the ground and another as
the reference) attached to the ears. After viewing each video, students rated their confusion level on a scale
from 1 to 7. These ratings were further normalized to indicate whether the students were confused or not.
This self-labeled confusion was used in conjunction with a pre-specified confusion label. Throughout the
data collection process, each student watched ten video clips, resulting in a total of 100 data points. Although
Enhancing EEG-based brain-computer interface systems through … (Ferdi Ahmed Yassine)
2048  ISSN: 2502-4752

there are over 12,000 rows in the dataset, considering each video clip as a single data point, there are over
120 rows sampled every 0.5 seconds within each data point. Notably, EEG data was collected only during the
middle 1-minute segment of each 2-minute video, with the first and last 30 seconds removed. The average
values of the highest frequency signals were reported over each 0.5-second interval. More details shown in
Table 2.

Table 2. Description of databases

Dataset 1 Dataset 2
Type Mental states Confusion situations
Subject 5 subjects 10 students
Time 5 hours for each subject 40 min for each subject
Equipment EMOTIV EPOC + 14 channels Mindset NeuroSky
Channel 14 electrode channels 3 electrode channels
Frequency 128 Hz 512 Hz
Classes 3 classes (focused/unfocused/drowsy) 2 classes (attention/mediation)
Files types MATLAB file CSV file
Size 584 MB 144

3.2. Machine learning

3.2.1. Linear discriminant analysis
Linear discriminant analysis (LDA) is proficient in resolving both binary and multiclass
classification challenges. It employs a linear classifier to allocate variables to specific classes. Due to its
linear nature, LDA inherits stringent assumptions regarding linearity and normality [24]. It is shares
similarities with linear regression and k-means clustering, yet it also differs significantly from both methods.
Unlike k-means clustering, which is an unsupervised classification technique, LDA is supervised. This means
LDA’s classification is trained on known data, making it a supervised learning method. LDA aims to find the
best linear function that can effectively separate data points into specific categories or groups. To achieve
this, LDA maximizes the variances between group means, ensuring that the differences between the means of
different classes are as large as possible. Simultaneously, LDA minimizes the variances within each group,
making the data points within the same class as similar to each other as possible. The main goal of LDA is to
strike a balance between enhancing inter-class variation and lowering intra-class variance [25].

3.2.2. Support vector machine (SVM)

SVM is a fundamental and crucial technique for classifying various data points. It categorizes these
points, also known as support vectors, by creating a hyperplane using the kernel function. There are different
types of kernel functions, including radial, radial-integral, polynomial, and linear kernels. A hyperplane is a
plane that passes through the centers of the data points and is responsible for properly separating classes
within the given dataset. It ensures the largest margin within the area bounded by the hyperplane. The
support vectors, belonging to subgroups +1 and -1, are the closest to the dividing hyperplane and the edge of
the slab. By correctly identifying support vectors, the margin can be maximized using appropriate
methods [26], [27].

3.2.3. K-nearest neighbors (K-NN)

K-NN is a straightforward supervised machine learning algorithm used for both classification and
regression tasks. It operates by referencing a database of data points categorized into different classes and
attempts to classify a new sample data point provided to it. K-NN is considered non-parametric as it doesn’t
assume anything about the underlying data distribution. K-NN offers several advantages: it is user-friendly,
cost-effective to construct, and adaptable for classes with diverse communication patterns. It can be highly
effective in certain scenarios where other methods might fail. However, there are drawbacks to KNN as well.
Classifying unclassified records can be expensive, requiring the calculation of distances to the k-nearest
neighbors. As the method becomes more computationally intensive with larger training sets, accuracy might
decrease, especially when dealing with numerous distracting or irrelevant elements. Additionally, KNN is a
slow learner as it computes distances across k-neighbors. It retains all the training data without generalizing,
making it less efficient in handling large datasets due to costly calculations. Moreover, higher dimensional
data can lead to reduced accuracy in defining regions, making it necessary to carefully consider the choice of
features when applying the K-NN algorithm [28].

