Enhanced Accuracy for Six Commands using

Sequential Hybrid P300/SSVEP BCI System

Mehreen Naeem a,, Muhammad Jawad Khan a, , Kashif Javed a, Saddaf Rubab a
a
National University of Sciences and Technology (NUST), Islamabad, Pakistan

Abstract
This study presents a hybrid brain computer interface (BCI) system that achieves
better accuracy based on event related potential signals. Following system based on
the P300-SSVEP hybrid sequential BCI system to decode six reactive brain
commands using ensemble classifier. For P300 we have used dataset of six images
stimulus. The feature vector is extracted from the denoised waves after filtered
through least mean square (LMS) filters. Extracted feature samples are fed into
support vector machine ensemble classifier model for classification. To minimize the
false detection rate, output from ensemble classifier trigger respective SSVEP
frequency stimulus. EEG signal is recorded at occipital region. After classification of
SSVEP signal command is send to control quadcopter. For BCI application, a virtual
quadcopter environment is created and controlled by proposed hybrid BCI system.
Results of the proposed study using LMS filter demonstrate that the sequential P300
and SSVEP model yields the combine accuracy of 96.98%.

Keywords: Brain-Computer Interface, Hybrid EEG, LMS Filter, Ensemble Classifier


Corresponding authors
Email address: [email protected]
Email address: [email protected]
1. Introduction

Brain-computer interface (BCI) gives a way to control one’s environment without

using the normal neuromuscular output pathways [1,2]. A BCI is an artificial
intelligence system that can identify a certain set of patterns generated by the brain
[3]. It can be categorized into non-invasive, invasive, and semi-invasive.
Electroencephalography (EEG) is a non-invasive BCI modality having better
temporal resolution and less environmental limitation [4].
The most used design for the creation of the non-invasive BCI system is event-
related potentials (ERP) [5]. Individuals who are severely disabled by disorders such
as spinal cord injuries, brainstem stroke, and muscular dystrophies might benefit from
BCI. This technology allows paralyzed people to control prosthetic limbs with their
brains. Improvement in such a system improves the quality of life for a person who is
unable to perform body movements. Other applications include the control of BCI
smart home, Wheelchair control [6], Prosthetics, Quadcopter control, Robotic
Augmentation [7, 8].
Practically most applications use visual evoked potentials to decode brain activity
[9]. It is initiated by visual stimuli, display on the screen and the signal is recorded
from the occipital cortex of the brain [10]. Among other modalities of EEG
approaches, event-related potentials signals, P300 and SSVEP have the advantage of
high information transfer rate (ITR) and high accuracy with minimal user training
[11]. Moreover, fewer EEG are required for the detection of P300 and SSVEP signal
[12]. The P300 potential is an event-related potential, elicited by rare and task-
relevant events of stimulus that can produce a positive deflection in the EEG at
300msec after the onset of the presented stimulus [13 – 15]. Y.Zhang in his research
used the different objects stimulus along with inverted face stimulus to evoke time-
locked P300 signal [16]. To exploit the advantage of each method, the hybrid BCI
system is used to improve accuracy and increase the number of commands [17, 18].
Mostly used hybrid systems are EEG-EOG [19, 20], P300-SSSEP [21], MI-SSVEP
[5], MI- P300 [22]. SSVEP conjunction with P300 is an ideal candidate to be used in
a hybrid BCI system, as both can be elicited by visual stimulus [23]. SSVEP is a
phase-locked response by the visual stimulus modulated at a certain repetitive stimuli

2
rate [24]. Usually, stimulus frequencies are higher than 6Hz [25]. P300 is difficult to
detect so it still required several trials of stimulus to obtain a strong signal whereas
SSVEP is detected in single trial [14].
The rest of this study is organized as follows. Section 2 investigates the related
work and presents a review on the previous methods for P300/SSVEP acquisition.
Section 3 provides details of the dataset and describes the design details of the
preprocessing process and proposed methodology. Section 4 presents the
experimental results. Section 5 outlines the conclusion and discusses on the
limitations of the proposed method.

