2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC)

October 11-14, 2020. Toronto, Canada

Complex Motor Imagery-based Brain-Computer Interface System:

A Comparison Between Different Classifiers
Seung-Bo Lee, Min-Kyung Jung, Hakseung Kim, Seong-Whan Lee and Dong-Joo Kim*
Department of Brain and Cognitive Engineering, Korea University, Seoul, South Korea

Abstract— Motor imagery (MI) classification is important as improving the mobility of patients with severe neuromuscular
the emerging research interest of brain computer interface (BCI) disorders [12, 13].
due to its potential about real-world application. Advancing
Performance of MI-BCI is mostly determined by efficacies
manipulation and control technology of external devices such as
robotics, the need of MI for complex and human-like movements of EEG feature extraction and movement classification
is growing. The two most important procedures that influence techniques, thus the two procedures are of utmost importance
the performance of MI-BCI are feature extraction and in developing reliable, high-powered MI-BCI systems. With
classification. Although there have been recent studies on feature the recent advancement of robotics manipulation and control
extraction for complex, there is no consensus on the classifier technology, complex and human-like movements can be more
suitable for complex MI. This study aimed to identify the best fluently executed by external devices. Thus, for MI-BCI to be
classifier for complex MI decoding. applied for controlling such devices, further advancements in
Electroencephalography (EEG) recordings measured during feature extraction and classification technologies should be
complex MI, which are hand grasping, spreading, pronation and made.
supination, were used for binary (grasp vs. twist) and
quaternary classification. Time domain parameter, which have
As for the former endeavor, recent investigations regarding
shown suitability for complex movement decoding in previous the efficacies of three EEG domains (i.e., temporal, spectral
works, was used as the EEG feature. Four types of ten machine and spatial) have identified that features from temporal
learning classifiers, which have been applied to MI-BCI, were domains could be better suited for classifying complex MI
compared. tasks [14-16]. As for the latter problem (i.e., classification), a
Shrinkage regularized linear discriminant analysis (SRLDA) wide range of machine learning classifier has been applied,
exhibited the best classification accuracy in both binary (92.8%) including but not limited to: variants of tree algorithms [17],
and quaternary (55.2%). In the case of training and testing time, support vector machines (SVM) [18], linear and quadratic
a small amount of time for real-time analysis were needed, discriminant analyses (LDA and QDA, respectively) [19],
except random forest and logistic regression. Naïve Bayes [20, 21], logistic regression [22], and k nearest
This study showed that SRLDA is an appropriate classifier neighbor (KNN) [21]. However, while various classifiers
for complex MI classification, due to its ability to handle having been applied to MI-BCI systems, there is no consensus
stationary and high dimensionality feature, TDP. The findings on the classifier suitable for complex MI.
suggest that complex MI-BCI could gain more benefit from
applying linear and shrinkage regularized model (i.e., SRLDA). This study aimed to identify the most suitable classifier for
decoding complex MI. The classification performances of
decision tree algorithms, SVM, discriminant analysis-based
I. INTRODUCTION classifiers and other methods, coupled with time-domain
parameter (TDP) [14], were compared against complex binary
Brain-computer interface (BCI) has been under extensive
and multi-class MI acquired with the use of a robotic arm
research due to its potential for providing an alternative
system. The remainder of this article is organized with 1)
communication pathway between the human brain and an
Method and Materials including data description, feature
external device [1-4]. Well-known measures of brain activity
extraction, decision tree, SVM, discriminant analysis-based,
for BCI include electroencephalography (EEG) [5-7], mag-
other classifiers, and statistical analysis, 2) Results including
netoencephalography (MEG) [8, 9], functional magnetic
accuracy comparison of each types of classifiers, performance
resonance imaging [10], and the near-infrared spectroscopy
comparison between types of classifiers, and execution time
(NIRS) [11]. Among those, EEG is the most frequently used
comparison, 3) Discussion and 4) Conclusion.
neurophysiological signal for BCI due to its relatively low cost,
convenience, superior temporal resolution, and non-
invasiveness [9]. Though EEG provides a wide variety of brain
activity features, the motor imagery (MI) has been at the center
of interest in the fields of BCI-mediated neurorehabilitation, or

