![TELKOMNIKA Telecommunication Computing Electronics and Control
Vol. 21, No. 5, October 2023, pp. 1068~1075
ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v21i5.24839 1068
Journal homepage: http://telkomnika.uad.ac.id
Electroencephalography-based brain-computer interface using
neural networks
Pham Van Huu Thien1,2
, Nguyen Ngoc Son2
1
Department of Electronics, Faculty of Electrical and Electronics, Quang Trung Vocational School, Ho Chi Minh City, Viet Nam
2
Department of Intelligent Systems, Faculty of Electronics Technology, Industrial University of Ho Chi Minh City, Ho Chi Minh City,
Viet Nam
Article Info ABSTRACT
Article history:
Received Nov 26, 2022
Revised Feb 27, 2023
Accepted Mar 07, 2023
This study aimed to develop a brain-computer interface that can control an
electric wheelchair using electroencephalography (EEG) signals. First, we used
the Mind Wave Mobile 2 device to capture raw EEG signals from the surface of
the scalp. The signals were transformed into the frequency domain using fast
Fourier transform (FFT) and filtered to monitor changes in attention and
relaxation. Next, we performed time and frequency domain analyses to
identify features for five eye gestures: opened, closed, blink per second,
double blink, and lookup. The base state was the opened-eyes gesture, and we
compared the features of the remaining four action gestures to the base state to
identify potential gestures. We then built a multilayer neural network to
classify these features into five signals that control the wheelchair’s
movement. Finally, we designed an experimental wheelchair system to test the
effectiveness of the proposed approach. The results demonstrate that the EEG
classification was highly accurate and computationally efficient. Moreover,
the average performance of the brain-controlled wheelchair system was over
75% across different individuals, which suggests the feasibility of this
approach.
Keywords:
Brain-controller wheelchair
Electroencephalography
Fast fourier transform
Neural networks
This is an open access article under the CC BY-SA license.
Corresponding Author:
Nguyen Ngoc Son
Department of Intelligent Systems, Faculty of Electronics Technology
Industrial University of Ho Chi Minh City, 12 Nguyen Van Bao, Go Vap District
Ho Chi Minh City 700000, Viet Nam
Email: nguyenngocson@iuh.edu.vn
1. INTRODUCTION
Brain-computer interfaces (BCI) are devices that acquire and analyze brain activities, then convert
them into output signals to control desired actions. BCI technology has found applications in several fields,
including rehabilitation robotics [1], [2], driver fatigue monitoring [3], [4], gaming [5], and cognitive
biometrics [6], [7]. Several techniques can be used to measure brain signals, such as electroencephalography
(EEG) [8], magnetoencephalography (MEG) [9], and functional magnetic resonance imaging (FMRI) [10].
Among these methods, EEG is widely used in many applications because it is cost-effective, easy to use with
minimal training, and can be used for mobile testing. EEG measures the electrical activity of the brain
through electrodes placed on the scalp and is divided into six main frequency bands: delta (0.5−4 Hz), theta
(4−8 Hz), alpha (8−13 Hz), beta (13−30 Hz), gamma (30−50 Hz), and high gamma (80−150 Hz). Two key
challenges in improving the performance of EEG classification are how to use spectral analysis techniques to
extract useful features from EEG signals and how to design a robust classifier.
With spectral analysis of the EEG signals, there are various techniques used. For example,
Adeli et al. [11] proposed the wavelet transform to analyze and characterize the EEG signal of the 3-Hz spike](https://image.slidesharecdn.com/1524839-230809011845-263c567b/85/Electroencephalography-based-brain-computer-interface-using-neural-networks-1-320.jpg)
![TELKOMNIKA Telecommun Comput El Control
Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien)
1069
and slow-wave epileptic discharges. Li et al. [12] proposed a quantum wavelet packet transformation
(QWPT) module to extract the wavelet packet energy entropy feature of the multi-channel and multi-sample
EEG signal. Rahman et al. [13] introduced a hybrid principal component analysis (PCA) and t-statistical
approach for feature extraction of emotion from multichannel EEG signals. Bajaj et al. [14] proposed a
tunable Q-factor wavelet transform (TQWT) to extract the non-stationary characteristic of the EEG signal for
identifying a neurological change in alertness and drowsiness states. A short-time Fourier transform (STFT)
was used to analyze EEG signals to obtain time-frequency representations [15], [16]. Li and Chen [17]
applied a fast Fourier transform (FFT) to generate the EEG matrix and PCA neural network to learn the
hidden information from the frequency matrix of EEG signals. FFT was also used to extract features of stroke
patients’ EEG signals [18]. The choice of the appropriate spectral analysis technique depends on the
complexity and characteristics of the EEG signals to be analyzed. FFT is a suitable method due to its time-shift
invariance in both time and frequency domains and its computational efficiency.
To design a powerful classifier, machine learning (ML) techniques have been applied for EEG
classification. For example, Sharma et al. [19] showed the performance of the multi-layered perceptron
(MLP) model classified the EEG signals with 90% accuracy and half the classification time compared to
traditional ML-based models (i.e., support vector machines, k-nearest neighbors, random forest, logistic
regression, and Bayes). In [15], [20]–[23] introduced a convolutional neural network (CNN) for EEG
classification. In addition, many other ML techniques were also used for classification such as extreme
learning machine (ELM) [3], [24], ensemble support vector learning (ESVL) [25], quantum machine learning
(QML) [12], long short-term memory (LSTM) and variants [26], [27], recurrent neural network (RNN) [28],
you only look once (YOLO) algorithm [29]. These classifiers can be trained on either raw EEG signals or
meaningful features extracted from pre-processed signals in the time and frequency domains. However,
a dataset is a set of images or series, so the computational cost is quite large. By selecting appropriate
features and classifiers, the classification performance of EEG signals can be greatly improved.
In this study, we use Mind Wave Mobile 2 to measure raw EEG signals from the surface of the scalp.
The signals are then transformed into the frequency domain using FFT. To extract meaningful features for
classification, we analyze the signals in both the time and frequency domains and obtain 9 features including
amplitude, peak frequency, phase, period, and root mean square for 5 action gestures: opened-eyes, closed-eyes,
blink per second, double blink, and lookup. Before analysis, noise in the appropriate frequency bands is removed
through filtering. Subsequently, these features are fed into MLP neural networks for classification of 5 signals to
control wheelchair movement: forward, backward, turn left, turn right, and stop. Finally, we test the experimental
wheelchair system on 5 different individuals to evaluate the effectiveness of our proposed approach.
The rest of this paper is structured as: in section 2, we present our methodology for feature extraction
and classification using neural networks. Section 3 provides an in-depth analysis of our EEG classification
results, along with the brain-controlled wheelchair. Finally, in section 4, we draw our conclusions based on the
findings of this study.