2. Related Work
In BCI, extraction of patterns containing different order information in the signal is
a challenging task [26]. There are several modalities for the extraction of features by
removing artifacts from the corrupted signal [27, 28]. In [29] feature vector for P300
is obtained by averaging a signal and spatial/ temporal filtering. In adaptive filter
output signal is subtracted from the raw EEG signal to eliminate the artifacts [30, 27].
Hybrid strategies for features extraction were used to exploit the advantage of each
method such as Empirical mode decomposition (EMD), Blind Source Separation
(BSS), Wavelet Transform, and Regression Methods [31]. Drawback of EMD is that
it is limited to numerical simulations and lack a formal mathematical framework [32].
In another related research, P300 signal is extracted using fusion of deep features with
the temporal features based on sparse auto encoder [33]. To obtain SSVEP features,
the power spectrum of the data was taken followed by the band-pass filtered using
FIR filter [34]. The identification of EEG signal relies on a classification algorithm
[12]. Selected feature samples are fed into a relative classifier that automatically
estimating the class of data. Mostly classifiers used in BCI systems are Linear
Discriminant Analysis (LDA), Support Vector Machine (SVM), Neural Networks
(NN) [35, 36], Multi-Layer Perceptron (MLP) [26] and deep learning methods [37]. A
data partitioning technique Ensemble classifier has been introduced to reduce the
classifier variability. The average classification accuracy of hemodynamic responses
using extreme machine learning is reported as 76.7% for two classes [38]. A

3
quadcopter in a 3D physical space was driven by means of a motor imagery EEG BCI
system [4] and accuracy acquired up to 90.5% of all valid targets [39].
Several Hybrid EEG techniques have been used in the BCI field to improve
accuracy by using simple machine learning algorithms [30, 40]. In most existing BCI,
the decoding of output from EEG relies on the single classification algorithm [41 –
43]. P300 signal is a weak signal, and it has required an average number of trials and
a strong classifier with more training for detection. The problem with a single
classifier is that it may cause over-fitting over extra training. So, we have used
ensemble classifiers to reduce overfitting.
Following research presents the performance of P300 classification in comparison
with the work of Turnip et al. [44]. In the present BCI research, we have developed a
sequential Hybrid P300/SSVEP system that will generate six commands accurately
and improve classification accuracy by using an ensemble machine learning
algorithm. Previously recorded P300 dataset of six images stimulus is used.
Preprocessing and filtration of signal is done using adaptive LMS filters. After which
features are extracted and then concatenating them to feed into ensemble SVM
(eSVM) classifier. The classified output is given to the SSVEP paradigm to flash the
stimulus. A person is asked to look at the stimulus for several seconds. Hence SSVEP
signal is recorded, and the features vector is extracted by taking power spectrum of a
signal after band-pass filtration. The classified output of the SSVEP signal is used to
control quadcopter in a virtual environment.

3. Methods
3.1 Datasets

The present study uses the dataset of Hoffman [45] for the six-choice signal
paradigm. The dataset is available on the website of the EPFL BCI group. The
experiment was performed with four disabled and four healthy. The disabled
subjects had limb muscle control abilities, and all were wheelchair-bound. Some
disabled subjects had varying communication, and some are suffered from
dysarthria. The healthy subjects were Ph.D. students and none of them had any
neurological deficits. In this experiment, users were asked to focus on the screen

4
 containing six images stimulus. The healthy subjects were Ph.D. students and none
of them had any neurological deficits. In this experiment, users were asked to focus
on the screen containing six images stimulus. The stimulus was images of a
television, a lamp, a window, a telephone, a door, and a radio. The images were
flashed in a random order such that one image at a time. Stimulus flashed for
100ms and in the next 300ms, none of the images was flashed. The total inter-
stimulus interval (ISI) was 400ms. The data was recorded at a 4048 Hz sampling
frequency from 32 electrodes according to the 10-20 system. The dataset contained
four sessions for each subject. Each session was consisting of six runs such that one
run for each of the six stimuli. One run consisted of an average of 135 target and
non-target trials. The entire data for one subject was consist of 3240 trials. Fig.1
[45] presents the protocol of data recording. The dataset which we are using for
SSVEP dataset recorded by Vilic [46]. The data was taken on six flickering stimuli
and recorded by placing electrodes at Oz for signal, Fz as a reference, and Fpz as
ground using the 10-20 system. A six-choice flickering stimulus panel was used for

Fig. 1. A timeline of P300 experiment. It consists of six images (telephone, television,

lamp, door, window, radio) stimulus flickering in random evoke P300 signal.

5
 a collection of a dataset. Each stimulus flash with multiple frequencies of 6Hz,
6.5Hz, 7Hz, 7.5Hz, 8.2Hz, 9.3Hz. The experiment was performed with five
subjects. The data was recorded in two sessions for each of the subjects. Each
session had ten trials of sixteen seconds. In Fig. 2 recording scenario is explained.