This research was supported in part by Institute for Information & and ICT, MSIT) under Grant 2019R1A2C1003399, and in part by Research
communications Technology Promotion(IITP) grant funded by the Korea Foundation of Korea (NRF) grant funded by the Korea government (Ministry
government(MSIT) (No. 2017-0-00432, Development of non-invasive of Science and ICT, MSIT) under Grant NRF-2020R1C1C1006773.
integrated BCI SW platform to control home appliances and external devices All authors are with the Department of Brain and Cognitive Engineering,
by user's thought via AR/VR interface), in part by Research Foundation of Korea University. (*corresponding author to provide phone: +82-2-3290-
Korea (NRF) grant funded by the Korea government (Ministry of Science 5929; fax: +82-2-3290-3970; e-mail: dongjookim@ korea.ac.kr).

978-1-7281-8526-2/20/$31.00 ©2020 IEEE 2496

Authorized licensed use limited to: Korea University. Downloaded on April 09,2024 at 09:52:01 UTC from IEEE Xplore. Restrictions apply.
 II. METHOD AND MATERIALS technique to be more resilient to noise and outliers than a
single decision tree classifier.
A. Dataset Description
EEG recordings from thirteen healthy subjects (all males, D. SVM Classifiers
right-handed, average age of 26.8 ± 1.8 years) were used. None • Linear SVM
of the subjects had known cognitive deficits and experience
with BCI experiments prior to this study. The study was Linear SVM [25] constructs an optimal hyperplane that
approved by the Institutional Review Board (IRB) at Korea maximizes the margin between the two divided planes. With
University (1040548-KU-IRB-17-172-A-2). The MI classes an input space X = [x1, x2, …, xn] and output space y ∈ [-1, 1],
included right-hand grasping, spreading, pronation, and the hyperplane separating the two classes can be defined as the
supination. At the initial stage of each MI trial, a computer following:
monitor displayed an empty screen for 2 seconds, followed by ⃗⃗ ∙ 𝑥 +b = 0 (2)
𝑤
a ‘Rest’ instruction for 3 seconds. The subject was asked to
imagine one of the four tasks of the movement in the following where w is “weight (i.e.,the orthogonal vector of determining
‘Visual cue’ period for 3 seconds. The subject performed MI the hyperplane)” and b is bias (i.e.,the distance from the
movement during the following ‘Motor imagery’ period for 4 original hyperplane). In terms of optimization problem,
seconds. EEG signals were measured with Brain Vision Lagrange multiplier can be applied as follows:
Recorder (BrainProduct GmbH) recording system. 64 1
min⁡( ‖𝑤‖2 + 𝐶 ∑𝑛𝑖=1 𝜉𝑖 ) (3)
Ag/AgCl electrodes following 10/20 international systems (all 2
the channels referenced to FCz and grounded to the FPz) were where ξ is the slack variable for handling non-linearly
used to record the signals. The signals were digitalized 1000 separable problem, and C is the cost function.
Hz. A band-pass filter of 8-30 Hz was applied. Details of the
experimental paradigm and data are reported in [14]. • Kernel SVM
Kernel SVM [26] includes a mapping function ɸ(x) for
mapping the input space into a higher dimension space where
B. Feature extraction the class can be more discriminative. The feature space is
In this study, TDP, which have shown suitability for defined as the inner product of the mapping function and the
complex movement decoding, was used as the EEG feature. decision function is represented as following:
TDP features are extracted from the time domain but can also
be inferred as frequency domain. Increment of localized neural 𝑦𝑡 = ∑𝑙𝑖=1 𝑎𝑖 𝑦𝑖 𝑘(𝑥𝑖 , 𝑥𝑡 ) + 𝑏 (4)
activity during MI can be distinguished by observing TDP. where k is kernel function and a is Lagrange multipliers for
The TDP is calculated as follows: solving optimization problem. In our study, Gaussian kernel
2
𝑑 𝑖 𝑥(𝑡) ‖𝑥𝑖 −𝑥𝑗 ‖
𝑇𝐷𝑃𝑖 = 𝑣𝑎𝑟 ( ) , 𝑖 = 0, 1, 2 (1) (exp⁡(− ), which is the most frequently used kernel,
𝑑𝑡 𝑖 2𝜎 2
were applied.
where var=variance, i=the number of derivatives. A total of
192 features from 64 channels and three i were used for MI
classification.
E. Discriminant analysis-based classifiers
• Linear discriminant analysis (LDA)
C. Decision tree classifiers LDA [27] is a generalization of Fisher’s discriminant. The
• Classification and regression tree (CART) classification of a given sample X is performed as following:
CART [23] is the most widely used model of the decision 𝑠𝑖𝑔𝑛(𝑤 𝑇 𝑥 + 𝑏) (5)
tree, and it uses a recursive portioning via splitting criteria to
where sign indicates signum function and w the projection
create nodes. CART includes four process steps that are 1)
vector. Classification performance of LDA significantly
building a tree using recursive splitting of nodes where
depends on the estimation of the class means and the common
splitting criteria reached, 2) stopping tree building after the
covariance matrix. If the amount of training data is insufficient,
learning dataset, 3) tree pruning, and 4) selecting an optimal
precise estimation of common covariance can be difficult. To
tree from the sequence of pruned trees, which are not overfitted.
solve this problem, analytic shrinkage regularization can be
The tree pruning is a process that removes insignificant tree
applied to LDA.
nodes, thereby achieving an appropriate fitting and low
computational complexity. • Shrinkage regularized LDA (SRLDA)
• Random forest Analytic shrinkage regularization [28] compensates for the
systemic errors that occur with low amount of data. The
Random forest [24] is a set of decision trees that uses
systematic error may distort the eigenvalue of the covariance
numerous independent decision trees that are created by
matrix. Large eigenvalues induce overestimated covariance
randomly chosen variables. Each created tree predicts a class
and small eigenvalues induce underestimated covariance.
differently and votes for the selected class. An overall
Shrinkage regularization technique counteract these systemic
prediction of random forest is determined by the majority of
errors as following estimation of covariance matrix:
the votes. This ensemble approach allows random forest

