Multilayer extreme learning machine for hand movement prediction based on electroencephalography

Bulletin of Electrical Engineering and Informatics
Vol. 9, No. 6, December 2020, pp. 2404~2410
ISSN: 2302-9285, DOI: 10.11591/eei.v9i6.2626  2404
Journal homepage: http://beei.org
Multilayer extreme learning machine for hand movement
prediction based on electroencephalography
Khairul Anam1
, Cries Avian2
, Muhammad Nuh3
1,2
Department of Electrical Engineering, Universitas Jember, Indonesia
3
Department of Biomedical Engineering, Institut Teknologi Sepuluh November, Indonesia
Article Info ABSTRACT
Article history:
Received Dec 10, 2019
Revised Mar 29, 2020
Accepted May 3, 2020
Brain computer interface (BCI) technology connects humans with machines
via electroencephalography (EEG). The mechanism of BCI is pattern recognition,
which proceeds by feature extraction and classification. Various feature
extraction and classification methods can differentiate human motor
movements, especially those of the hand. Combinations of these methods
can greatly improve the accuracy of the results. This article explores
the performances of nine feature-extraction types computed by a multilayer
extreme learning machine (ML-ELM). The proposed method was tested on
different numbers of EEG channels and different ML-ELM structures.
Moreover, the performance of ML-ELM was compared with those of ELM,
Support Vector Machine and Naive Bayes in classifying real and imaginary
hand movements in offline mode. The ML-ELM with discrete wavelet
transform (DWT) as feature extraction outperformed the other classification
methods with highest accuracy 0.98. So, the authors also found that
the structures influenced the accuracy of ML-ELM for different task,
feature extraction used and channel used.
Keywords:
Classification methods
Electroencephalography
Extreme learning machine
Movement prediction
Multilayer ELM
This is an open access article under the CC BY-SA license.
Corresponding Author:
Khairul Anam,
Department of Electrical Engineering,
Universitas Jember,
Jl. Kalimantan No.37, Sumbersari, Jember, East Java, Indonesia.
Email: khairul@unej.ac.id
1. INTRODUCTION
Technological advances, especially in Brain–Computer Interfaces (BCIs), play a crucial role in
assisting patients with physical disabilities or paralysis [1]. BCI detects the brain activity in the cortex as
electroencephalography (EEG) signals. One of the most popular BCI applications is classifying human motor
movements, especially hand movements, for various assistive robots [2, 3]. Hand movements can be
recognized and classified by various recognition/classification methods [4, 5] that distinguish between right-
and left-hand movements in real-time [6, 7] or in imagery [8, 9]. The pattern recognition method
must be optimized before implementing the EEG signal on an assistive robot that controls the robot
hand/robot therapy.
In EEG signal processing, the signal information is first extracted and then classified [9]. Among the most
popular classification methods are the support vector machine (SVM) [10, 11], linear discriminant analysis [12, 13],
artificial neural networks [14, 15], and fuzzy algorithms [16, 17]. The accuracy obtained in each
classification method also varies in each study, like the results obtained by Farooq et al. [18]. In this research,
the classification method used the principal component analysis (PCA) and independent component analysis
(ICA) as feature extraction with the K nearest neighbor (KNN) as a classifier which obtained an accuracy

Bulletin of Electr Eng & Inf ISSN: 2302-9285 
Multilayer extreme learning machine for hand movement prediction based on… (Khairul Anam)
2405
rate of 80.6%. Besides that, Kevric et al. [19] also use KNN as its classifier with discrete wavelet transform
(DWT) as feature extraction. From this study, the accuracy of the system reaches 92.8%. Deep learning was
also used in this case, one of which was a study conducted by Korhan et al. [20] who tried to solve problems
using common spatial filter (CSP) and convolutional neural network (CNN). In this study, the system reaches
an accuracy of 93.75%. However, the existing classification methods often struggle to meet the following
requirements: (1) ease of use, (2) low computational complexity, and (3) high accuracy [1, 21].
One classification methods that can achieve these criteria is the extreme learning machine (ELM) [22].