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 3, June 2024: 2045-2054
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  2049

3.2.4. Decision tree (DT)

DT are hierarchical structures used for classifying instances based on their feature values. Each node
in a decision tree represents a feature of the instance to be classified, and each branch corresponds to a
possible value the node can have. The classification process starts at the root node, where instances are sorted
based on their feature values [29]. In the context of data mining and machine learning, DT learning involves
using a decision tree as a predictive model. It maps observations about an item to conclusions about the
item’s target value. These tree models are also known as classification trees or regression trees [30]. Decision
tree classifiers often utilize post-pruning techniques to enhance their performance. These techniques involve
evaluating decision trees by using a validation set and removing nodes, assigning them the most common
class of the training instances they are associated with [29].

3.2.5. Random forest (RF)

The RF algorithm utilizes the collective strength of multiple DT to make decisions in the field of
machine learning [31]. It comprises a set of n decision trees, each generating distinct results for a given input.
In this context, the model’s output is determined by the majority of outcomes from these n decision trees.
Random forest serves as a notable illustration of ensemble learning [32]. It has the capability to address
classification and regression (CART) challenges by employing the bootstrap clustering technique, commonly
known as bagging [33].

4. EXPERIMENTAL SETUP
To showcase our work, we utilized the Python environment (version 3.11.3) on a HP EliteBook
laptop equipped with an 8th generation Intel Core i7 processor, 16 GB of RAM, and running Windows 11
(64-bit). In the initial stages of both preprocessing and feature extraction, we employed power spectral
density (PSD), a method used in signal processing and physics to describe the distribution of power over
different frequencies in a signal. It provides information about how the power of a signal is distributed across
its frequency components. PSD is particularly useful in analyzing signals that vary over time. The PSD is
typically calculated for a continuous signal or a discrete signal. For a continuous signal, the PSD S(f) is
defined as the Fourier transform of the autocorrelation function R(t) of the signal.
∞
𝑆(𝑓) = ∫−∞ 𝑅(𝑡)𝑒 −𝑗2𝜋𝑓𝑡 𝑑𝑡 (1)

Here, f represents frequency, R(t) is the autocorrelation function, and t is the time lag. For a discrete
signal, the PSD can be estimated using methods such as the periodogram, which is a tool for estimating the
spectral density of a signal. The data underwent normalization, a process aimed at reorganizing it to facilitate
the application of machine learning algorithms. Data normalization is a technique used in data preprocessing
to scale and standardize the features of a dataset. The goal of normalization is to bring the values of different
features into a similar range, preventing some features from dominating others in machine learning
algorithms that are sensitive to the scale of the input features. This normalization served the dual purpose of
eliminating repetitive and disorganized data while ensuring uniformity across all records and fields. In
addition to PSD and normalization, the SMOTE technique (synthetic minority oversampling technique) is
also an added value in this work. It is a common technique in machine learning to address class imbalance by
creating artificial samples for the minority class, and it had a major role and magical effect. In performing
algorithms with data and improving results. In Python we use the “imbalanced learning” library, commonly
referred to as “non-learning”, to implement SMOTE.
Our work was carried out on each database in a separate program, and these steps are common to
them: Initially, in our Python program, we imported data files, utilizing the MATLAB format for the first
dataset and csv format for the second. Subsequently, we amalgamated all elements from these files into a
single table. We proceeded with table level partitioning, defining inputs and outputs, and subsequently fed
the data into our machine learning algorithms. During the classification phase for the first database, the
classifier was trained to categorize values as 0 (indicating drowsiness), 1 (representing unfocused attention),
or 2 (indicating focused attention). For the second database, the classifier discerned values as 0 (Attention
<=50) or 1 (Attention>50). In all experiments, we split the datasets into training sets (80%) and test sets
(20%), Table 3 shows the number of cases used for training and testing. The classification process involved a
range of classifiers, including LDA, SVM, KNN, DT and RF. Unlike many studies that rely on default
algorithm parameters, we fine-tuned and modified these settings multiple times to enhance the performance
of each algorithm. Finally, the selection of the most suitable machine learning algorithm was guided by the
accuracy measure, with the algorithm demonstrating the highest accuracy being chosen. After selection, we
have a ready-made model with an optimal algorithm. Once we’ve selected and finetuned the optimal model,

Enhancing EEG-based brain-computer interface systems through … (Ferdi Ahmed Yassine)

2050  ISSN: 2502-4752

we’ve taken the extra step to export it into Python (.py) format for computer usage. Additionally, we’ve
saved the model in widely used formats such as joblib (.sav) and pickle (.pkl), making it ready for integration
into any Android application. This step is driven by the fact that mobile phones are the most ubiquitous
communication devices in history [34], and mobile networks enjoy global coverage and are currently the
most widely used network type [35]. Moreover, the vast majority of people now possess a mobile phone
equipped with internet or network connectivity [36]. By taking this step, we’ve ensured that our work is
highly versatile and can be utilized on a wide range of devices.