2497
Authorized licensed use limited to: Korea University. Downloaded on April 09,2024 at 09:52:01 UTC from IEEE Xplore. Restrictions apply.
 Σ̃(γ) = (1 − γ)Σ̂ + γ
𝑡𝑟𝑎𝑐𝑒⁡Σ
𝐼 (6)
̂ where 𝑁𝑐 is the number of correctly classified trials and 𝑁𝑖 is
𝑑 the number of incorrectly classified trials. Precision and recall
where γ is an optimizing parameter, Σ̃ is the original estimation are given in equation (10).
of the covariance matrix, d indicates the dimensionality of 𝑡𝑝 𝑡𝑝
feature space and I denotes the identity matrix. 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = ; 𝑅𝑒𝑐𝑎𝑙𝑙 = (10)
𝑡𝑝+𝑓𝑝 𝑡𝑝+𝑓𝑛

• Quadratic discriminant analysis (QDA) where tp is true positive, fp is false positive, and fn is false
negative. F-measure is given in equation (11).
QDA [29] is an extension of the LDA. This classifier
2×𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛×𝑟𝑒𝑐𝑎𝑙𝑙
considers quadratic terms of feature, so that the classification 𝐹 − 𝑚𝑒𝑎𝑠𝑢𝑟𝑒 = (11)
𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑟𝑒𝑐𝑎𝑙𝑙
is defined as follows:
Hyper-parameters from all classifiers were optimized by grid
𝑠𝑖𝑔𝑛(𝑥 𝑇 𝐴𝑥 + 𝑤 𝑇 𝑥 + 𝑏) (6)
search [33]. 5-fold cross-validation were used.
where A is the quadratic projection matrix. The quadratic terms
enable the construction of non-linear decision boundaries that
are necessary for optimal classification when the distribution III. RESULTS
of class feature is normal.
A. Accuracy comparison of each types of classifiers
Figures 1 and 2 illustrates the binary (hand grasping vs.
F. Other classifiers wrist-twisting) and quaternary classification (grasp vs. open vs.
left twist vs. right twist) accuracies of all the employed
• Naïve Bayes classifiers, grouped by the types of classifiers (i.e., tree, SVM,
Naïve Bayes [30] is one of the probabilistic classifiers that discriminant analyses, and others). For all classifiers, the
uses vectorized probability to predict classes. Assuming the degradation of accuracy as per an increase in the number of
features of the vector 𝑥 are independent, the probability of x classes was noted. Among each classifier group, the random
being included in class ck can be calculated as follows forest, LSVM, SRLDA, and LR showed the highest accuracy.
according to the Bayes rule: SRLDA showed the best accuracy in both binary (92.8%) and
quaternary (55.2%) classification.
∏𝑑
𝑖=1 𝑃(⃗⃗⃗
𝑥𝑖 |𝑐𝑘 )
𝑃(𝑐𝑘 |𝑥 ) = 𝑃(𝑐𝑘 ) (7)
𝑃(𝑥)

where d is the number of features.

• Logistic regression (LR)
LR [31] is developed to predict the probability with linear
combination of sigmoid function, assuming that the input
space X = [x1, x2, …, xn] is independent and the output y ∈ [-
1, 1] is dependent variable. The probability of the y, when y
belongs to class 1, is represented as following:
𝐶+∑𝑛𝑖=1 𝑎𝑖 𝑥𝑖
𝑒
𝑃(𝑦 = 1|𝑋) = 𝐶+∑𝑛
(8)
1+𝑒 𝑖=1 𝑎𝑖 𝑥𝑖

where C is constant intercept and a is the regression coefficient

related to the independent variable X. Regression coefficient
parameters are estimated by maximum likelihood estimation.
• K-nearest neighbor (KNN)
The KNN [32] is one of the simplest classification algorithms,
and it uses the distance of feature vector to predict class. A
feature vector is classified with the class of the majority in the
trained KNN. In this study, Euclidean metric, which is the Figure 1. Binary classification accuracy of each types of classifiers. (A)
Decision tree algorithms. (B) Support vector machines. (C) Discriminant
most common method for the distance function, was used for analysis-based classifiers. (D) Other methods. Error bars represent the
KNN classification. standard deviation.
G. Performance evaluation
In this study, accuracy, precision, recall, and f-measure
were used to evaluate the performance of classification. For binary classification, the average accuracies of
Classification accuracy is computed as following equation (9): decision tree algorithms, SVM, discriminant analysis-based
𝑁𝑐 classifiers, and other classifier groups were 80.3%, 82.4%,
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (9) 87.5%, and 78.7%, respectively. As of particular note, the
𝑁𝑐 +𝑁𝑖
linear models showed better performance than that of non-
linear models (LSVM (90.8%) > KSVM (73.9%) and LDA
(86.2%) > QDA (83.6%)). The classifiers that showed over 90%

2498
Authorized licensed use limited to: Korea University. Downloaded on April 09,2024 at 09:52:01 UTC from IEEE Xplore. Restrictions apply.
of accuracy were SRLDA and LSVM. The classifier with the
lowest accuracy was found to be Naïve Bayes classifier, which
was 21.1% higher than the chance rate (50.0%).