2. METHOD
2.1. EEG-based brain-controlled wheelchair approach
The diagram in Figure 1 illustrates the process of measuring, extracting, analyzing, and classifying
EEG signals. It includes components such as data acquisition, feature extraction, classifier-based MLP neural
networks, and a wheelchair system. In which, the MindWave Mobile 2 is used to measure EEG signals from
the surface of the scalp, and to transmit power spectrum data to a computer via Bluetooth. The analysis is
focused on 5 action gestures: opened-eyes, closed-eyes, blink per second, double-blink, and gaze direction
(look-up). The features of the EEG signals are analyzed in both the time and frequency domains and are then
used as inputs for the neural network classifier. The output of the classifier is sent as control signals through
RF networks to operate a wheelchair.
2.2. EEG feature extraction
After collecting the EEG data, it was analyzed and filtered to monitor changes in attention and
relaxation. The frequency domain was transformed using FFT, and a set of potential features was extracted,
which included 4 time-domain and 5 frequency-domain features. The time-domain EEG feature extraction
comprised the maximum, minimum, mean, and root mean square (RMS) of frequency components less than
30 Hz. In the frequency domain, the EEG feature extraction involved the mean and maximum of the alpha
band within the range of 8−13 Hz, the mean and maximum frequencies of the delta and theta bands in the
range of 1−7 Hz, and the peak frequency. These 9 features were then inputted into the MLP neural network,
which classified them into 5 control signals for the wheelchair.](https://image.slidesharecdn.com/1524839-230809011845-263c567b/85/Electroencephalography-based-brain-computer-interface-using-neural-networks-2-320.jpg)
![ ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 5, October 2023: 1068-1075
1070
Figure 1. Block diagram of EEG-based brain-controlled wheelchair
2.3. EEG classification using MLP neural model
This section employs MLP neural networks for EEG classification. The architecture of the MLP
neural classifier includes the input layer 𝑥(𝑘) = [𝑥1(𝑘), 𝑥2(𝑘) , . . . , 𝑥9(𝑘)] represents the 9 features of the
EEG signal. The output layer 𝑦(𝑘) = [𝑦1(𝑘), 𝑦2(𝑘) , . . . , 𝑦5(𝑘)] consists of the 5 output signals of the
classified neural network, namely opened-eyes, closed-eyes, blinks per second, double blink, and look-up,
respectively. 𝑣𝑗𝑛 denotes the weighting value of the input layer, 𝑤𝑚𝑗 represents the weighting value of the
hidden layer. 𝑓𝑗(. ) denotes a sigmoid function at the hidden layer, 𝐹(. ) represents a linear activation
function at the output layer, respectively. We determine the output neural model as:
𝑦𝑖(𝑣, 𝑤) = 𝐹𝑖(∑ 𝑤𝑖𝑗𝑓𝑗(∑ 𝑣𝑗𝑙𝑥𝑙
𝑚
𝑙=1 )
𝑞
𝑗=1 ) (1)
In the training process, both the input vector 𝑥 and the output vector 𝑦𝑟𝑒𝑓 are known and the
synaptic weights (𝑣, 𝑤) are adapted to obtain appropriate functional mappings from the input 𝑥 to the output
𝑦𝑟𝑒𝑓. Generally, the adaptation process can be carried out by minimizing the network error function which is
based on a measure of closeness in terms of a mean sum of square error (MSSE) criterion:
𝑀𝑆𝑆𝐸((𝑣, 𝑤) , 𝑍𝑁) =
1
2𝑁
∑ [𝑦𝑟𝑒𝑓(𝑘) − 𝑦(𝑘|(𝑣, 𝑤))]
𝑇
[𝑦𝑟𝑒𝑓(𝑘) − 𝑦(𝑘|(𝑣, 𝑤))]
𝑁
𝑡=1 (2)
Where, the training data 𝑍𝑁
is specified by 𝑍𝑁
= {[𝑥(𝑘), 𝑦𝑟𝑒𝑓(𝑘)]|𝑘 = 1, . . . , 𝑁}.
3. RESULTS AND DISCUSSION
3.1. Extracted features dataset
The data was collected by measuring each person’s actions for 30 minutes, corresponding to five
different gestures: opening their eyes, closing their eyes, blinking per second, double blinking, and looking up.
EEG raw data was collected and pre-processed to remove electrical signal noise at 30 Hz. The detailed
analysis of one participant’s features is as: the base state is the opened-eyes gesture, and the features of the
remaining four action gestures were compared to the base state to identify potential gestures. Figure 2
displays both the EEG raw signal in the time domain and the EEG signal in the frequency domain after FFT.
Figure 2. EEG signal pattern when opened-eyes
After analyzing numerous signal samples, it was found that the “eyes-closed” features differ from the
eyes-opened features in the alpha frequency range (8−13 Hz), particularly in the frequency range of 9−11 Hz,
as depicted in Figure 3(a) and Figure 3(b). Table 1 shows the extracted features, including the maximum and
mean amplitude of the alpha frequency range (9−11 Hz). We see that mean amplitude of eyes-closed signal is
4.5 times larger than eyes-opened signal.](https://image.slidesharecdn.com/1524839-230809011845-263c567b/85/Electroencephalography-based-brain-computer-interface-using-neural-networks-3-320.jpg)
![TELKOMNIKA Telecommun Comput El Control
Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien)
1073
Figure 7. MSE in training MLP neural model for EEG classification
3.3. Controlling wheelchair
3.3.1. Experimental wheelchair configurations
In Figure 1, a flowchart of EEG data acquisition based on MindWave Mobile 2 to control a
wheelchair is shown. The wheelchair is capable of moving in five different directions, namely forward,
backward, left turn, right turn, and stopping. The experimental wheelchair’s architecture is presented in
Figure 8(a) and a flowchart of EEG classification is shown in Figure 8(b).
In this system, [H1] and [H2] are permanent magnet motors that operate on 24 volts, with a power
rating of 250 watts, and are equipped with a 75 RPM reduction gearbox. [H3] comprises a pair of power
supplies that provide power to the wheelchair system. [H4] is a microcontroller board that transmits and
receives EEG data from the computer via the ZigBee RF transceiver module [H8]. It also controls the speed of
the two motors for movement and measures signals from the ultrasonic sensor [H5] located at the front and back of
the wheelchair to avoid obstacles. [H6] is an integrated circuit consisting of two H-bridges that utilize Mosfet
IRF3205 to drive two motors with a maximum current of 10 A for each motor (maximum current of up to 30 A).
(a) (b)
Figure 8. Wheelchair systems and flowchart of EEG classification: (a) photo of wheelchair and (b) flowchart
3.3.2. Flowchart to control a wheelchair
In sub-section 3.2, we extracted 9 features and utilized the MLP neural model to classify them into 5
gestures, namely opened-eyes, closed eyes, blink per second, double-blink, and look-up. However, in reality,
only 3 gestures, namely closed eyes, look-up, and double-blink, are suitable for controlling a wheelchair.