Table 3. Distribution of datasets

Instances Value
Data 1 Training 16,236 (80%)
Testing 4,060 (20%)
Data 2 Training 598 (80%)
Testing 200 (20%)

5. RESULTS AND DISCUSSION

5.1. Accuracy
In our results evaluation, we utilized the accuracy metric, a fundamental measure for evaluating
classification models. Accuracy represents the proportion of correct predictions made by the model.
Formally, accuracy is calculated (2).

(𝑇𝑜𝑡𝑎𝑙 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠)

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (2)
(𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐶𝑜𝑟𝑟𝑒𝑐𝑡 𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠)

In the context of classification, accuracy can also be expressed in terms of true positives (TP), true
negatives (TN), false positives (FP) and false negatives (FN) as (3).
(𝑇𝑁+𝑇𝑃)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (3)
(𝑇𝑃+𝐹𝑃+𝑇𝑁+𝐹𝑁)

Where, TP is instances where the model correctly predicts the positive class, TN is instances where the
model correctly predicts the negative class, FP is instances where the model incorrectly predicts the positive
class and FN is instances where the model incorrectly predicts the negative class. Accuracy provides a clear
and intuitive measure of the model’s overall correctness in its predictions. The classification results for all
classifiers for each of the two databases used are shown in Table 4.
For the first, the highest accuracy was obtained for RF and DT (100%), followed by KNN (95%),
then LDA (78%), then SVM (75.5%). As for the second database, the highest accuracy was obtained using
RF (86.47%), followed by SVM (65.09%), then KNN (73.37%), then LDA (65.71%), then DT (78.07%).
Figure 2 represents a flow chart for using the machine learning techniques used in this study and comparing
the percentage of accuracy between them, so that Figure 2(a) represents an explanation of the results of
mental states data, and Figure 2(b) represents confusion situations.

Table 4. Machine learning models’ accuracies are typically expressed as percentages

SVM DT RF KNN LDA
DATA 1 75.5% 100% 100% 95% 78 %
DATA 2 65.09% 78.07% 86.47% 73.37% 65.71%

5.2. Confusion matrix

The confusion matrix, a summary of the machine learning model’s performance on the test data
used to predict category scores for input instances, serves as an additional tool to evaluate the effectiveness of
our classification model. This matrix displays the proportions of true positives (TP), false positives (FP), true
negatives (TN) and false negatives (FN) generated by the model in the test data. It helps understand the
unpredictability of classification model predictions, enabling us to identify the types of errors that have been
made. The confusion matrix results for the top algorithms that achieved the best accuracy for both datasets
are shown in Figure 3. Whereas the results for the mental states data are shown in Figure 3(a), Figure 3(b),
and the result for the confusion states data are shown in Figure 3(c). When discussing the accuracy results,
we find that the RF classifier is the best for both databases, as it achieved 100% accuracy for the first

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 3, June 2024: 2045-2054
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  2051

database and 86.47% accuracy for the second database. RF is an ensemble learning method that combines the
predictions of multiple decision trees, which often produces robust and accurate results, especially for
complex data sets. For the first database, both the RF and DT classifiers achieved perfect accuracy (100%),
indicating that these models were able to accurately classify all data points in the dataset. KNN achieved an
accuracy of 95%, which indicates that it performed well but may have encountered some difficulties in some
cases where the nearest neighbors were not representative of the class. LDA and SVM achieved lower
accuracy of 78% and 75.5%, respectively. These results indicate that the data in the first database may not
have clear linear separation, making it difficult for linear classifiers such as LDA and SVM to perform the
same as tree-based methods. For the second database, RF still performs the best with an accuracy of 86.47%,
indicating its robustness across different datasets. SVM, which is known for handling complex decision
boundaries, achieved an accuracy of 65.09%. KNN and LDA achieved an accuracy of 73.37% and 65.71%,
respectively. The decision tree was the least accurate at 78.07%.