Figure 3. Classification performance comparison between types of

classifiers. (A) Binary (hand grasping vs. wrist twisting) classification. (B)
Quaternary (grasp vs. open vs. pronation vs. supination) classification.

Figure 2. Quaternary classification accuracy of each types of classifiers.

C. Performance comparison with statistical analysis
(A) Decision tree algorithms. (B) Support vector machines. (C) To investigate the efficiency of classifiers, execution times
Discriminant analysis-based classifiers. (D) Other classifiers. Error bars of individual MI classification procedures including train and
represent the standard deviation.
test models were compared. Table 1 contains the average
execution time of classifiers. Ensemble (random forest) and
combination (LR) models had extremely long training time
In the quaternary classification, average accuracies of periods. Naïve Bayes, KNN, and LSVM showed the shortest
decision tree algorithms, SVM, discriminant-analysis based training time in that order. Most classifiers took longer training
classifiers, and other classifier groups were 43.7%, 44.1%, time as increasing the number of classes. As the number of
48.4%, and 41.9%, respectively. Again, linear models yielded classes increases, the training time increased in proportion. As
better performance than those of non-linear models (LSVM for the test execution, all classifiers required a very short time
(48.5%) > KSVM (39.6%) and LDA (46.7%) > QDA (43.3%)). except for the random forest model. Contrary to training, there
Only one classifier (i.e., SRLDA, Figure 2) showed accuracy was no correlation between the number of classes and the test
over 50%. The classifier with the lowest accuracy was Naïve time.
Bayes classifier.

TABLE I. EXECUTION TIMES OF TEN CLASSIFIERS DURING TRAINING

B. Performance comparison between types of classifiers AND TESTING FOR COMPLEX MOTOR IMAGERY CLASSIFICATION.

Detailed performance reports of classifiers that yielded the Model Task Train (ms) Test (ms)
highest accuracy among each classifier groups (i.e., random
Binary 6.71 0.14
forest, LSVM, SRLDA, and LR) are provided in Figure 3. CART
Among the four classifiers, the SRLDA had highest precision, Quaternary 9.75 0.2
recall and F-measure for both the binary and quaternary Binary 106.29 9.26
Random forest
classification tasks. SRDLA also showed the least difference Quaternary 141.93 11.51
between precision and recall (1.2% in binary and 0.9% in Binary 3.83 0.09
LSVM
quaternary classification), which indicates a balanced Quaternary 12.45 0.29
classification performance. Binary 4.66 1
KSVM
Quaternary 17.18 1.06
Binary 18.61 0.29
LDA
Quaternary 17.84 0.26
Binary 29.66 0.23
SRLDA
Quaternary 33.86 0.22
Binary 12.49 0.68
QDA
Quaternary 15.52 0.57
Naïve Bayes Binary 2.22 0.26