Two gestures consisting of opened eyes and a blink per second are not feasible, since these gestures are
infrequent human occurrences. In light of this, the authors propose a method of controlling the wheelchair
based on the 3 above gesture signals, as demonstrated in Table 4. In which, 3 gestures including closed-eyes,
look-up, and double-blink are denoted as A, B, and C, respectively. Figure 8(b) shows a flowchart to read
EEG signals via MindWave Mobile 2 and classification based on MLP neural model on the embedded
computer to send signals to control the wheelchair.](https://image.slidesharecdn.com/1524839-230809011845-263c567b/85/Electroencephalography-based-brain-computer-interface-using-neural-networks-6-320.jpg)
![ ISSN: 1693-6930
TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 5, October 2023: 1068-1075
1074
Table 4. EEG signal frame to control a wheelchair
No
Control signal
Wheelchair
Frame 1 Frame 2
1 B B Move forward
2 B C Turn left
3 C B Turn right
4 C C Move backward
5 A Stop
3.3.3. Wheelchair performance results
The wheelchair is operated by 5 participants, in which 1 person is got data set for training and
validating MLP neural model in part 3.2, 1 disabled person, and 3 normal people. Each person controls the
wheelchair 20 times. Table 5 shows the wheelchair performance results of 5 participants.
Based on the above results, an accuracy of 78% was attained for forward and backward movement, with
72% for turning right, 68% for turning left, and 70% for stopping. The highest accuracy of 92% was recorded for
the 5th
individual whose sample data was used for training and validating the MLP neural classifier. Conversely,
the lowest accuracy of 50% was achieved by the 1st
individual due to the limitations of using brain waves to control
eye movements. The 2nd
to 4th
participants were not familiar with using brain waves to control eye gestures and
had yet to undergo MLP neural network training, resulting in accuracies ranging from 68% to 80%. Although
the wheelchair performed well overall, all participants should be trained before using it to increase accuracy.
Table 5. The wheelchair performance results
Operation
Disabled person Normal people not yet training Trained
Avarage
1 2 3 4 5
Move forward 6/10 7/10 8/10 8/10 10/10 78%
Move backward 5/10 7/10 9/10 8/10 10/10 78%
Turn left 4/10 6/10 8/10 7/10 9/10 68%
Turn right 4/10 7/10 8/10 8/10 9/10 72%
Stop 6/10 7/10 7/10 7/10 8/10 70%
Avarage 50% 68% 80% 76% 92% 73.2%
4. CONCLUSION
The study successfully extracted 9 features from 5 distinct action gestures, namely opened-eyes,
closed-eyes, blink per second, double blink, and lookup, both in the time and frequency domains. These features
were then processed and classified using MLP neural networks to effectively control the movement of a
wheelchair, including forward, backward, left, right, and stopping. The experiment aimed to evaluate the
effectiveness of this approach, which resulted in a remarkable 99% accuracy in EEG classification performance.
Moreover, the average performance of the brain-controlled wheelchair achieved by 5 individuals was over 75%,
indicating the system’s robustness and practicality. However, due to the EEG device’s fixed position with only
one pair of sensors, it could only receive brain waves related to concentration and meditation, limiting its ability
to receive signals from other brain regions. Additionally, collecting EEG data from stimuli that cause eye
fatigue is not ideal. To address these limitations, future work will use EEG sensors that measure multiple points,
specifically in the posterior brain regions, to directly collect Mu waves related to EEG motor imagery signals to
better classification of motor intentions and enhanced control of the wheelchair.
REFERENCES
[1] S. Y. Gordleeva et al., “Real-time EEG–EMG human–machine interface-based control system for a lower-limb exoskeleton,”
IEEE Access, vol. 8, pp. 84070–84081, 2020, doi: 10.1109/ACCESS.2020.2991812.
[2] T. A. Izzuddin, N. M. Safri, and M. A. Othman, “Mental imagery classification using one-dimensional convolutional neural
network for target selection in single-channel BCI-controlled mobile robot,” Neural Computing and Applications, vol. 33,
pp. 6233–6246, 2021, doi: 10.1007/s00521-020-05393-6.
[3] Z. Jin, G. Zhou, D. Gao, and Y. Zhang, “EEG classification using sparse Bayesian extreme learning machine for brain–computer
interface,” Neural Computing and Applications, vol. 32, pp. 6601–6609, 2020, doi: 10.1007/s00521-018-3735-3.
[4] T. Tuncer, S. Dogan, and A. Subasi, “EEG-based driving fatigue detection using multilevel feature extraction and iterative hybrid
feature selection,” Biomedical Signal Processing and Control, vol. 68, 2021, doi: 10.1016/j.bspc.2021.102591.
[5] N. Robinson, T. W. J. Chester, and Smitha K. G., “Use of Mobile EEG in Decoding Hand Movement Speed and Position,” IEEE
Transactions on Human-Machine Systems, vol. 51, no. 2, pp. 120–129, 2021, doi: 10.1109/THMS.2021.3056274.
[6] E. Maiorana, “Deep learning for EEG-based biometric recognition,” Neurocomputing, vol. 410, pp. 374–386, 2020,
doi: 10.1016/j.neucom.2020.06.009.
[7] M. Wang, J. Hu, and H. A. Abbass, “BrainPrint: EEG biometric identification based on analyzing brain connectivity graphs,”
Pattern Recognition, vol. 105, 2020, doi: 10.1016/j.patcog.2020.107381.
[8] J. C. Henry, “Electroencephalography: basic principles, clinical applications, and related fields,” Neurology, vol. 67, no. 11, 2006,
doi: 10.1212/01.wnl.0000243257.85592.9a.](https://image.slidesharecdn.com/1524839-230809011845-263c567b/85/Electroencephalography-based-brain-computer-interface-using-neural-networks-7-320.jpg)
![TELKOMNIKA Telecommun Comput El Control
Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien)
1075
[9] C. D. Gratta, V. Pizzella, F. Tecchio, and G. L. Romani, “Magnetoencephalography-a noninvasive brain imaging method with 1
ms time resolution,” Reports on Progress in Physics, vol. 64, 2001, doi: 10.1088/0034-4885/64/12/204.
[10] D. J. Heeger and D. Ress, “What does fMRI tell us about neuronal activity?,” Nature Reviews Neuroscience, vol. 3, pp. 142–151,
2002, doi: 10.1038/nrn730.
[11] H. Adeli, Z. Zhou, and N. Dadmehr, “Analysis of EEG records in an epileptic patient using wavelet transform,” Journal of
Neuroscience Methods, vol. 123, no. 1, pp. 69–87, 2003, doi: 10.1016/S0165-0270(02)00340-0.
[12] Y. Li, R. -G. Zhou, R. Xu, J. Luo, and S. -X. Jiang, “A quantum mechanics-based framework for EEG signal feature extraction
and classification,” IEEE Transactions on Emerging Topics in Computing, vol. 10, no. 1, pp. 211-222, 2022,
doi: 10.1109/TETC.2020.3000734.
[13] M. A. Rahman, M. F. Hossain, M. Hossain, and R. Ahmmed, “Employing PCA and t-statistical approach for feature extraction
and classification of emotion from multichannel EEG signal,” Egyptian Informatics Journal, vol. 21, no. 1, pp. 23–35, 2020,
doi: 10.1016/j.eij.2019.10.002.
[14] V. Bajaj, S. Taran, S. K. Khare, and A. Sengur, “Feature extraction method for classification of alertness and drowsiness states
EEG signals,” Applied Acoustics, vol. 163, 2020, doi: 10.1016/j.apacoust.2020.107224.