(a) (b)

Figure 2. Flowchart of using ML techniques for (a) mental states and (b) confusion situations

(a) (b) (c)

Figure 3. Confusion matrices of the best classification algorithms obtaining the (a) best accuracy, (b) mental
states database, and (c) confusion situations database

The lower accuracy of all classifiers in the second database compared to the first database may
indicate that the second dataset is more challenging, perhaps due to higher dimensionality, noise, or class
imbalance, which may affect the performance of machine learning models. After RF achieved excellent
results, it has proven that it is the ideal choice for classification problems of all types (multiple and binary)
and that it is highly efficient when it comes to this type of data, which is represented by EEG signals and
their various types. Therefore, it was chosen as an ideal classifier that can be relied upon in the classification
stage through the BCI to determine states of attention (focused, unfocused, and sleepy) and to determine
whether confusion exists or not, for the first database and the second database, respectively. This study
investigated the effects of improving the performance of classifiers at the classification stage in EEG-based
BCI systems. While previous studies have used machine learning classifiers, they have not exploited
classifiers effectively, paid appropriate attention to selecting the optimal classifier, nor have they shown
widespread interest in improving the performance of the models they have used. We found that improving
Enhancing EEG-based brain-computer interface systems through … (Ferdi Ahmed Yassine)
2052  ISSN: 2502-4752

classification results is related to working on pre-normalizing the data and improving the hyperparameters of
the machine learning algorithms. The method proposed in this study tends to have an unusually high
percentage of “accuracy” similar to that achieved by the RF algorithm and DT. Our study indicates that high
performance accuracy of BCI systems is not only associated with poor performance in the stages that precede
classification. The proposed method may benefit from the characteristics of the original data without
negatively affecting its quality or reducing it is values, in contrast to some previous research. This study
explored a comprehensive optimization method for brain-computer interface systems based on EEG with the
use of efficient machine learning classification algorithms. Despite the positive findings of this paper, more
in-depth studies may be needed to ensure that the research steps have positive outcomes for all, or at least
most, work.
Our results update this data and make it recommended and of unlimited use. This also gives
effectiveness to BCI systems and the possibility of rapid progress in research. Our outcomes encourage
others to build on our findings, as the work steps in this study can be exploited on similar EEG data for the
purpose of improving performance. Recent observations indicate that focusing on data preparation, such as
normalizing it, prior imbalance, and optimizing the hyperparameters of machine learning algorithms, has a
very significant impact in improving the results. In comparison with published works that relied on the same
databases used in our study, as shown in the Table 5, our obtained results achieved better results, proving the
effectiveness and superiority of our model, and this is what we sought in this study. Our final results provide
conclusive evidence that this phenomenon is related to the change that occurs before and during the use of
machine learning tools, and not due to improvement being limited to only one stage.

Table 5. Comparison with research work using the same datasets

Work Classification method Best accuracy
[22] KNN, ANFIS, SVM 91.72
[37] KNN 97.5
[38] CNN 96.40%
[39] SVM, KNN, QDA 95.39%
Data1 [40] XGBoost, RF, KNN 98%
[41] RF, KNN, SVM 96%
[42] RF, SVM, XGBoost, Neural Networks 99.9%
[43] Optimizable Ensemble 97.8%
Our contribution SVM, DT, RF, KNN, LDA 100%
[23] Specific classifiers, independent classifier 67%
[44] LSTM 73.3%
[45] KNN 73.33%
[46] AlexNet, Custom CNN with Dropout 65%
Data 2 [47] SVM 59.1%
[48] RF, XGBoost, LightGBM, Catboost 64.75%
[49] GTN, RNN, GCN 53.67%
[50] fMGTN, GRU, TTNN, GCN 56.10%
Our contribution SVM, DT, RF, KNN, LDA 86.47%