3.2 Data Processing

Prior to feature extraction, we have preprocessed the signals to get the best possible
result. For P300, the dataset of 32 channels is given. A Signal is downsampled from
2048 Hz to 256 Hz followed by the band-pass (BFP) filtration. A Butterworth band-
pass filter with cutoff frequencies of 1 Hz to 20 Hz is used to remove the irrelevant
information from the signal. Furthermore, to reduce the effect of outliers such as large
amplitude outliers due to eye blinking and muscle movement, the data is processed by
windsorization and standardization method. In our experiment, we are using four
electrodes as signal electrodes Fz, Cz, Pz, and Oz. As P300 signal is strongly
generated on the parietal region so by averaging Cz Pz and Oz we obtained the
required signal. Averaged signal is subtracted from the frontal channel Fz to obtain
the activity signal. The window of 1000ms is used (starting at the stimulus onset) for
the extraction of single trials.

Fig. 2. SSVEP dataset structure for five subjects. Each consists of two sessions and
further divided in ten trial each.

6
 Preprocessing of SSVEP is quite simple as compared to P300. SSVEP is detected
in the frequency domain as the range of stimulus frequencies is known. According to
given stimulus frequencies, the EEG raw signal is filtered by bandpass filters between
4Hz to 11Hz. The filtered output that is in the time domain, is converted into the
frequency domain by taking Fast Fourier Transform (FFT) as shown in Fig 3. By
analyzing the power spectrum of the FFT signal, the dominant frequency having
maximum amplitude is considered as the frequency of the stimulus on which the
subject has focused on.
The feature extractions intend to extract the important information from brain
signals. We have used Adaptive Least Means Square (LMS) filters to extract the P300
features. By using preprocessed signal, LMS changes its filter coefficient based on the
desired or reference signal by reducing the least mean square error. In our case desired
reference signal is the known noise signal. This delayed version of the noisy signal is
extracted by finding common signal between target and non-target signal of P300.
The algorithm tries to find the common noise in both the reference signal and filter
output signal, by estimating the error between the two. In Fig. 4 x(k) is the one trial
signal of 1000ms, y(k) is the filtered output, d(k) represents the reference noise signal.
The whole process is repeated for M channels and N trails.

𝑦 = 𝑤 × 𝑥 (1)
𝑒 = 𝑑 − 𝑦 (2)

𝑤 = 𝑤 + (µ × 𝑒 × 𝑥) (3)

Fig. 3. SSVEP signal classification followed by pre-processing. Bandpass filters, Fast

Fourier Transform and power spectrum density functions are used to extract the noise
free signal.

7
Fig. 4. Least Mean Square Filter (LMS) design for P300 features extraction followed
by pre-processing of signal. Bandpass filter, down-sampling, windsorizing and
channel selection methods are the steps involved in the pre-processing of the data.

In mentioned equations, w denotes the learning weights, e is the error between the
reference signal d and the filter output y, where the µ is the learning parameter When
error gets minimized, output y which is filtered signal is subtracted from the initial
input signal to get the desired signal as shown in Fig 5. The LMS filtered output is
then analysis by statistical methods. Following statistical measurement as a feature
vector are computed: positive peak, positive peak latency, negative peak, negative
peak latency, peak to peak, peak to peak latency, and root mean square (RMS) value.
SSVEP is a periodic response generated when a subject has focused on some
repetitive stimulus. It is confined to a specific set of stimulus frequencies, so instead
of the time domain, it is analyzed in the frequency domain. The power spectrum of a
signal is taken followed by FFT. The frequency having the highest power respective
to their amplitude is considered as a feature for the classification process as shown in
Fig 6. The red dot in a figure shows the highest power of the frequency in a signal.

8
 Fig. 5. Filtered output from LMS filter for target and non-target P300

Fig. 6. Power spectrum of signal for SSVEP feature extraction. Red dot shows the
highest power of the frequency in a signal.