[15] C. Huang, Y. Xiao, and G. Xu, “Predicting Human Intention-Behavior through EEG signal analysis using Multi-Scale CNN,” IEEE/ACM
Transactions on Computational Biology and Bioinformatics, vol. 18, no. 5, pp. 1722-1729, 2021, doi: 10.1109/TCBB.2020.3039834.
[16] R. R. -Aguilar, J. A. O. -López, I. O. -Pineda, and S. S. -Urrieta, “Feature extraction from EEG spectrograms for epileptic seizure
detection,” Pattern Recognition Letters, vol. 133, pp. 202–209, 2020, doi: 10.1016/j.patrec.2020.03.006.
[17] M. Li and W. Chen, “FFT-based deep feature learning method for EEG classification,” Biomedical Signal Processing and
Control, vol. 66, 2021, doi: 10.1016/j.bspc.2021.102492.
[18] S. Fawaz, K. S. Sim, and S. C. Tan, “Encoding Rich Frequencies for Classification of Stroke Patients EEG Signals,” IEEE
Access, vol. 8, pp. 135811–135820, 2020, doi: 10.1109/ACCESS.2020.3011185.
[19] R. Sharma, M. Kim, and A. Gupta, “Motor imagery classification in brain-machine interface with machine learning algorithms:
Classical approach to multi-layer perceptron model,” Biomedical Signal Processing and Control, vol. 71, 2022,
doi: 10.1016/j.bspc.2021.103101.
[20] S. Raghu, N. Sriraam, Y. Temel, S. V. Rao, and P. L. Kubben, “EEG based multi-class seizure type classification using convolutional
neural network and transfer learning,” Neural Networks, vol. 124, pp. 202–212, 2020, doi: 10.1016/j.neunet.2020.01.017.
[21] G. Dai, J. Zhou, J. Huang, and N. Wang, “HS-CNN: a CNN with hybrid convolution scale for EEG motor imagery classification,”
Journal of Neural Engineering, vol. 17, no. 1, 2020, doi: 10.1088/1741-2552/ab405f.
[22] A. M. Roy, “A CNN model with feature integration for MI EEG subject classification in BMI,” bioRxiv, 2022,
doi: 10.1101/2022.01.05.475058.
[23] A. M. Roy, “Adaptive transfer learning-based multiscale feature fused deep convolutional neural network for EEG MI
multiclassification in brain–computer interface,” Engineering Applications of Artificial Intelligence, vol. 116, 2022,
doi: 10.1016/j.engappai.2022.105347.
[24] Y. Zhang et al., “Multi-kernel extreme learning machine for EEG classification in brain-computer interfaces,” Expert Systems
with Applications, vol. 96, pp. 302–310, 2018, doi: 10.1016/j.eswa.2017.12.015.
[25] J. Luo, X. Gao, X. Zhu, B. Wang, N. Lu, and J. Wang, “Motor imagery EEG classification based on ensemble support vector
learning,” Computer Methods and Programs in Biomedicine, vol. 193, 2020, doi: 10.1016/j.cmpb.2020.105464.
[26] X. Hu, S. Yuan, F. Xu, Y. Leng, K. Yuan, and Q. Yuan, “Scalp EEG classification using deep Bi-LSTM network for seizure
detection,” Computers in Biology and Medicine, vol. 124, 2020, doi: 10.1016/j.compbiomed.2020.103919.
[27] P. Wang, A. Jiang, X. Liu, J. Shang, and L. Zhang, “LSTM-based EEG classification in motor imagery tasks,” IEEE Transactions on
Neural Systems and Rehabilitation Engineering, vol. 26, no. 11, pp. 2086–2095, 2018, doi: 10.1109/TNSRE.2018.2876129.
[28] E. D. Übeyli, “Analysis of EEG signals by implementing eigenvector methods/recurrent neural networks,” Digital Signal
Processing, vol. 19, no. 1, pp. 134–143, 2009, doi: 10.1016/j.dsp.2008.07.007.
[29] A. M. Roy, R. Bose, and J. Bhaduri, “A fast accurate fine-grain object detection model based on YOLOv4 deep neural network,”
Neural Computing and Applications, vol. 34, pp. 3895-3921, 2022, doi: 10.1007/s00521-021-06651-x.
BIOGRAPHIES OF AUTHORS
Pham Van Huu Thien received his M.Sc. degree in the Faculty of Electronics
Technology, Industrial University of Ho Chi Minh City, Viet Nam in 2021. Currently, he is a
Ph.D. candidate at Industrial University of Ho Chi Minh City, Viet Nam. His major research
interests are embedded systems, digital signal processing, and neural networks. He can be
contacted at email: thienqt1976@gmail.com.
Nguyen Ngoc Son received his M.Sc. and Ph.D. degrees in the Faculty of Electrical
and Electronics Engineering from Ho Chi Minh City University of Technology in 2012 and 2017,
respectively. He is currently a lecturer at the Faculty of Electronics Technology, Industrial
University of Ho Chi Minh City, Viet Nam. His current research interests include artificial
intelligence, robotics, identification and intelligent control, and the internet of things. He can be
contacted at email: nguyenngocson@iuh.edu.vn.](https://image.slidesharecdn.com/1524839-230809011845-263c567b/85/Electroencephalography-based-brain-computer-interface-using-neural-networks-8-320.jpg)
Electroencephalography-based brain-computer interface using neural networks
- 1. TELKOMNIKA Telecommunication Computing Electronics and Control Vol. 21, No. 5, October 2023, pp. 1068~1075 ISSN: 1693-6930, DOI: 10.12928/TELKOMNIKA.v21i5.24839 1068 Journal homepage: http://telkomnika.uad.ac.id Electroencephalography-based brain-computer interface using neural networks Pham Van Huu Thien1,2 , Nguyen Ngoc Son2 1 Department of Electronics, Faculty of Electrical and Electronics, Quang Trung Vocational School, Ho Chi Minh City, Viet Nam 2 Department of Intelligent Systems, Faculty of Electronics Technology, Industrial University of Ho Chi Minh City, Ho Chi Minh City, Viet Nam Article Info ABSTRACT Article history: Received Nov 26, 2022 Revised Feb 27, 2023 Accepted Mar 07, 2023 This study aimed to develop a brain-computer interface that can control an electric wheelchair using electroencephalography (EEG) signals. First, we used the Mind Wave Mobile 2 device to capture raw EEG signals from the surface of the scalp. The signals were transformed into the frequency domain using fast Fourier transform (FFT) and filtered to monitor changes in attention and relaxation. Next, we performed time and frequency domain analyses to identify features for five eye gestures: opened, closed, blink per second, double blink, and lookup. The base state was the opened-eyes gesture, and we compared the features of the remaining four action gestures to the base state to identify potential gestures. We then built a multilayer neural network to classify these features into five signals that control the wheelchair’s movement. Finally, we designed an experimental wheelchair system to test the effectiveness of the proposed approach. The results demonstrate that the EEG classification was highly accurate and computationally efficient. Moreover, the average performance of the brain-controlled wheelchair system was over 75% across different individuals, which suggests the feasibility of this approach. Keywords: Brain-controller wheelchair Electroencephalography Fast fourier transform Neural networks This is an open access article under the CC BY-SA license. Corresponding Author: Nguyen Ngoc Son Department of Intelligent Systems, Faculty of Electronics Technology Industrial University of Ho Chi Minh City, 12 Nguyen Van Bao, Go Vap District Ho Chi Minh City 