2499
Authorized licensed use limited to: Korea University. Downloaded on April 09,2024 at 09:52:01 UTC from IEEE Xplore. Restrictions apply.
 Quaternary 2.03 0.23 the combination of other well-known feature extraction and
Binary 233.3 0.23 classification techniques would be desirable.
LR
Quaternary 521.85 0.29
Binary 2.32 2.77
KNN
Quaternary 2.96 3 V. CONCLUSION
This study attempted to identify the most suitable machine
IV. DISCUSSION learning classifier for a complex MI-based BCI system. The
This study is an attempt to identify the most suitable results suggest that SRLDA is the most appropriate classifier
classifier for a complex MI-based BCI system. In this when coupled with TDP, due to its ability to handle stationary
endeavor, a total of 10 different classifiers were employed. and high dimensionality of TDP. SRLDA with TDP reached
Discriminant analysis-based classifiers, especially SRLDA, an average accuracy of 92.8% and 55.2% in binary and
showed the best classification performance with relatively quaternary classification, respectively. The findings of this
short execution time. SRLDA also showed the most study with similar designs and larger dataset could promote
appropriate balance between precision and recall. The findings
from including combinations of feature extraction and
of this study indicate that the SRLDA should be the most
suitable for classifying complex MI; considering the classification methods for confidence in results.
importance of trade-off between classification performance
and execution time, LSVM might be the next option.
REFERENCES
Throughout the investigation, linear models showed better
[1] J. Wolpaw and E. W. Wolpaw, Brain-computer interfaces:
classification performance than that of non-linear models. In
principles and practice. OUP USA, 2012.
general, nonlinear models exhibit superior fitting accuracy due [2] J. Shin et al., "Open access dataset for EEG+ NIRS single-trial
to its capacity to detect complex non-linear relationships classification," vol. 25, no. 10, pp. 1735-1745, 2017.
between features and class [34]. For this conceptual advantage, [3] Y. Chen et al., "A high-security EEG-based login system with
it is expected that non-linear models would yield better RSVP stimuli and dry electrodes," IEEE Transactions on
performance for the classification of complex MI, which was Information Forensics and Security, vol. 11, no. 12, pp. 2635-
2647, 2016.
not the case. These seemingly paradoxical findings could be [4] K.-T. Kim, H.-I. Suk, and S.-W. Lee, "Commanding a brain-
explained by stationary characteristics of TDP. TDP was controlled wheelchair using steady-state somatosensory evoked
designed to resolve non-stationarities of EEG via temporal potentials," IEEE Transactions on Neural Systems and
derivatives [35], and for the approximation of stationary data, Rehabilitation Engineering, vol. 26, no. 3, pp. 654-665, 2016.
increasing the simplicity of the linear model can be particularly [5] W.-Y. Hsu and Y.-N. J. J. o. n. m. Sun, "EEG-based motor
imagery analysis using weighted wavelet transform features," vol.
effective [36]. Thus, when coupled with TDP for the 176, no. 2, pp. 310-318, 2009.
classification of complex MI, the performance of linear [6] M.-H. Lee, J. Williamson, D.-O. Won, S. Fazli, and S.-W. Lee,