9
3.3 Classification
An average of trials is needed for the appropriate P300 signal detection. As P300 is
difficult to detect in a single trial, so strong classifier is needed for classification. We
are using ensemble SVM for the classification of P300 of a single trial. The
architecture of model is shown in Fig. 7. A good surface decision can be made by
using ensemble SVM if the input features are well organized. We have 3240 samples
for the single subject. After shuffling and standardization, data is divided into training
and test data in the ratio of 4:1. Training data is further split into three equal divisions
by the method of Bootstrap Aggregation commonly called bagging. In this method
samples from observation are selected randomly with replacement. After the splitting
of training data into three subsets, an SVM model has developed from each

Fig. 7. Ensemble Classification model for P300 detection. Training data split into
Boostrap samples

10
concatenated features dataset. SVM model with the polynomial kernel of order two is
used for training. Kernel function is expressed as:

𝐾(𝑥𝑖 , 𝑥𝑗 ) = (1 + 𝑥𝑖′ 𝑥𝑗 )
2 (4)

The output from all the standalone models of the same types is gathered. The final
prediction is based on the aggregation of prediction of each base model. For testing
and validation, the test data is passed through each three trained models of SVM and
predicted. By the voting method, final prediction can be made as:

𝑌 = 𝑚𝑎𝑥 (𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛 (𝑀1 , 𝑀2 , 𝑀3 ), 𝑥) (5)

Where M1, M2, M3 are the trained models of SVM, and x is the test sample.
SSVEP Classification is done by SVM with kernel. It is easy to detect, so with
little training of data can give remarkable results. In AVI SSVEP dataset, we have 100
samples, 80 random samples are separate out for training. Classification of SSVEP
followed by preprocessing is shown in Fig. 8.

3.4 The Proposed Method

We have implemented an offline sequential hybrid P300/SSVEP system that
generates six commands. Six images stimulus oddball paradigm dataset is used to
generate the P300 signal. extracted and fed into the ensemble model. The output from
the classifier triggers the respective SSVEP frequency signal. After preprocessing and
filtering, extracted features are classified by a simple SVM algorithm. After LMS
filtration followed by preprocessing, features are classified by a simple SVM
algorithm. Flight control of quadcopter is one of the applications of BCI. The

Fig. 8. SSVEP classification workflow using SVM learner

11
classified command decoded from sequential hybrid BCI system is transmitted with a
fixed time interval to the quadcopter. The generated output is used to control the
virtual quadcopter in a 3D environment by MATLAB simulation. Fig. 9 provides a
workflow of the proposed scheme for quadcopter control. A Graphical User Interface
(GUI) is developed, that displays the control of quadcopter. The aim of our project is
to perform maximum operations by using lesser number of outputs. The total control
commands of quadcopter are eight out of which we have formulated six commands
for the flight of quadcopter: ascend/descend movement, forward/backward movement,
clockwise and counterclockwise rotation. Following mentioned commands cover all
the possible movements of flight. The remaining right and left movements are
covered by turning the quadcopter in the desired direction and then move forward.
The display panel on the GUI shows the respective commands for all the six outputs
generated from the hybrid model shown in Fig. 10.

4. Experiment
In this proposed scheme of sequential hybrid BCI model, P300 signal
preprocessing step includes the channel selection, windowing, standardization and
LMS filter. The filtered output for eight subjects with target and non-target P300
results are shown in Fig. 11. The positive deflection of P300 signal of target stimulus
is observed nearly at 300ms in graphs for subjects 1, 2, 3 and 8. For Subjects 4, 5, 6

Fig. 9. Structure of adopted methodology. Hybrid sequential P300 and SSVEP signal
architecture used to control flight of the quadcopter

12
Fig. 10. Graphical User Interface for quadcopter flight control. It consists of a panel
having list of commands and trajectory plot

and 7 peaks are detected after and before 300ms.

The ensemble SVM classifier is used for classification purpose. We have
implemented the three models of SVM, and final prediction is made by the
aggregation of the prediction form each model. Receiver operating characteristic
(ROC) curve plots are used to analyze the threshold value for classifier. It illustrates
the performance of classification model at all classification thresholds shown in
Fig.12. The accuracy for eight subjects using eSVM model is relative higher than
90%. Subject 5 achieved minimum accuracy of 94% and subject 7 achieved highest
accuracy among all subject with the percentage of 97.84.
SSVEP data consist of six stimulus frequencies. The raw data is of single channel
Oz. Its pre-processing is done by band-pass filtering. The Fig. 13 and Fig. 14 shows
the result of band-pass filtering and power spectrum followed by the FFT for all six
frequencies. The red marked point in the figures represents the maximum power of
the frequency in the signal.
The classification of SSVEP dataset is done by SVM. Total number of samples for
a subject are 100 which is split into 80% training data and 20% test data. The average
accuracy for all subjects is reported on test set as 94.6%. The output was used to test

13
Fig. 11. LMS filter results for eight subjects. Target signal was shown with blue and
non-target signal with red line

14
Fig. 12. ROC plots for three SVM models

15
 the movement of a virtual quadcopter. The multiple trajectories were executed in
MATLAB stimulation. The trajectories followed by the quadcopter were shown in
Fig. 15.