700000, Viet Nam Email: [email protected] 1. INTRODUCTION Brain-computer interfaces (BCI) are devices that acquire and analyze brain activities, then convert them into output signals to control desired actions. BCI technology has found applications in several fields, including rehabilitation robotics [1], [2], driver fatigue monitoring [3], [4], gaming [5], and cognitive biometrics [6], [7]. Several techniques can be used to measure brain signals, such as electroencephalography (EEG) [8], magnetoencephalography (MEG) [9], and functional magnetic resonance imaging (FMRI) [10]. Among these methods, EEG is widely used in many applications because it is cost-effective, easy to use with minimal training, and can be used for mobile testing. EEG measures the electrical activity of the brain through electrodes placed on the scalp and is divided into six main frequency bands: delta (0.5−4 Hz), theta (4−8 Hz), alpha (8−13 Hz), beta (13−30 Hz), gamma (30−50 Hz), and high gamma (80−150 Hz). Two key challenges in improving the performance of EEG classification are how to use spectral analysis techniques to extract useful features from EEG signals and how to design a robust classifier. With spectral analysis of the EEG signals, there are various techniques used. For example, Adeli et al. [11] proposed the wavelet transform to analyze and characterize the EEG signal of the 3-Hz spike
- 2. TELKOMNIKA Telecommun Comput El Control Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien) 1069 and slow-wave epileptic discharges. Li et al. [12] proposed a quantum wavelet packet transformation (QWPT) module to extract the wavelet packet energy entropy feature of the multi-channel and multi-sample EEG signal. Rahman et al. [13] introduced a hybrid principal component analysis (PCA) and t-statistical approach for feature extraction of emotion from multichannel EEG signals. Bajaj et al. [14] proposed a tunable Q-factor wavelet transform (TQWT) to extract the non-stationary characteristic of the EEG signal for identifying a neurological change in alertness and drowsiness states. A short-time Fourier transform (STFT) was used to analyze EEG signals to obtain time-frequency representations [15], [16]. Li and Chen [17] applied a fast Fourier transform (FFT) to generate the EEG matrix and PCA neural network to learn the hidden information from the frequency matrix of EEG signals. FFT was also used to extract features of stroke patients’ EEG signals [18]. The choice of the appropriate spectral analysis technique depends on the complexity and characteristics of the EEG signals to be analyzed. FFT is a suitable method due to its time-shift invariance in both time and frequency domains and its computational efficiency. To design a powerful classifier, machine learning (ML) techniques have been applied for EEG classification. For example, Sharma et al. [19] showed the performance of the multi-layered perceptron (MLP) model classified the EEG signals with 90% accuracy and half the classification time compared to traditional ML-based models (i.e., support vector machines, k-nearest neighbors, random forest, logistic regression, and Bayes). In [15], [20]–[23] introduced a convolutional neural network (CNN) for EEG classification. In addition, many other ML techniques were also used for classification such as extreme learning machine (ELM) [3], [24], ensemble support vector learning (ESVL) [25], quantum machine learning (QML) [12], long short-term memory (LSTM) and variants [26], [27], recurrent neural network (RNN) [28], you only look once (YOLO) algorithm [29]. These classifiers can be trained on either raw EEG signals or meaningful features extracted from pre-processed signals in the time and frequency domains. However, a dataset is a set of images or series, so the computational cost is quite large. By selecting appropriate features and classifiers, the classification performance of EEG signals can be greatly improved. In this study, we use Mind Wave Mobile 2 to measure raw EEG signals from the surface of the scalp. The signals are then transformed into the frequency domain using FFT. To extract meaningful features for classification, we analyze the signals in both the time and frequency domains and obtain 9 features including amplitude, peak frequency, phase, period, and root mean square for 5 action gestures: opened-eyes, closed-eyes, blink per second, double blink, and lookup. Before analysis, noise in the appropriate frequency bands is removed through filtering. Subsequently, these features are fed into MLP neural networks for classification of 5 signals to control wheelchair movement: forward, backward, turn left, turn right, and stop. Finally, we test the experimental wheelchair system on 5 different individuals to evaluate the effectiveness of our proposed approach. The rest of this paper is structured as: in section 2, we present our methodology for feature extraction and classification using neural networks. Section 3 provides an in-depth analysis of our EEG classification results, along with the brain-controlled wheelchair. Finally, in section 4, we draw our conclusions based on the findings of this study. 2. METHOD 2.1. EEG-based brain-controlled wheelchair approach The diagram in Figure 1 illustrates the process of measuring, extracting, analyzing, and classifying EEG signals. It includes components such as data acquisition, feature extraction, classifier-based MLP neural networks, and a wheelchair system. In which, the MindWave Mobile 2 is used to measure EEG signals from the surface of the scalp, and to transmit power spectrum data to a computer via Bluetooth. The analysis is focused on 5 action gestures: opened-eyes, closed-eyes, blink per second, double-blink, and gaze direction (look-up). The features of the EEG signals are analyzed in both the time and frequency domains and are then used as inputs for the neural network classifier. The output of the classifier is sent as control signals through RF networks to operate a wheelchair. 2.2. EEG feature extraction After collecting the EEG data, it was analyzed and filtered to monitor changes in attention and relaxation. The frequency domain was transformed using FFT, and a set of potential features was extracted, which included 4 time-domain and 5 frequency-domain features. The time-domain EEG feature extraction comprised the maximum, minimum, mean, and root mean square (RMS) of frequency components less than 30 Hz. In the frequency domain, the EEG feature extraction involved the mean and maximum of the alpha band within the range of 8−13 Hz, the mean and maximum frequencies of the delta and theta bands in the range of 1−7 Hz, and the peak frequency. These 9 features were then inputted into the MLP neural network, which classified them into 5 control signals for the wheelchair.