classification models such as LSVM, LDA, and SRLDA could "A high performance spelling system based on EEG-EOG signals
equal, or even surpass those of non-linear classification with visual feedback," IEEE Transactions on Neural Systems and
models, as indicated in our findings. Rehabilitation Engineering, vol. 26, no. 7, pp. 1443-1459, 2018.
[7] D.-O. Won, H.-J. Hwang, D.-M. Kim, K.-R. Müller, and S.-W.
Among the linear classifiers, SRLDA had superior Lee, "Motion-based rapid serial visual presentation for gaze-
discriminant ability of TDP from EEG during complex MI. independent brain-computer interfaces," IEEE Transactions on
Neural Systems and Rehabilitation Engineering, vol. 26, no. 2, pp.
This seems to be due to the subject and session variability of 334-343, 2017.
EEG and large dimensionality of TDP. Individualized EEG [8] J. Mellinger et al., "An MEG-based brain–computer interface
data might not constantly support robust brain decoding (BCI)," Neuroimage, vol. 36, no. 3, pp. 581-593, 2007.
because of intra-subject and session to session variability [37, [9] M. Ahn et al., "Gamma band activity associated with BCI
38]. Thus, the EEG trials are constrained in same subject and performance: simultaneous MEG/EEG study," Frontiers in
session, so that they usually have small sample size. However, human neuroscience, vol. 7, p. 848, 2013.
[10] J.-H. Lee, J. Ryu, F. A. Jolesz, Z.-H. Cho, and S.-S. J. N. l. Yoo,
multi-channel EEG data has TDP of large dimensionality, "Brain–machine interface via real-time fMRI: preliminary study
because the dimensionality of TDP is determined by multiples on thought-controlled robotic arm," vol. 450, no. 1, pp. 1-6, 2009.
of channels. Analytic shrinkage regularization was proposed [11] A. F. Abdelnour and T. J. N. Huppert, "Real-time imaging of
to solve large dimensionality of feature and small sample size human brain function by near-infrared spectroscopy using an
problem, which is critical problem for standard LDA [39, 40]. adaptive general linear model," vol. 46, no. 1, pp. 133-143, 2009.
[12] J. Jin et al., "An adaptive P300-based control system," Journal of
Therefore, SRLDA is the most suitable classifier for complex neural engineering, vol. 8, no. 3, p. 036006, 2011.
MI using TDP, due to linearity and analytic shrinkage [13] H. I. Suk and S. W. Lee, "Subject and class specific frequency
regularization. bands selection for multiclass motor imagery classification,"
International Journal of Imaging Systems and Technology, vol.
Several limitations of this study should be noted. Firstly, 21, no. 2, pp. 123-130, 2011.
this research was performed in an off-line environment with [14] S.-B. Lee, H.-J. Kim, H. Kim, J.-H. Jeong, S.-W. Lee, and D.-J.
retrospectively measured data; a real-time and on-line Kim, "Comparative analysis of features extracted from EEG
environment could yield different findings. Secondly, detailed spatial, spectral and temporal domains for binary and multiclass
motor imagery classification," Information Sciences, vol. 502, pp.
statistical analyses were not available due to the small sample 190-200, 2019.
size of the experimental data, which warrants further [15] M. Mohseni, V. Shalchyan, M. Jochumsen, and I. K. Niazi,
validation with a larger sample size. Finally, TDP is the only "Upper limb complex movements decoding from pre-movement
feature utilized in this study; further comparative analyses with