Fig. 13. SSVEP Pre-processing at occipital region for frequencies 6Hz, 6.5Hz and 7 Hz.

Fig. 14. SSVEP Pre-processing at occipital region for frequencies 7.5Hz, 8.2H and, 9.3
Hz.

16
 Fig. 15. Quadcopter trajectories using commands generated from hybrid model.

5. Discussion
In this research, we have decoded six active BCI commands by using sequential
hybrid P300/SSVEP. Available dataset of P300 and SSVEP was used. P300 dataset of
Hoffmann was composed of six images stimulus. If target P300 is truly classified by

17
Table 1. Accuracies reported in previous related research

Papers Stimulus Average

Accuracy

Kunda et al. 6x6 P300 Speller 95.5%

(P300)

Katyal et al. 2x4 P300 paradigm 92.30 %

(P300/SSVEP)

Merino et al. 6 flickering stimuli 85%

(SSVEP)

Proposed Method 6 images paradigm / flickering stimuli 96.98%

(P300/SSVEP)

ensemble learner e.g., an oddball is generated in response to stimulus 1 then from

the SSVEP dataset, the frequency signal listed on position one is selected. SSVEP
signal after pre-processing is further classified by the SVM model and command is
generated. The final output is fed to the quadcopter in a visual environment. In the
results, P300 signal classification accuracy is enhanced by using ensemble learner.
Moreover, the LMS filter helps to prominent the P300 wave. Previously Turnip et al.
[44] presented a P300 features extracted model based on adaptive recursive (AR)
filter on the same dataset. The drawback of the mentioned model was to feed a
reference signal to the filter in the form of a desired signal respective to input. It is
just like that you converted the signal to your desired output which must be predicted
from classification. In contrast to this, we have used LMS filters with a noise
reference signal to the filter in the form of a desired signal respective to input. It is
just like that you converted the signal to your desired output which must be predicted
from classification. In contrast to this, we have used LMS filters with a noise
reference signal, measured by a common noise in target and non-target signal. Fig. 11
represents the result of LMS filter output for eight subjects. By using noise signal as a
reference, the target and non-target signal is precisely filtered out.
Numerous researches have been carried out to measure and classify the single-trial
P300 response for the purpose of quick detection by improving accuracy in hybrid

18
with SSVEP [42], [47] shown in Table 1. Kunda et al. had implemented sparse
autoencoder for deep features extraction with an ensemble of SVMs on the BNCI
Horizon dataset and achieved an average accuracy of 95.5% [33]. In the present study,
an ensemble SVM was used with LMS filter for P300 classification with 96.63%
accuracy. The comparison of 8 subjects with eLDA model and eSVM model are
represented in a graph Fig. 16. For P300 single classifier may cause an overfitting
problem over more training. The ensemble classifier reduces the effect of overfitting
by making a prediction from the aggregation of multiple models [48].
BCI-based SSVEP has a high information transfer rate (ITR) with a less training
period. Ojha et al. had used empirical mode decomposition (EMD) decomposition
mode on AVI SSVEP dataset with a classification accuracy of 81.45% [49]. In
another research maximum of 87.5% accuracy on the same dataset was reported [50].
By using a frequency-based approach we have achieved 94.6% classification accuracy
by SVM learner. SSVEP required training for its detection. If we combine P300 and
SSVEP less training for its detection. If we combine P300 and SSVEP modalities
there is less chance of false detection.

Classification Accuracy
100 97.2% 96.7% 97.26% 97.43% 97.84% 97.2%
95.45%
94%
95

90
85.04% 84.61% 84.82%
85 83.7% 82.72% 83.33%
81.58% 82.24%

80

75

70
S1 S2 S3 S4 S5 S6 S7 S8

eLDA eSVM

Fig. 16. Comparison of classification accuracies for eight subjects

19
6. Conclusion
We have employed a sequential hybrid P300/SSVEP system that generates six
commands accurately. Six images stimulus oddball paradigm is used to generate the
P300 signal. The LMS filter used with common noise as a reference provides a noise-
free signal. After LMS filtration followed by pre-processing, features are extracted
and fed into the ensemble SVM model. Classified output triggers the SSVEP
frequency signal respective to ensemble output. After signal acquisition and filtering,
extracted features are classified by the SVM algorithm. The generated output is used
to control the virtual quadcopter on MATLAB simulation. The collective accuracy for
Hybrid Model achieved using the ensemble classifier was 96.98%, relatively
improved than other conventional classifiers.