- 3. ISSN: 1693-6930 TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 5, October 2023: 1068-1075 1070 Figure 1. Block diagram of EEG-based brain-controlled wheelchair 2.3. EEG classification using MLP neural model This section employs MLP neural networks for EEG classification. The architecture of the MLP neural classifier includes the input layer 𝑥(𝑘) = [𝑥1(𝑘), 𝑥2(𝑘) , . . . , 𝑥9(𝑘)] represents the 9 features of the EEG signal. The output layer 𝑦(𝑘) = [𝑦1(𝑘), 𝑦2(𝑘) , . . . , 𝑦5(𝑘)] consists of the 5 output signals of the classified neural network, namely opened-eyes, closed-eyes, blinks per second, double blink, and look-up, respectively. 𝑣𝑗𝑛 denotes the weighting value of the input layer, 𝑤𝑚𝑗 represents the weighting value of the hidden layer. 𝑓𝑗(. ) denotes a sigmoid function at the hidden layer, 𝐹(. ) represents a linear activation function at the output layer, respectively. We determine the output neural model as: 𝑦𝑖(𝑣, 𝑤) = 𝐹𝑖(∑ 𝑤𝑖𝑗𝑓𝑗(∑ 𝑣𝑗𝑙𝑥𝑙 𝑚 𝑙=1 ) 𝑞 𝑗=1 ) (1) In the training process, both the input vector 𝑥 and the output vector 𝑦𝑟𝑒𝑓 are known and the synaptic weights (𝑣, 𝑤) are adapted to obtain appropriate functional mappings from the input 𝑥 to the output 𝑦𝑟𝑒𝑓. Generally, the adaptation process can be carried out by minimizing the network error function which is based on a measure of closeness in terms of a mean sum of square error (MSSE) criterion: 𝑀𝑆𝑆𝐸((𝑣, 𝑤) , 𝑍𝑁) = 1 2𝑁 ∑ [𝑦𝑟𝑒𝑓(𝑘) − 𝑦(𝑘|(𝑣, 𝑤))] 𝑇 [𝑦𝑟𝑒𝑓(𝑘) − 𝑦(𝑘|(𝑣, 𝑤))] 𝑁 𝑡=1 (2) Where, the training data 𝑍𝑁 is specified by 𝑍𝑁 = {[𝑥(𝑘), 𝑦𝑟𝑒𝑓(𝑘)]|𝑘 = 1, . . . , 𝑁}. 3. RESULTS AND DISCUSSION 3.1. Extracted features dataset The data was collected by measuring each person’s actions for 30 minutes, corresponding to five different gestures: opening their eyes, closing their eyes, blinking per second, double blinking, and looking up. EEG raw data was collected and pre-processed to remove electrical signal noise at 30 Hz. The detailed analysis of one participant’s features is as: the base state is the opened-eyes gesture, and the features of the remaining four action gestures were compared to the base state to identify potential gestures. Figure 2 displays both the EEG raw signal in the time domain and the EEG signal in the frequency domain after FFT. Figure 2. EEG signal pattern when opened-eyes After analyzing numerous signal samples, it was found that the “eyes-closed” features differ from the eyes-opened features in the alpha frequency range (8−13 Hz), particularly in the frequency range of 9−11 Hz, as depicted in Figure 3(a) and Figure 3(b). Table 1 shows the extracted features, including the maximum and mean amplitude of the alpha frequency range (9−11 Hz). We see that mean amplitude of eyes-closed signal is 4.5 times larger than eyes-opened signal.
- 4. TELKOMNIKA Telecommun Comput El Control Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien) 1071 When analysing the blink action gesture, it becomes evident that the blink per second and double blink features differ from the opened-eyes features in the time domain and in the frequency range of 1−7 Hz, as shown in Figure 4(a), Figure 4(b), Figure 4(c), Figure 5(a), Figure 5(b) and Figure 5(c). Table 2 provides the feature extraction details for the opened-eyes, blink, and double blink gestures, including the maximum and mean amplitude, peak frequency (fmax) in the 1−7 Hz range, and maximum, minimum, and mean amplitude in the time domain. The results show that the features are sufficiently different to distinguish between the blink per second and double blink features. Only one gaze direction gesture called the look-up, which involves moving the gaze from the center to the top. The EEG signal analysis in the frequency domain does not differ from the base state of opened-eyes, except in the time domain, as shown in Figure 6(a) and Figure 6(b). To extract features in this case, a low-pass filter was utilized to eliminate the component signal with a frequency greater than 30 Hz. After completing the filtering process, the root mean square (RMS) amplitude in the time domain was 141.96 and 1817.51 for opened-eyes and look-up, respectively. Table 1. Opened-eyes and closed-eyes feature extraction Gesture Alpha signal in (9-11Hz) range Max (dB) Mean (dB) Opened-eyes 2.70 2.17 Closed-eyes 14.09 9.80 (a) (b) Figure 3. EEG alpha signal pattern in (9−11 Hz) domain: (a) opened-eyes and (b) closed-eyes (a) (b) (c) Figure 4. EEG pattern of opened-eyes, blink, double blink in (1-7 Hz) domain: (a) opened-eyes, (b) blink per second, and (c) double blink Table 2. Opened-eyes, blink, and double-blink feature extraction Gesture 1−7 Hz frequency domain (dB) Time-domain (uV) Max Mean fmax Max Min Mean Opened-eyes 42.65 10.27 1 Hz 106.00 -13.00 42.65 Blink 72.54 44.23 3 Hz 440.00 -280.00 44.39 Double blink 264.98 64.62 6 Hz 551.00 -297.00 47.51 In summary, a total of 300 samples were collected from 1 participant shows as Table 3. The dataset was split into two categories: 210 samples for training data and 90 samples for validating and testing data. The table includes 𝑥1 and 𝑥2, which represent the maximum and mean amplitude in the 9−11 Hz range, respectively.
- 5. ISSN: 1693-6930 TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 5, October 2023: 1068-1075 1072 𝑥3, 𝑥4, and 𝑥5 indicate the maximum, mean, and peak frequency in the 1−7 Hz range, respectively. 𝑥6, 𝑥7, 𝑥8, and 𝑥9 denote the maximum, minimum, mean, and RMS in the time domain, respectively. Finally, 𝑦1 to 𝑦5 are the five gestures considered in this study, namely opened-eyes, closed-eyes, blink per second, double-blink, and look-up. (a) (b) (c) Figure 5. EEG pattern of opened-eyes, blink, double blink in the time domain: (a) opened-eyes, (b) blink per second, and (c) double blink (a) (b) Figure 6. EEG pattern of opened eyes and look-up in the time domain: (a) opened-eyes and (b) look-up Table 3. The data set extracted features for classification No 𝑥-input 𝑦-output 𝑥1 𝑥2 𝑥3 𝑥4 𝑥5 𝑥6 𝑥7 𝑥8 𝑥9 1 2 3 4 5 1 6 3 44 12 1 120 -40 44 216 1 0 0 0 0 … … … … … … … … … … - - - - - 60 10 8 40 13 1 106 -65 40 194 1 0 0 0 0 61 7 5 40 10 1 115 -40 40 455 0 1 0 0 0 … … … … … … … … … … - - - - - 120 8 7 54 12 1 167 -28 54 454 0 1 0 0 0 121 46 34 98 71 4 664 -350 44 293 0 0 1 0 0 … … … … … … … … … … - - - - - 180 87 75 93 52 3 691 -291 49 268 0 0 1 0 0 181 94 75 187 103 6 614 -305 55 358 0 0 0 1 0 … … … … … … … … … … - - - - - 240 91 63 248 90 5 677 -289 56 323 0 0 0 1 0 241 11 6 35 15 1 182 -110 35 1563 0 0 0 0 1 … … … … … … … … … … - - - - - 300 11 8 53 12 1 214 -156 53 1944 0 0 0 0 1 3.2. EEG classification results In this section, the neural network toolbox of MATLAB is utilized to perform EEG classification using the data set described in Table 3. After a trial and error run, the number of suitably hidden layer neurons is determined to be 9. The MLP neural model is trained using 20 epochs, with mean squared error (MSE) value of 6.10−9 obtained for the training data and 1.10−8 for the testing data. The MSE value for EEG classification is 0.0065597. Figure 7 depicts the MSE achieved during the training of the MLP model. Following successful training, the MLP neural model is employed for EEG classification to control the electrical wheelchair.