2500
Authorized licensed use limited to: Korea University. Downloaded on April 09,2024 at 09:52:01 UTC from IEEE Xplore. Restrictions apply.
 EEG signals using wavelet common spatial patterns," Computer [35] C. Vidaurre, N. Krämer, B. Blankertz, and A. Schlögl, "Time
methods and programs in biomedicine, vol. 183, p. 105076, 2020. domain parameters as a feature for EEG-based brain–computer
[16] Y. Bian, H. Qi, L. Zhao, D. Ming, T. Guo, and X. Fu, interfaces," Neural Networks, vol. 22, no. 9, pp. 1313-1319, 2009.
"Improvements in event-related desynchronization and [36] A. Hapfelmeier, B. Pfahringer, and S. Kramer, "Pruning
classification performance of motor imagery using instructive incremental linear model trees with approximate lookahead,"
dynamic guidance and complex tasks," Computers in biology and IEEE Transactions on Knowledge and Data Engineering, vol. 26,
medicine, vol. 96, pp. 266-273, 2018. no. 8, pp. 2072-2076, 2013.
[17] E. Gokgoz and A. Subasi, "Comparison of decision tree [37] A. Kondacs and M. Szabó, "Long-term intra-individual variability
algorithms for EMG signal classification using DWT," of the background EEG in normals," Clinical Neurophysiology,
Biomedical Signal Processing and Control, vol. 18, pp. 138-144, vol. 110, no. 10, pp. 1708-1716, 1999.
2015. [38] C.-S. Wei, Y.-P. Lin, Y.-T. Wang, C.-T. Lin, and T.-P. Jung, "A
[18] S. Singla, S. Garsha, and S. Chatterjee, "Characterization of subject-transfer framework for obviating inter-and intra-subject
classifier performance on left and right limb motor imagery using variability in EEG-based drowsiness detection," NeuroImage, vol.
support vector machine classification of EEG signal for left and 174, pp. 407-419, 2018.
right limb movement," in 2016 5th International Conference on [39] J. Kalina and J. D. Tebbens, "Algorithms for Regularized Linear
Wireless Networks and Embedded Systems (WECON), 2016, pp. Discriminant Analysis," in BIOINFORMATICS, 2015, pp. 128-
1-4: IEEE. 133.
[19] T. Omar, Z. Wassim, and B. M. Mohamed, "Brain-computer [40] J. Schäfer and K. Strimmer, "A shrinkage approach to large-scale
interface: Frequency domain approach using the linear and the covariance matrix estimation and implications for functional
quadratic discriminant analysis," in 2014 1st International genomics," Statistical applications in genetics and molecular
Conference on Advanced Technologies for Signal and Image biology, vol. 4, no. 1, 2005.
Processing (ATSIP), 2014, pp. 346-349: IEEE.
[20] G. Sagee and S. Hema, "EEG feature extraction and classification
in multiclass multiuser motor imagery brain computer interface u
sing Bayesian Network and ANN," in 2017 International
Conference on Intelligent Computing, Instrumentation and