The ELM was applied to EEG classification by Tan et al. [23]. They reported that ELM delivers
excellent performance compared to other classification methods. However, other studies have reported poor
performance stability in ELM. The poor performance is caused by the randomness of assigning weights (w)
to the neurons and the threshold between the input layer and the hidden layer [24]. These problems degrade
the accuracy of ELM classification. However, the problems found in ELM have been resolved by
the existence of a multilayer extreme learning machine (ML-ELM). With its multi-layered structure,
ML-ELM overcomes the problems found in ELM. In particular, the data are transformed in each layer,
accelerating the convergence and minimizing the errors [24]. Research conducted by Duan et al. [25] has
proven the superiority of this ML-ELM. In that study, researchers used the PCA and linear discriminant
analysis (LDA) as feature extraction with the ML-ELM as the classifier. The results obtained indicate that the
accuracy rate of this system is 94.2%. But unfortunately, Duan et al. [25] only explored the use of
ML-ELM in the extraction feature, even though many feature extractions can be used to improve
the accuracy of the classification of EEG signals such as power spectral density (PSD) and also DWT.
Therefore, to this end, the authors explore various feature extraction methods employed in previous studies,
and integrate them into a multilayer (ML) ELM classifier.
In this exploration study, the performance of hand classification will be improved by using several
step. First, the ML-ELM was paired with nine feature extractions that are popularly used in EEG.
After selecting the best feature extraction–classifier pairs, we tracked the performances of the optimized
ML-ELM classifier with different numbers of channels and different structures. To demonstrate the superior
performance of ML-ELM, the classifier was competed against the ELM, SVM and Naive Bayes (NB)
classifiers. Each of the three classifiers performed nine types of feature extractions. The nine-feature
extraction was selected because in our previous research, the features used were limited to power
features [23]. Feature extraction in EEG has many variations [26] which can be combined with ML-ELM.
This opportunity must be exploited because different feature extraction and classification methods can pool
their specific advantages to improve the accuracy of the result. The accuracy performances of
the classification methods were evaluated by the K-fold technique. The ML-ELM outperformed the other
classifiers. When paired with discrete wavelet density, ML-ELM delivered its highest performance, with the
highest classification accuracy of 0.98 (98%).
2. RESEARCH METHOD
2.1. Multilayer extreme learning machine
As explained in the previous section, ML-ELM is ELM with more than one hidden layer [27]. The last
hidden layer generally processes the ELM algorithm. To qualify as ML-ELM, some articles require the ELM
to possess more than two hidden layers [24]. ML-ELM differs from ELM in fundamental ways. ML-ELM is
based on the ELM auto-encoder [28], and is algorithmically similar to deep learning. The ELM auto-encoder
operates on the weights (w) of the hidden layers during training with unsupervised learning. However, unlike
deep learning, there is no fine-tuning in ML-ELM. The activation function in ML-ELM can be linear or
nonlinear. If the number of nodes Lk in the k-th hidden layer equals the number of nodes Lk-1 in the (k-1)-th
hidden layer, g is a linear function; otherwise, g is a nonlinear connection such as a sigmoidal function.
The output matrix Hk
of hidden layer k is given by:
) ) (1)
The input layer x can be considered as the 0-th hidden layer (i.e., k = 0). The output of the connection between
the last hidden layer and the output node t is analytically calculated by the regularized least-squares method.
2.2. Dataset and system construction
The data in this research were provided by Hohyun Cho from the School of Electrical Engineering
and Computer Science, Gwangju Institute of Science and Technology, South Korea [29]. The dataset consists
of 52 data with two types of hand movements: real right- and left-hand movements, and imaginary right- and
left-hand movements. The dataset contains 64 channels of EEG data. In this study, the EEG signals in this

 ISSN: 2302-9285
Bulletin of Electr Eng & Inf, Vol. 9, No. 6, December 2020 : 2404 – 2410
2406
dataset were extracted by several signal processing methods. The methods were evaluated on nine popular
feature extraction types used in existing studies as shown in Table 1.
Table 1. The nine feature-extraction types employed in the present study
No Feature Extraction Equation Information
1 Bandpower [30]
∫ )
idxDelta is the power at a specific
frequency (bandpass).
2 Hurst Coeffecient [31] ) ) T is the sampling period, R/S is
the rescaled range value.
3 Detrended Fluctuation
Analysis (DFA) [32] ) √ ∑ ) ))
X(k) is the EEG in the time series
data, ) is the local trend.
4 Hjorth [31] Activity =
Mobility = √ ) )
Complexity =
is the average signal power or
variance. and are variants of
the first and second signal
derivatives, respectively.
5 Power Spectral Density
(PSD) [30] ̂ ) |∑ ) ) |
∑ | )|
̂ ) ∑ ̂ )
{xd(n)} is a sequence with d =
1,2,3…L signal intervals, each of
length M. U stands for
the normalization factor of power
in the window function. w(n)
denotes the windowed data.