6. CONCLUSION
In this study, an important stage was worked on, which is classification, using two databases of
different types, original, open source and available to everyone, with different classification (binary and
multi-section). In our work, we relied on the latest and best machine learning algorithms most used in this
field. In our work, we were keen to improve the performance of each algorithm by changing and modifying
the input data for each of them several times until we reached the best. We then evaluated the effectiveness of
the developed classifiers by measuring the accuracy percentage and then selecting the best one and
displaying it in the confusion matrix. DT, RF, LDA, KNN, and SVM are five classifiers used in this work to
classify our data. The RF classifier achieved the best results on both databases, with an accuracy of 100% on
the first, and more than 86 % on the second. This makes it recommended as a suitable, effective, and
ready-to-use classifier for researchers interested in working on the same databases used in this study. As an
idea for subsequent work, we can rely on the idea of amplifying this same data to test the efficiency of deep
learning techniques on it, and then modifying the inputs of these algorithms in order to improve them as well.

REFERENCES
[1] S. Mantri, V. Dukare, S. Yeole, D. Patil, and V. M. Wadhai, “A survey : fundamental of EEG,” International Journal of Advance
Research in Computer Science and Management Studies, vol. 1, no. 4, pp. 83–89, 2013.
[2] G. Pfurtscheller et al., “Current trends in Graz brain-computer interface (BCI) research,” IEEE Transactions on Rehabilitation
Engineering, vol. 8, no. 2, pp. 216–219, Jun. 2000, doi: 10.1109/86.847821.

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 3, June 2024: 2045-2054
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  2053