- 6. TELKOMNIKA Telecommun Comput El Control Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien) 1073 Figure 7. MSE in training MLP neural model for EEG classification 3.3. Controlling wheelchair 3.3.1. Experimental wheelchair configurations In Figure 1, a flowchart of EEG data acquisition based on MindWave Mobile 2 to control a wheelchair is shown. The wheelchair is capable of moving in five different directions, namely forward, backward, left turn, right turn, and stopping. The experimental wheelchair’s architecture is presented in Figure 8(a) and a flowchart of EEG classification is shown in Figure 8(b). In this system, [H1] and [H2] are permanent magnet motors that operate on 24 volts, with a power rating of 250 watts, and are equipped with a 75 RPM reduction gearbox. [H3] comprises a pair of power supplies that provide power to the wheelchair system. [H4] is a microcontroller board that transmits and receives EEG data from the computer via the ZigBee RF transceiver module [H8]. It also controls the speed of the two motors for movement and measures signals from the ultrasonic sensor [H5] located at the front and back of the wheelchair to avoid obstacles. [H6] is an integrated circuit consisting of two H-bridges that utilize Mosfet IRF3205 to drive two motors with a maximum current of 10 A for each motor (maximum current of up to 30 A). (a) (b) Figure 8. Wheelchair systems and flowchart of EEG classification: (a) photo of wheelchair and (b) flowchart 3.3.2. Flowchart to control a wheelchair In sub-section 3.2, we extracted 9 features and utilized the MLP neural model to classify them into 5 gestures, namely opened-eyes, closed eyes, blink per second, double-blink, and look-up. However, in reality, only 3 gestures, namely closed eyes, look-up, and double-blink, are suitable for controlling a wheelchair. Two gestures consisting of opened eyes and a blink per second are not feasible, since these gestures are infrequent human occurrences. In light of this, the authors propose a method of controlling the wheelchair based on the 3 above gesture signals, as demonstrated in Table 4. In which, 3 gestures including closed-eyes, look-up, and double-blink are denoted as A, B, and C, respectively. Figure 8(b) shows a flowchart to read EEG signals via MindWave Mobile 2 and classification based on MLP neural model on the embedded computer to send signals to control the wheelchair.
- 7. ISSN: 1693-6930 TELKOMNIKA Telecommun Comput El Control, Vol. 21, No. 5, October 2023: 1068-1075 1074 Table 4. EEG signal frame to control a wheelchair No Control signal Wheelchair Frame 1 Frame 2 1 B B Move forward 2 B C Turn left 3 C B Turn right 4 C C Move backward 5 A Stop 3.3.3. Wheelchair performance results The wheelchair is operated by 5 participants, in which 1 person is got data set for training and validating MLP neural model in part 3.2, 1 disabled person, and 3 normal people. Each person controls the wheelchair 20 times. Table 5 shows the wheelchair performance results of 5 participants. Based on the above results, an accuracy of 78% was attained for forward and backward movement, with 72% for turning right, 68% for turning left, and 70% for stopping. The highest accuracy of 92% was recorded for the 5th individual whose sample data was used for training and validating the MLP neural classifier. Conversely, the lowest accuracy of 50% was achieved by the 1st individual due to the limitations of using brain waves to control eye movements. The 2nd to 4th participants were not familiar with using brain waves to control eye gestures and had yet to undergo MLP neural network training, resulting in accuracies ranging from 68% to 80%. Although the wheelchair performed well overall, all participants should be trained before using it to increase accuracy. Table 5. The wheelchair performance results Operation Disabled person Normal people not yet training Trained Avarage 1 2 3 4 5 Move forward 6/10 7/10 8/10 8/10 10/10 78% Move backward 5/10 7/10 9/10 8/10 10/10 78% Turn left 4/10 6/10 8/10 7/10 9/10 68% Turn right 4/10 7/10 8/10 8/10 9/10 72% Stop 6/10 7/10 7/10 7/10 8/10 70% Avarage 50% 68% 80% 76% 92% 73.2% 4. CONCLUSION The study successfully extracted 9 features from 5 distinct action gestures, namely opened-eyes, closed-eyes, blink per second, double blink, and lookup, both in the time and frequency domains. These features were then processed and classified using MLP neural networks to effectively control the movement of a wheelchair, including forward, backward, left, right, and stopping. The experiment aimed to evaluate the effectiveness of this approach, which resulted in a remarkable 99% accuracy in EEG classification performance. Moreover, the average performance of the brain-controlled wheelchair achieved by 5 individuals was over 75%, indicating the system’s robustness and practicality. However, due to the EEG device’s fixed position with only one pair of sensors, it could only receive brain waves related to concentration and meditation, limiting its ability to receive signals from other brain regions. Additionally, collecting EEG data from stimuli that cause eye fatigue is not ideal. To address these limitations, future work will use EEG sensors that measure multiple points, specifically in the posterior brain regions, to directly collect Mu waves related to EEG motor imagery signals to better classification of motor intentions and enhanced control of the wheelchair. REFERENCES [1] S. Y. Gordleeva et al., “Real-time EEG–EMG human–machine interface-based control system for a lower-limb exoskeleton,” IEEE Access, vol. 8, pp. 84070–84081, 2020, doi: 10.1109/ACCESS.2020.2991812. [2] T. A. Izzuddin, N. M. Safri, and M. A. Othman, “Mental imagery classification using one-dimensional convolutional neural network for target selection in single-channel BCI-controlled mobile robot,” Neural Computing and Applications, vol. 33, pp. 6233–6246, 2021, doi: 10.1007/s00521-020-05393-6. [3] Z. Jin, G. Zhou, D. Gao, and Y. Zhang, “EEG classification using sparse Bayesian extreme learning machine for brain–computer interface,” Neural Computing and Applications, vol. 32, pp. 6601–6609, 2020, doi: 10.1007/s00521-018-3735-3. [4] T. Tuncer, S. Dogan, and A. Subasi, “EEG-based driving fatigue detection using multilevel feature extraction and iterative hybrid feature selection,” Biomedical Signal Processing and Control, vol. 68, 2021, doi: 10.1016/j.bspc.2021.102591. [5] N. Robinson, T. W. J. Chester, and Smitha K. G., “Use of Mobile EEG in Decoding Hand Movement Speed and Position,” IEEE Transactions on Human-Machine Systems, vol. 51, no. 2, pp. 120–129, 2021, doi: 10.1109/THMS.2021.3056274. [6] E. Maiorana, “Deep learning for EEG-based biometric recognition,” Neurocomputing, vol. 410, pp. 374–386, 2020, doi: 10.1016/j.neucom.2020.06.009. [7] M. Wang, J. Hu, and H. A. Abbass, “BrainPrint: EEG biometric identification based on analyzing brain connectivity graphs,” Pattern Recognition, vol. 105, 2020, doi: 10.1016/j.patcog.2020.107381. [8] J. C. Henry, “Electroencephalography: basic principles, clinical applications, and related fields,” Neurology, vol. 67, no. 11, 2006, doi: 10.1212/01.wnl.0000243257.85592.9a.