6 Mean Absolute Value
(MAV) [33] ∑| |
x is a signal in a large sample of
size N, i is the sequence index.
7 Standard Deviation [33]
√
∑ ̅ )
̅ is the average of x.
8 Petrosian Fractal
Dimension (PFD) [31] ( ( ))
K is the number of samples in the
signal. denotes the number of
signal changes in the signal
derivative.
9 Discrete Wavelet
Transform (DWT) [34]
[ ] [ ] ∑ [ ] [ ]
[ ] [ ] ∑ [ ] [ ]
[ ] is a discrete-time signal,
[ ] is first passed through a half-
band high-pass filter and then
through h[ ] in the low-pass filter.
d1 and a1 are level-1
approximation coefficients.
[ ] and [ ] are the
outputs of the high-pass and low-
pass filters, respectively, after
subsampling.
Prior to extraction, the EEG signal was preprocessed through a bandpass filter operated at 8-14 Hz
with a sampling frequency of 512 Hz. For the present study, the authors chose the data on subjects 3, 5, 9, 14,
15, 35, 41, 43, 46, and 52 from the provided dataset. The preprocessed and extracted EEG signals were then
classified by various types of extraction methods applied in the literature. The classification methods
are required to discriminate between right- and left-hand movements. This study compared the results of
three classification methods: ELM, ML-ELM, SVM and NB. Figure 1 is a flowchart of the work process.
The proposed system is implemented in Python 3.6.5.

2407
Figure 1. Methods adopted in the present research
2.3. Testing methods and parameters
The method was tested for different numbers of channels and different structures of the ML-ELM.
The testing methods are detailed by:
Testing Method 1: The dataset was classified by pairs of classification methods. Altogether, there were 36
pairs of classification methods (nine feature extractions performed by four classifiers). Each pair was
required to classify the five selected dataset subjects on 3, 5, 9, 15, 41, or 64 channels. The details of
the channels were as follows: 3 channels: C3, Cz, C4; 5 channels: C3, C1, Cz, C2, C4; 9 channels: C3, C2,
CZ, C2, C4, FZ, FCz, CP2, Pz; 15 channels: FC3, FC1, FCz, FC2, FC4, C3, C1, Cz, C2, C4, CP3, CP1, CPZ,
CP2, CP4; 41 channels: AF3,AFZ,AF4,F5,FC3,FC1, FCz, FC2, FC4, FC6, C5, C3, C1, Cz, C2, C4, C5, CP5,
CP3, CP1, CPZ, CP2, CP4, CP6, P5, P3, P1, Pz, P2, P4, P6, PO3, POz, PO4; and 64 channels: all channels in
the dataset. The number of neurons was 1000 in the ELM classifier and 1000 in each layer (five layers) of
the ML-ELM classifier.
Testing Method 2: Using the two feature extractions yielding the highest accuracy and the two number of
the optimal channels, the system was tested for different structures of the ML-ELM. From an initial sequence
of two layers, the ML-ELM was built to a maximum of eight hidden layers with a maximum neuron number
of 100 per layer. Before adding a new hidden layer, each neuron in the present layer was increased by ten
neurons. All tests were performed on the ten subjects selected from the dataset. The results in the following
section report the average accuracy over all subjects and all channel used, as obtained in the 5-fold evaluation.
3. RESULTS AND DISCUSSION
This section reports and discusses the results under the test scenarios described in the previous section.
3.1. Testing method 1
The results of the ML-ELM, ELM, SVM and NB methods are shown in Figures 2, 3, and 4,
respectively. Figures 2 and 3 are the performance data for each pair of classification methods. The results
contained in the Figures are the result of the accumulation of all subjects used and the entire channel used.
The accumulated results are then averaged based on feature extraction and classifier. After that, the results
are displayed in a graph contained in Figures 2 and 3. The results obtained indicate that PSD and DWT have
higher performance compared to other extraction features. This claim is found in the ELM and ML-ELM
classifier types. In the Movement type, PSD has the best accuracy on the combination with ML-ELM 0.92.
As for DWT, the performance is owned by ELM with an accuracy of 0.95.
For imagery types, ML-ELM has the best accuracy on PSD with an accuracy of 0.93. In DWT,
ML-ELM also has the best performance with a value of 0.96, while ELM has an accuracy of 0.958. Related
to the performance of each classifier, then analyzing of head-to-head comparison data is needed for each task.