[3] L. F. Nicolas-Alonso and J. Gomez-Gil, “Brain computer interfaces, a review,” Sensors, vol. 12, no. 2, pp. 1211–1279, Jan. 2012,
doi: 10.3390/s120201211.
[4] K. Hirai, M. Hirose, Y. Haikawa, and T. Takenaka, “The development of Honda humanoid robot,” in Proceedings. 1998 IEEE
International Conference on Robotics and Automation (Cat. No.98CH36146), IEEE, pp. 1321–1326. doi:
10.1109/ROBOT.1998.677288.
[5] R. A. Brooks, C. Breazeal, M. Marjanović, B. Scassellati, and M. M. Williamson, “The cog project: building a humanoid robot,”
in International Workshop on Computation for Metaphors, Analogy, and Agents, 1999, pp. 52–87. doi: 10.1007/3-540-48834-0_5.
[6] K. Blinowska and P. Durka, “Electroencephalography (EEG),” in Wiley Encyclopedia of Biomedical Engineering, Wiley, 2006.
doi: 10.1002/9780471740360.ebs0418.
[7] E. Başar, “Brain oscillations in neuropsychiatric disease,” Dialogues in Clinical Neuroscience, vol. 15, no. 3, pp. 291–300, Sep.
2013, doi: 10.31887/DCNS.2013.15.3/ebasar.
[8] D. Millett, “Hans Berger: from psychic energy to the EEG,” Perspectives in Biology and Medicine, vol. 44, no. 4, pp. 522–542,
Sep. 2001, doi: 10.1353/pbm.2001.0070.
[9] M. Abo-Zahhad, S. M. Ahmed, and S. N. Abbas, “A new EEG Acquisition protocol for biometric identification using eye
blinking signals,” International Journal of Intelligent Systems and Applications, vol. 7, no. 6, pp. 48–54, May 2015,
doi: 10.5815/ijisa.2015.06.05.
[10] R. Vinayakumar, M. Alazab, K. P. Soman, P. Poornachandran, A. Al-Nemrat, and S. Venkatraman, “Deep learning approach for
intelligent intrusion detection system,” IEEE Access, vol. 7, pp. 41525–41550, 2019, doi: 10.1109/ACCESS.2019.2895334.
[11] P. Rani, C. Liu, N. Sarkar, and E. Vanman, “An empirical study of machine learning techniques for affect recognition in human–
robot interaction,” Pattern Analysis and Applications, vol. 9, no. 1, pp. 58–69, May 2006, doi: 10.1007/s10044-006-0025-y.
[12] F. Nasoz, K. Alvarez, C. L. Lisetti, and N. Finkelstein, “Emotion recognition from physiological signals using wireless sensors for
presence technologies,” Cognition, Technology and Work, vol. 6, no. 1, pp. 4–14, Feb. 2004, doi: 10.1007/s10111-003-0143-x.
[13] C. Conati, “Probabilistic assessment of user’s emotions in educational games,” Applied Artificial Intelligence, vol. 16, no. 7–8,
pp. 555–575, Aug. 2002, doi: 10.1080/08839510290030390.
[14] G. F. Wilson and C. A. Russell, “Real-time assessment of mental workload using psychophysiological measures and artificial
neural networks,” Human Factors: The Journal of the Human Factors and Ergonomics Society, vol. 45, no. 4, pp. 635–644, Dec.
2003, doi: 10.1518/hfes.45.4.635.27088.
[15] R. A. Fisher, “The use of multiple measurements in taxonomic problems,” Annals of Eugenics, vol. 7, no. 2, pp. 179–188, Sep.
1936, doi: 10.1111/j.1469-1809.1936.tb02137.x.
[16] Y.-H. Yu, P.-C. Lai, L.-W. Ko, C.-H. Chuang, B.-C. Kuo, and C.-T. Lin, “An EEG-based classification system of Passenger’s
motion sickness level by using feature extraction/selection technologies,” in The 2010 International Joint Conference on Neural
Networks (IJCNN), IEEE, Jul. 2010, pp. 1–6. doi: 10.1109/IJCNN.2010.5596739.
[17] A. Bhattacharyya, R. K. Tripathy, L. Garg, and R. B. Pachori, “A novel multivariate-multiscale approach for computing eeg
spectral and temporal complexity for human emotion recognition,” IEEE Sensors Journal, vol. 21, no. 3, pp. 3579–3591, Feb.
2021, doi: 10.1109/JSEN.2020.3027181.
[18] L. E. Brown, I. Tsamardinos, and C. F. Aliferis, “A novel algorithm for scalable and accurate Bayesian network learning,” Studies
in Health Technology and Informatics, vol. 107, pp. 711–715, 2004, doi: 10.3233/978-1-60750-949-3-711.
[19] M. Macas, M. Vavrecka, V. Gerla, and L. Lhotska, “Classification of the emotional states based on the EEG signal processing,” in