- 8. TELKOMNIKA Telecommun Comput El Control Electroencephalography-based brain-computer interface using neural networks (Pham Van Huu Thien) 1075 [9] C. D. Gratta, V. Pizzella, F. Tecchio, and G. L. Romani, “Magnetoencephalography-a noninvasive brain imaging method with 1 ms time resolution,” Reports on Progress in Physics, vol. 64, 2001, doi: 10.1088/0034-4885/64/12/204. [10] D. J. Heeger and D. Ress, “What does fMRI tell us about neuronal activity?,” Nature Reviews Neuroscience, vol. 3, pp. 142–151, 2002, doi: 10.1038/nrn730. [11] H. Adeli, Z. Zhou, and N. Dadmehr, “Analysis of EEG records in an epileptic patient using wavelet transform,” Journal of Neuroscience Methods, vol. 123, no. 1, pp. 69–87, 2003, doi: 10.1016/S0165-0270(02)00340-0. [12] Y. Li, R. -G. Zhou, R. Xu, J. Luo, and S. -X. Jiang, “A quantum mechanics-based framework for EEG signal feature extraction and classification,” IEEE Transactions on Emerging Topics in Computing, vol. 10, no. 1, pp. 211-222, 2022, doi: 10.1109/TETC.2020.3000734. [13] M. A. Rahman, M. F. Hossain, M. Hossain, and R. Ahmmed, “Employing PCA and t-statistical approach for feature extraction and classification of emotion from multichannel EEG signal,” Egyptian Informatics Journal, vol. 21, no. 1, pp. 23–35, 2020, doi: 10.1016/j.eij.2019.10.002. [14] V. Bajaj, S. Taran, S. K. Khare, and A. Sengur, “Feature extraction method for classification of alertness and drowsiness states EEG signals,” Applied Acoustics, vol. 163, 2020, doi: 10.1016/j.apacoust.2020.107224. [15] C. Huang, Y. Xiao, and G. Xu, “Predicting Human Intention-Behavior through EEG signal analysis using Multi-Scale CNN,” IEEE/ACM Transactions on Computational Biology and Bioinformatics, vol. 18, no. 5, pp. 1722-1729, 2021, doi: 10.1109/TCBB.2020.3039834. [16] R. R. -Aguilar, J. A. O. -López, I. O. -Pineda, and S. S. -Urrieta, “Feature extraction from EEG spectrograms for epileptic seizure detection,” Pattern Recognition Letters, vol. 133, pp. 202–209, 2020, doi: 10.1016/j.patrec.2020.03.006. [17] M. Li and W. Chen, “FFT-based deep feature learning method for EEG classification,” Biomedical Signal Processing and Control, vol. 66, 2021, doi: 10.1016/j.bspc.2021.102492. [18] S. Fawaz, K. S. Sim, and S. C. Tan, “Encoding Rich Frequencies for Classification of Stroke Patients EEG Signals,” IEEE Access, vol. 8, pp. 135811–135820, 2020, doi: 10.1109/ACCESS.2020.3011185. [19] R. Sharma, M. Kim, and A. Gupta, “Motor imagery classification in brain-machine interface with machine learning algorithms: Classical approach to multi-layer perceptron model,” Biomedical Signal Processing and Control, vol. 71, 2022, doi: 10.1016/j.bspc.2021.103101. [20] S. Raghu, N. Sriraam, Y. Temel, S. V. Rao, and P. L. Kubben, “EEG based multi-class seizure type classification using convolutional neural network and transfer learning,” Neural Networks, vol. 124, pp. 202–212, 2020, doi: 10.1016/j.neunet.2020.01.017. [21] G. Dai, J. Zhou, J. Huang, and N. Wang, “HS-CNN: a CNN with hybrid convolution scale for EEG motor imagery classification,” Journal of Neural Engineering, vol. 17, no. 1, 2020, doi: 10.1088/1741-2552/ab405f. [22] A. M. Roy, “A CNN model with feature integration for MI EEG subject classification in BMI,” bioRxiv, 2022, doi: 10.1101/2022.01.05.475058. [23] A. M. Roy, “Adaptive transfer learning-based multiscale feature fused deep convolutional neural network for EEG MI multiclassification in brain–computer interface,” Engineering Applications of Artificial Intelligence, vol. 116, 2022, doi: 10.1016/j.engappai.2022.105347. [24] Y. Zhang et al., “Multi-kernel extreme learning machine for EEG classification in brain-computer interfaces,” Expert Systems with Applications, vol. 96, pp. 302–310, 2018, doi: 10.1016/j.eswa.2017.12.015. [25] J. Luo, X. Gao, X. Zhu, B. Wang, N. Lu, and J. Wang, “Motor imagery EEG classification based on ensemble support vector learning,” Computer Methods and Programs in Biomedicine, vol. 193, 2020, doi: 10.1016/j.cmpb.2020.105464. [26] X. Hu, S. Yuan, F. Xu, Y. Leng, K. Yuan, and Q. Yuan, “Scalp EEG classification using deep Bi-LSTM network for seizure detection,” Computers in Biology and Medicine, vol. 124, 2020, doi: 10.1016/j.compbiomed.2020.103919. [27] P. Wang, A. Jiang, X. Liu, J. Shang, and L. Zhang, “LSTM-based EEG classification in motor imagery tasks,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 26, no. 11, pp. 2086–2095, 2018, doi: 10.1109/TNSRE.2018.2876129. [28] E. D. Übeyli, “Analysis of EEG signals by implementing eigenvector methods/recurrent neural networks,” Digital Signal Processing, vol. 19, no. 1, pp. 134–143, 2009, doi: 10.1016/j.dsp.2008.07.007. [29] A. M. Roy, R. Bose, and J. Bhaduri, “A fast accurate fine-grain object detection model based on YOLOv4 deep neural network,” Neural Computing and Applications, vol. 34, pp. 3895-3921, 2022, doi: 10.1007/s00521-021-06651-x. BIOGRAPHIES OF AUTHORS Pham Van Huu Thien received his M.Sc. degree in the Faculty of Electronics Technology, Industrial University of Ho Chi Minh City, Viet Nam in 2021. Currently, he is a Ph.D. candidate at Industrial University of Ho Chi Minh City, Viet Nam. His major research interests are embedded systems, digital signal processing, and neural networks. He can be contacted at email: [email protected]. Nguyen Ngoc Son received his M.Sc. and Ph.D. degrees in the Faculty of Electrical and Electronics Engineering from Ho Chi Minh City University of Technology in 2012 and 2017, respectively. He is currently a lecturer at the Faculty of Electronics Technology, Industrial University of Ho Chi Minh City, Viet Nam. His current research interests include artificial intelligence, robotics, identification and intelligent control, and the internet of things. He can be contacted at email: [email protected].