Acknowledgements

This research work was supported by National University of Sciences and

Technology, Pakistan.

References

[1] Z.T. Al-qaysi, B.B. Zaidan, A.A. Zaidan, M.S. Suzani, Computer Methods
and Programs in Biomedicine A review of disability EEG based wheelchair
control system : Coherent taxonomy , open challenges and recommendations,
Comput. Methods Programs Biomed. 164 (2018) 221–237.
https://doi.org/10.1016/j.cmpb.2018.06.012.

[2] A. Nourmohammadi, M. Jafari, T.O. Zander, A Survey on Unmanned Aerial

Vehicle Remote Control Using Brain – Computer Interface, (2018) 1–12.

[3] L.F. Nicolas-Alonso, J. Gomez-Gil, Brain computer interfaces, a review,

Sensors. 12 (2012) 1211–1279. https://doi.org/10.3390/s120201211.

20
[4] M. Aljalal, R. Djemal, S. Ibrahim, Robot Navigation Using a Brain Computer
Interface Based on Motor Imagery, J. Med. Biol. Eng. (2018).
https://doi.org/10.1007/s40846-018-0431-9.

[5] X. Duan, S. Xie, X. Xie, Y. Meng, Z. Xu, Quadcopter Flight Control Using a
Non-invasive Multi-Modal Brain Computer Interface, 13 (2019) 1–11.
https://doi.org/10.3389/fnbot.2019.00023.

[6] Fernández-Rodríguez, F. Velasco-Álvarez, R. Ron-Angevin, Review of real

brain-controlled wheelchairs, J. Neural Eng. 13 (2016).
https://doi.org/10.1088/1741-2560/13/6/061001.

[7] Q. Liu, C. Wang, B. Zhang, L. Duan, X. Zhang, T. Sun, W. Shang, Y. Shen,

Z. Lin, X. Li, W. Li, W. Li, Y. Wang, M.G. Fujie, Z. Wu, Kinematics
Analysis and Motion Planning of a Redundant Robotic Manipulator for
Surgical Intervention, (2017) 5–10.

[8] R. Sharma, EEG Signal Based Movement Control for Mobile Robots, (2019).

[9] G. Schalk, J. Mellinger, Brain Sensors and Signals, (2010) 9–36.

[10] A. Rakshit, A. Konar, S. Member, A.K. Nagar, A Hybrid Brain-Computer

Interface for Closed- Loop Position Control of a Robot Arm, 7 (2020) 1344–
1360.

[11] E. Yin, T. Zeyl, S. Member, R. Saab, T. Chau, S. Member, A Hybrid Brain –

Computer Interface Based on the Fusion of P300 and SSVEP Scores, 23
(2015) 693–701.

[12] J. Zhao, W. Li, M. Li, Comparative Study of SSVEP- and P300- Based
Models for the Telepresence Control of Humanoid Robots, (2015) 1–18.
https://doi.org/10.1371/journal.pone.0142168.

[13] R. Fazel-rezai, A Review of P300 , SSVEP , and Hybrid P300 / SSVEP

Brain-Computer Interface Systems, (2013).

21
 https://doi.org/10.5772/56135.This.

[14] M.R. Meshriky, S. Eldawlatly, G.M. Aly, An Intermixed Color Paradigm for
P300 Spellers : A Comparison with Gray-scale Spellers, (2017).
https://doi.org/10.1109/CBMS.2017.123.

[15] Y. Zhang, Q. Zhao, J. Jing, X. Wang, human faces A novel BCI based on
ERP components sensitive to configural processing of human faces, (2012).
https://doi.org/10.1088/1741-2560/9/2/026018.

[16] B.H. Kim, M. Kim, S. Joevaluated, Quadcopter fl ight control using a low-
cost hybrid interface with EEG-based classi fi cation and eye tracking,
Comput. Biol. Med. (2014) 1–11.
https://doi.org/10.1016/j.compbiomed.2014.04.020.

[17] M.J. Khan, K. Hong, Hybrid EEG–fNIRS-Based Eight-Command Decoding

for BCI: Application to Quadcopter Control, 11 (2017).
https://doi.org/10.3389/fnbot.2017.00006.