2009 9th International Conference on Information Technology and Applications in Biomedicine, IEEE, Nov. 2009, pp. 1–4.
doi: 10.1109/ITAB.2009.5394429.
[20] G. Chen and R. Hou, “A new machine double-layer learning method and its application in non-linear time series forecasting,” in 2007
International Conference on Mechatronics and Automation, IEEE, Aug. 2007, pp. 795–799. doi: 10.1109/ICMA.2007.4303646.
[21] W. Y. Huang, X. Q. Shen, and Q. Wu, “Classify the number of EEG current sources using support vector machines,” Proceedings
of 2002 International Conference on Machine Learning and Cybernetics, vol. 4, 2002, doi: 10.1109/icmlc.2002.1175348.
[22] Ç. İ. Acı, M. Kaya, and Y. Mishchenko, “Distinguishing mental attention states of humans via an EEG-based passive BCI using
machine learning methods,” Expert Systems with Applications, vol. 134, pp. 153–166, Nov. 2019, doi: 10.1016/j.eswa.2019.05.057.
[23] H. Wang, Y. Li, X. Hu, Y. Yang, Z. Meng, and K. M. Chang, “Using EEG to improve massive open online courses feedback
interaction,” CEUR Workshop Proceedings, vol. 1009, pp. 59–66, 2013.
[24] T. C. Nokeri, “Dimension reduction and multivariate analysis using linear discriminant analysis,” in Data Science Revealed,
Berkeley, CA: Apress, 2021, pp. 117–128. doi: 10.1007/978-1-4842-6870-4_6.
[25] H. Zhou, “Linear discriminant analysis,” in Learn Data Mining Through Excel, Berkeley, CA: Apress, 2023, pp. 53–70.
doi: 10.1007/978-1-4842-9771-1_4.
[26] D. Garrett, D. A. Peterson, C. W. Anderson, and M. H. Thaut, “Comparison of linear, nonlinear, and feature selection methods for
EEG signal classification,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 11, no. 2, pp. 141–144,
Jun. 2003, doi: 10.1109/TNSRE.2003.814441.
[27] M. R. N. Kousarrizi, A. A. Ghanbari, M. Teshnehlab, M. A. Shorehdeli, and A. Gharaviri, “Feature extraction and classification
of EEG signals using wavelet transform, SVM and artificial neural networks for brain computer interfaces,” in 2009 International
Joint Conference on Bioinformatics, Systems Biology and Intelligent Computing, IEEE, 2009, pp. 352–355.
doi: 10.1109/IJCBS.2009.100.
[28] S. Ray, “A quick review of machine learning algorithms,” in 2019 International Conference on Machine Learning, Big Data,
Cloud and Parallel Computing (COMITCon), IEEE, Feb. 2019, pp. 35–39. doi: 10.1109/COMITCon.2019.8862451.
[29] S. B. Kotsiantis, “Supervised machine learning: A review of classification techniques,” Informatica (Ljubljana), vol. 31, no. 3, pp.
249–268, 2007.
[30] T. Hastie, R. Tibshirani, and J. Friedman, The elements of statistical learning, vol. 173, no. 3. in Springer Series in Statistics, vol.
173. New York, NY: Springer New York, 2009. doi: 10.1007/978-0-387-84858-7.
[31] M. I. Prasetiyowati, N. U. Maulidevi, and K. Surendro, “Determining threshold value on information gain feature selection to
increase speed and prediction accuracy of random forest,” Journal of Big Data, vol. 8, no. 1, p. 84, Dec. 2021,
doi: 10.1186/s40537-021-00472-4.
[32] M. G. HR, A. MV, G. P. S, and V. S, “Development of anti-phishing browser based on random forest and rule of extraction
framework,” Cybersecurity, vol. 3, no. 1, Dec. 2020, doi: 10.1186/s42400-020-00059-1.
[33] T. A. Rather, S. Kumar, and J. A. Khan, “Multi-scale habitat selection and impacts of climate change on the distribution of four
sympatric meso-carnivores using random forest algorithm,” Ecological Processes, vol. 9, no. 1, Dec. 2020, doi: 10.1186/s13717-
020-00265-2.
[34] G. Abdelkader, A. Alipacha, and N. H. Said, “Evolutionary approach to secure mobile telecommunication networks,” Indonesian
Journal of Electrical Engineering and Computer Science, vol. 24, no. 1, Oct. 2021, doi: 10.11591/ijeecs.v24.i1.pp357-366.
Enhancing EEG-based brain-computer interface systems through … (Ferdi Ahmed Yassine)
2054  ISSN: 2502-4752