Control Technologies (ICICICT), 2017, pp. 938-943: IEEE.
[21] S. Bhaduri, A. Khasnobish, R. Bose, and D. Tibarewala,
"Classification of lower limb motor imagery using K Nearest
Neighbor and Naïve-Bayesian classifier," in 2016 3rd
International Conference on Recent Advances in Information
Technology (RAIT), 2016, pp. 499-504: IEEE.
[22] Y. Li, J. Wu, and J. Yang, "Developing a logistic regression model
with cross-correlation for motor imagery signal recognition," in
The 2011 IEEE/ICME International Conference on Complex
Medical Engineering, 2011, pp. 502-507: IEEE.
[23] L. Breiman, J. Friedman, R. Olshen, and C. Stone, "Classification
and regression trees, wadsworth international group, belmont, ca,
1984," Case Description Feature Subset Correct Missed FA
Misclass, vol. 1, pp. 1-3, 1993.
[24] Y. Qi, "Random forest for bioinformatics," in Ensemble machine
learning: Springer, 2012, pp. 307-323.
[25] C. J. J. D. m. Burges and k. discovery, "A tutorial on support
vector machines for pattern recognition," vol. 2, no. 2, pp. 121-
167, 1998.
[26] M. A. Oskoei and H. J. I. T. B. E. Hu, "Support vector machine-
based classification scheme for myoelectric control applied to
upper limb," vol. 55, no. 8, pp. 1956-1965, 2008.
[27] F. Lotte et al., "A review of classification algorithms for EEG-
based brain–computer interfaces: a 10 year update," vol. 15, no.
3, p. 031005, 2018.
[28] G. Bauernfeind, D. Steyrl, C. Brunner, and G. R. Müller-Putz,
"Single trial classification of fnirs-based brain-computer interface
mental arithmetic data: a comparison between different
classifiers," in 2014 36th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society, 2014, pp.
2004-2007: IEEE.
[29] T. Hastie, R. Tibshirani, and J. Friedman, The elements of
statistical learning: data mining, inference, and prediction.
Springer Science & Business Media, 2009.
[30] I. Rish, "An empirical study of the naive Bayes classifier," in
IJCAI 2001 workshop on empirical methods in artificial
intelligence, 2001, vol. 3, no. 22, pp. 41-46.
[31] D. G. Kleinbaum, K. Dietz, M. Gail, M. Klein, and M. Klein,
Logistic regression. Springer, 2002.
[32] L. E. Peterson, "K-nearest neighbor," Scholarpedia, vol. 4, no. 2,
p. 1883, 2009.
[33] J. Bergstra and Y. Bengio, "Random search for hyper-parameter
optimization," Journal of machine learning research, vol. 13, no.
Feb, pp. 281-305, 2012.
[34] D. Ouali, F. Chebana, and T. B. Ouarda, "Fully nonlinear
statistical and machine‐learning approaches for hydrological
frequency estimation at ungauged sites," Journal of Advances in
Modeling Earth Systems, vol. 9, no. 2, pp. 1292-1306, 2017.

2501
Authorized licensed use limited to: Korea University. Downloaded on April 09,2024 at 09:52:01 UTC from IEEE Xplore. Restrictions apply.