It can be known by adding up the overall accuracy contained in the nine extraction features used. The best
performance data can be seen in Figure 4. The results show that ML-ELM has the best performance
compared to the other three classifiers. The total accuracy obtained by ML-ELM on the nine feature
extractions are 5.47 for movement task and 5.28 for imagery task. Even so, ELM has almost the same
performance as ML-ELM, with a value of 5.40 for imagery and 5.25 for movement.

 ISSN: 2302-9285
2408
Figure 2. Performance results of each clasification
method pairs in movement task
Figure 3. Performance results of each clasification
method pairs in imagery task
Figure 4. Total accuracy of classifier performace in each task
3.2. Testing method 2
This test was performed for different structures of ML-ELM. The test was intended to determine
the effect of changing the number of hidden layers and neurons on the accuracy results. The results
are presented in Figures 5 and 6. Labels such as "PSD-I-15" describe the feature extraction type, type of
movement (“I” or “M”)-number of channels (15 or 41). The numbers along the x-axis are the combined
numbers of hidden layers and neurons. The number of tens symbolizes the number of hidden layers and
the unit number symbolizes the number of neurons. For example, 23 denotes three hidden layers with 1000
neurons in that layer and 30 neurons in the fourth layer. The selected numbers of channels (15 and 41)
provided the highest accuracy in the previous test. The results of the present test are the average accuracies
over the ten selected subjects.
Figure 5. DWT performance in structure change of
ML-ELM
Figure 6. PSD performance in structure change of
ML-ELM

2409
From the results obtained indicate that the ML-ELM structure has a significant effect on the level of
accuracy obtained. This proves that there is a need for tuning the ML-ELM structure in different types of
tasks, feature extraction and the number of channels to get the best performance. For example in the task
movement data with DWT extraction feature using 41 channels, it is found that the average of the accuracy
of 10 subjects, the best performance is located in the 7th ML-ELM structure or with the structure [1000, 700]
which means 1000 is the first number of neurons and 700 for neurons second. Whereas in other structures, it
has fluctuating accuracy. The lowest value obtained is in the 49th structure (1000 neurons for layers 1 to 5
and 900 neurons for layer 6) with an accuracy of 0.9. Whereas the type of task imagery with DWT extraction
feature uses 41 channels, the best structure is in structure number 17 (1000 neurons for layers 1 to 2 and 700
neurons in layer 3) with an accuracy of 0.984. When viewed in the use of a different number of channels,
the best structure that is produced is also different. For example, on the use of channel number 15, DWT on
imagery task has the best performance on structure number 6, and on the movement task has the best
performance on structure number 19. This is also experienced on PSD feature extraction both on using
the different number of channels and various tasks. However, when compared, DWT still has the best
performance compared to PSD. When examined, the phenomenon that occurs in the form of fluctuating
accuracy values is caused by underfitting and overfitting problems. This can be observed from Figure 5
and Figure 6. If with the addition of structure, the accuracy rate is getting better, then the system
is underfitting. Conversely, if the addition of structure precisely decreases the accuracy level, it means the system
is overfitting.
4. CONCLUSION
The results confirmed the superior classification performance of multilayer extreme learning
machine (ML-ELM) over ELM, Support Vector Machine and Naive Bayes on almost every feature-
extraction evaluation. Moreover, ML-ELM delivered its best results with DWT as feature extraction.
The accuracy of ML-ELM depended on the number of channels used, the feature extraction, and the task.
However, the accuracy was sometimes degraded by underfitting and overfitting problems, causing accuracy
fluctuations in the results of different structure change. In this article, the combinations of feature extractions
and classification methods were somewhat limited. In future research, the underfitting and overfitting
problem will be resolved by combining ML-ELM with optimization algorithms. The second line of research
is related to combinations of feature extractions. Instead of a single feature extraction, ML-ELM can be
combined with several feature extractions. This is expected to improve the accuracy by increasing the ease of
differentiating the data. Finally, one could apply additional filters.
ACKNOWLEDGEMENTS
We would like to thank the Directorate Research and Development, The Ministry of Research
and Higher Education, the Republic of Indonesia for research funds under Post-doctoral Research Scheme
Contract no. 175/SP2H/LT/DRPM/2019.
REFERENCES
[1] L. Bi, X. Fan and Y. Liu, "EEG-Based Brain-Controlled Mobile Robots: A Survey," in IEEE Transactions on
Human-Machine Systems, vol. 43, no. 2, pp. 161-176, March 2013.
[2] O. Rahma, R. Hendradi and F. Ama, "Classifying imaginary hand movement through electroencephalograph signal for
neuro-rehabilitation," Walailak J Sci Technol, vol. 16, no. 12, pp. 943-953, 2019.