[35] G. Abdelkader, A. Adda, and H. S. Naima, “Chao Athentication and ciphering approach to secure mobile networks,”
International Journal of Networks and Communications, vol. 10, no. 1, pp. 20–32, 2020, doi: 10.5923.j.ijnc.20201001.03.
[36] G. Abdelkader, H. S. Naima, and A. P. Adda, “Secure authentication approach based new mobility management schemes for mobile
communication,” Journal of Information Processing Systems, vol. 13, no. 1, pp. 152–173, 2017, doi: 10.3745/JIPS.03.0064.
[37] R. S. Kumar, K. K. Srinivas, A. Peddi, and P. A. H. Vardhini, “Artificial intelligence based human attention detection through brain
computer interface for health care monitoring,” in 2021 IEEE International Conference on Biomedical Engineering, Computer and
Information Technology for Health (BECITHCON), IEEE, Dec. 2021, pp. 42–45. doi: 10.1109/BECITHCON54710.2021.9893646.
[38] D. Zhang, D. Cao, and H. Chen, “Deep learning decoding of mental state in non-invasive brain computer interface,” in
Proceedings of the International Conference on Artificial Intelligence, Information Processing and Cloud Computing, New York,
NY, USA: ACM, Dec. 2019, pp. 1–5. doi: 10.1145/3371425.3371441.
[39] E. Akman Aydin, “EEG sinyalleri kullanılarak zihinsel iş yükü seviyelerinin sınıflandırılması,” Politeknik Dergisi, vol. 24, no. 2,
pp. 681–689, Jun. 2021, doi: 10.2339/politeknik.794655.
[40] K. Suwida, S. C. Hidayati, and R. Sarno, “Application of machine learning algorithm for mental state attention classification
based on electroencephalogram signals,” in 2023 International Conference on Computer Science, Information Technology and
Engineering (ICCoSITE), IEEE, Feb. 2023, pp. 354–358. doi: 10.1109/ICCoSITE57641.2023.10127825.
[41] A. Al-Nafjan and M. Aldayel, “Predict students’ attention in online learning using EEG data,” Sustainability, vol. 14, no. 11, May
2022, doi: 10.3390/su14116553.
[42] Y. Wang, R. Nahon, E. Tartaglione, P. Mozharovskyi, and V.-T. Nguyen, “Optimized preprocessing and Tiny ML for Attention
State Classification,” in 2023 IEEE Statistical Signal Processing Workshop (SSP), IEEE, Jul. 2023, pp. 695–699.
doi: 10.1109/SSP53291.2023.10207930.
[43] S. K. Khare, V. Bajaj, N. B. Gaikwad, and G. R. Sinha, “Ensemble wavelet decomposition-based detection of mental states using
electroencephalography signals,” Sensors, vol. 23, no. 18, p. 7860, Sep. 2023, doi: 10.3390/s23187860.
[44] Z. Ni, A. C. Yuksel, X. Ni, M. I. Mandel, and L. Xie, “Confused or not confused?,” in Proceedings of the 8th ACM International
Conference on Bioinformatics, Computational Biology,and Health Informatics, New York, NY, USA: ACM, Aug. 2017,
pp. 241–246. doi: 10.1145/3107411.3107513.
[45] H. Jamielatuththooah, “Klasifikasi video pembelajaran daring yang membingungkan siswa dengan algoritma k-star nearest
neighbor,” MATHunesa: Jurnal Ilmiah Matematika, vol. 9, no. 1, pp. 22–26, Jan. 2021, doi: 10.26740/mathunesa.v9n1.p22-26.
[46] I. P. A. E. D. Udayana, M. Sudarma, and N. W. S. Ariyani, “Detecting excessive daytime sleepiness with CNN and commercial grade
EEG,” Lontar Komputer : Jurnal Ilmiah Teknologi Informasi, vol. 12, no. 3, Nov. 2021, doi: 10.24843/LKJITI.2021.v12.i03.p06.
[47] D. Ganepola, M. Maduranga, and I. Karunaratne, “Comparison of machine learning optimization techniques for EEG-based
confusion emotion recognition,” in 2023 IEEE 17th International Conference on Industrial and Information Systems (ICIIS),
IEEE, Aug. 2023, pp. 341–346. doi: 10.1109/ICIIS58898.2023.10253515.
[48] M. J. Cheon, D. H. Lee, J. W. Park, H. J. Choi, J. S. Lee, and O. Lee, “CTGAN VS TGAN? Which one is more suitable for
generating synthetic EEG data,” Journal of Theoretical and Applied Information Technology, vol. 99, no. 10, 2021.
[49] Y. L. Xu, K. Konstantinidis, and D. P. Mandic, “Graph tensor networks: An intuitive framework for designing large-scale neural
learning systems on multiple domains,” ArXiv (Preprint), Mar. 2023.
[50] Y. L. Xu, K. Konstantinidis, and D. P. Mandic, “Tensor networks for multi-modal non-euclidean data,” ArXiv (Preprint), Mar.
2021.

BIOGRAPHIES OF AUTHORS

Ferdi Ahmed Yassine he is a Ph.D. student in computer science, specializing in

artificial intelligence and it is applications, at Tahri Mohammed University (UTMB), Algeria.
He obtained a Master’s Diploma in Engineering in Computer Science, specializing in artificial
intelligence and decision making, from Tahri Mohammed University in Algeria in 2019. His
research interests focus on the field of artificial intelligence, such as machine learning, deep
learning, brain computer interface and data analysis. He can be contacted at email:
ferdi.ahmed@univ-bechar.dz.

Ghazli Abdelkader he is a lecturer at Tahri Mohammed University in Bechar,

Algeria. He obtained an engineering diploma in computer science from University of Science
and Technology in Oran (USTO), Algeria in 2005. He obtained a teaching diploma in computer
science from University of Science and Technology in Oran (USTO), Algeria in 2009. His
research interests are artificial intelligence, mobile networks, cyber security, brain computer
interface, and embedded systems. He can be contacted at email: ghazek@gmail.com.

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 3, June 2024: 2045-2054