[18] L. Anh, T. Nguyen, Q.C. Lam, H.K. Nguyen, Developing a Wheelchair

System Controlled Based on EEG Signal and Eye-Direction, 19 (2019) 115–
122.

[19] J. Ma, Y. Zhang, A. Cichocki, F. Matsuno, A Novel EOG / EEG Hybrid

Human – Machine Interface Adopting Eye Movements and ERPs :
Application to Robot Control, IEEE Trans. Biomed. Eng. 62 (2015) 876–889.
https://doi.org/10.1109/TBME.2014.2369483.

[20] J. Li, J. Pu, H. Cui, X. Xie, S. Xu, T. Li, Y. Hu, An Online P300 Brain –
Computer Interface Based on Tactile Selective Attention of Somatosensory
Electrical Stimulation, J. Med. Biol. Eng. (2018).
https://doi.org/10.1007/s40846-018-0459-x.

[21] Y. Yu, Z. Zhou, Y. Liu, E. Yin, N. Zhang, Z. Wang, Self-Paced Operation of

a Wheelchair based on a Hybrid Brain-Computer Interface Combining Motor

22
 Imagery and P300 potential, 4320 (2017) 1–10.
https://doi.org/10.1109/TNSRE.2017.2766365.

[22] A.M. Norcia, L.G. Appelbaum, J.M. Ales, B.R. Cottereau, B. Rossion, The
steady-state visual evoked potential in vision research : A review, 15 (2015)
1–46. https://doi.org/10.1167/15.6.4.doi.

[23] Z. Wang, Y. Yu, M. Xu, Y. Liu, E. Yin, Z. Zhou, Towards a Hybrid BCI
Gaming Paradigm Based on Motor Imagery and SSVEP Towards a Hybrid
BCI Gaming Paradigm Based on Motor Imagery and SSVEP, Int. J. Human–
Computer Interact. 00 (2018) 1–9.
https://doi.org/10.1080/10447318.2018.1445068.

[24] M. Xu, H. Qi, B. Wan, T. Yin, Z. Liu, A hybrid BCI speller paradigm
combining P300 potential and the SSVEP, (2013).
https://doi.org/10.1088/1741-2560/10/2/026001.

[25] Z. Oralhan, A New Paradigm for Region-Based P300 Speller in Brain

Computer Interface, IEEE Access. 7 (2019) 106618–106627.
https://doi.org/10.1109/ACCESS.2019.2933049.

[26] F. Lotte, M. Congedo, A. Lécuyer, F. Lamarche, F. Lotte, M. Congedo, A.

Lécuyer, F. Lamarche, B. Arnaldi, A review of classification algorithms for
EEG-based brain – computer interfaces To cite this version : HAL Id : inria-
00134950 A Review of Classification Algorithms for EEG-based Brain-
Computer Interfaces, (2007).

[27] P.S. Kumar, R. Arumuganathan, K. Sivakumar, C. Vimal, Removal of

Artifacts from EEG Signals using Adaptive Filter through Wavelet
Transform, (2008) 2138–2141.

[28] R. Spies, R. Gassert, A penalized time-frequency band feature selection and

classification procedure for improved motor intention decoding in
multichannel EEG, J. Neural Eng. (2019) aaf046.

23
 https://doi.org/10.1088/1741-2552/aaf046.

[29] J. Long, J. Wang, T. Yu, An Efficient Framework for EEG Analysis with
Application to Hybrid Brain Computer Interfaces Based on Motor Imagery
and P300, 2017 (2017).

[30] A. Katyal, R. Singla, Biomedical Signal Processing and Control A novel

hybrid paradigm based on steady state visually evoked potential & P300 to
enhance information transfer rate, Biomed. Signal Process. Control. 59 (2020)
101884. https://doi.org/10.1016/j.bspc.2020.101884.

[31] X. Jiang, G. Bian, Removal of Artifacts from EEG Signals : A Review, (2019)
1–18. https://doi.org/10.3390/s19050987.

[32] E. Deléchelle, J. Lemoine, O. Niang, E. Deléchelle, J. Lemoine, O. Niang,

Empirical mode decomposition : an analytical approach for sifting process To
cite this version : Empirical Mode Decomposition : An Analytical Approach
for Sifting Process, (2012).

[33] S. Kundu, S. Ari, ScienceDirect P300 based character recognition using

sparse autoencoder with ensemble of SVMs, Integr. Med. Res. 39 (2019)
956–966. https://doi.org/10.1016/j.bbe.2019.08.001.

[34] Y. Chen, S. Chen, I.A.E. Zaeni, C. Wu, applied sciences Fuzzy Tracking and
Control Algorithm for an SSVEP-Based BCI System, (2016) 1–11.
https://doi.org/10.3390/app6100270.

[35] S. Naseem, K. Javed, M.J. Khan, S. Rubab, M.A. Khan, Y. Nam, Integrated
CWT-CNN for epilepsy detection using multiclass EEG dataset, Comput.
Mater. Contin. 69 (2021) 471–486.
https://doi.org/10.32604/cmc.2021.018239.

[36] S. Kundu, S. Ari, Biomedical Signal Processing and Control P300 based
character recognition using convolutional neural network and support vector
machine, Biomed. Signal Process. Control. 55 (2020) 101645.

24
 https://doi.org/10.1016/j.bspc.2019.101645.

[37] A. Sarwar, K. Javed, M.J. Khan, S. Rubab, O.Y. Song, U. Tariq, Enhanced
accuracy formotor imagery detection using deep learning for BCI, Comput.
Mater. Contin. 68 (2021) 3825–3840.
https://doi.org/10.32604/cmc.2021.016893.

[38] X. Yin, B. Xu, C. Jiang, Y. Fu, Classification of Hemodynamic Responses

Associated With Force and Speed Imagery for a Brain-Computer Interface,
(2015). https://doi.org/10.1007/s10916-015-0236-0.

[39] K. Lafleur, K. Cassady, A. Doud, K. Shades, E. Rogin, B. He, Quadcopter

control in three-dimensional space using a noninvasive motor imagery-based
brain-computer interface, J. Neural Eng. 10 (2013).
https://doi.org/10.1088/1741-2560/10/4/046003.

[40] H. Wang, Y. Li, J. Long, An asynchronous wheelchair control by hybrid EEG

– EOG brain – computer interface, (2014). https://doi.org/10.1007/s11571-
014-9296-y.

[41] N. V Manyakov, N. Chumerin, A. Combaz, M.M. Van Hulle, Comparison of

Classification Methods for P300 Brain-Computer Interface on Disabled
Subjects, 2011 (2011). https://doi.org/10.1155/2011/519868.

[42] F.E.M. Labib, I.A. Fouad, MULTIPLE CLASSIFICATION TECHNIQUES

P300 BCI SYSTEM, 32 (2020) 1–12.
https://doi.org/10.4015/S1016237220500106.

[43] A. Turnip, S.S. Hutagalung, J. Pardede, D. Soetraprawata, P300 Detection

Using a Multilayer Neural Network Classifier Based on Adaptive Feature
Extraction, 2 (2013) 63–75. https://doi.org/10.5923/j.ijbcs.20130205.01.

[44] U. Hoffmann, J. Vesin, T. Ebrahimi, K. Diserens, An efficient P300-based

brain – computer interface for disabled subjects, 167 (2008) 115–125.
https://doi.org/10.1016/j.jneumeth.2007.03.005.

25
[45] A. Vilic, L. Revision, The AVI SSVEP Dataset, (2013) 1–6.

[46] L. Meriño, T. Nayak, P. Kolar, G. Hall, Z. Mao, J. Pack, Y. Huang,

Asynchronous control of unmanned aerial vehicles using a steady-state visual
evoked potential-based brain computer interface, Brain-Computer Interfaces.
2621 (2017) 1–14. https://doi.org/10.1080/2326263X.2017.1292721.

[47] M. Wasim, M. Sajjad, F. Ramzan, U.G. Khan, W. Mahmood, A Review and

Classification of Widely used Offline Brain Datasets, 9 (2018) 399–408.

[48] Y. Lee, H. Kim, Biomedical Signal Processing and Control A data

partitioning method for increasing ensemble diversity of an, Biomed. Signal
Process. Control. 39 (2018) 53–63.
https://doi.org/10.1016/j.bspc.2017.07.025.

[49] M.K. Ojha, M.K. Mukul, Detection of Target Frequency from SSVEP Signal
Using Empirical Mode Decomposition for SSVEP Based BCI, Wirel. Pers.
Commun. 116 (2021) 777–789. https://doi.org/10.1007/s11277-020-07738-9.

[50] A.F. Demir, H. Arslan, I. Uysal, Bio-Inspired Filter Banks for Frequency
Recognition of SSVEP-Based Brain-Computer Interfaces, IEEE Access. 7
(2019) 160295–160303. https://doi.org/10.1109/ACCESS.2019.2951327.

26