[3] S. A. Dewangga, H. Tjandrasa, D. Herumurti, "Robot motion control using the emotiv EPOC EEG system," Bull
Electr Eng Inf, vol. 7, no. 2, pp. 279-285, 2018.
[4] R. Martín-Clemente, J. Olias, D. B. Thiyam, A. Cichocki and S. Cruces, "Information theoretic approaches for motor-imagery
BCI systems: review and experimental comparison," Entropy, vol. 20, no. 1, pp. 7:1-29, 2018.
[5] N. E. Md Isa, A. Amir, M. Z. Ilyas and M. S. Razalli, "Motor imagery classification in brain computer interface
(BCI) based on EEG signal by using machine learning technique," Bull Electr Eng Inf, vol. 8, no. 1, pp. 269-275, 2019.
[6] A. Dey, S. Bhattacharjee and D. Samanta, "Recognition of motor imagery left and right hand movement using
EEG," 2016 IEEE International Conference on Recent Trends in Electronics, Information & Communication
Technology (RTEICT), Bangalore, pp. 426-430, 2016.
[7] K. P. Thomas, N. Robinson, K. G. Smitha and A. P. Vinod, "EEG-based Discriminative Features During Hand
Movement Execution and Imagination," 2018 15th International Conference on Control, Automation, Robotics and
Vision (ICARCV), Singapore, pp. 883-888, 2018.
[8] S. A. El-aal, R. A. Ramadan and N. I. Ghali, "EEG signals of motor imagery classification using adaptive neuro-
fuzzy inference system," Advances in Nature and Biologically Inspired Computing, vol. 419, pp. 105-116, 2016.

 ISSN: 2302-9285
2410
[9] A. Schwarz, P. Ofner, J. Pereira, A. I. Sburlea and G. R. Müller-Putz, "Decoding natural reach and grasp actions
from human EEG," J Neural Eng, vol. 15, no. 1, pp. 1-23, 2017
[10] X. Tang, J. Yang and H. Wan, "A hybrid SAE and CNN classifier for motor imagery EEG classification," In
Artificial Intelligence and Algorithms in Intelligent Systems, Proceedings of 7th Computer Science On-line
Conference, vol 2, pp. 265-278, 2018.
[11] J. Moreira, M. Moreira, N. Pombo, B. M. C. Silva and N. M. Garcia, "Identification of real and imaginary
movements in EEG using machine learning models," In CMBEBIH: International Conference on Medical and
Biological Engineering, Banja Luka, pp. 469-474, 2019.
[12] N. G. Ozmen, L. Gumusel and Y. Yang, "A biologically inspired approach to frequency domain feature extraction
for EEG classification," Comput Math Methods Med, vol. 2018, pp. 19-24, 2018.
[13] S. Razi, M. R. K. Mollaei and J. Ghasemi, "A novel method for classification of BCI multi-class motor imagery
task based on Dempster-Shafer theory," Inf Sci, vol. 484, pp. 14-26, 2019.
[14] A. Javed, Mohsin I. Tiwana, Moazzam I. Tiwana, N. Rashid, J. Iqbal, and U. S. Khan, "Recognition of finger
movements using EEG signals for control of upper limb prosthesis using logistic regression," Biomed Res, vol. 28,
no. 17, pp. 7361-7369, 2017.
[15] S. Bhattacharya, K. Bhimraj, R. J. Haddad and M. Ahad, "Optimization of EEG-based imaginary motion
classification using majority-voting," SoutheastCon 2017, Charlotte, NC, pp. 1-5, 2017,
[16] T. Nguyen, I. Hettiarachchi, A. Khosravi, S. M. Salaken, A. Bhatti and S. Nahavandi, "Multiclass EEG data
classification using fuzzy systems," 2017 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE), Naples,
pp. 1-6, 2017.
[17] R. Chatterjee, T. Bandyopadhyay, D. K. Sanyal and D. Guha, "Comparative analysis of feature extraction
techniques in motor imagery EEG signal classification," In: Somani A., Srivastava S., Mundra A., Rawat S. (eds)
Proceedings of First International Conference on Smart System, Innovations and Computing. Smart Innovation,
Systems and Technologies, vol. 79, pp. 73-83, 2018.
[18] E. J. Rechy-Ramirez and H. Hu, "Bio-signal based control in assistive robots: a survey," Digit Commun Netw,
vol. 1, no. 2, pp. 85-101, 2015.
[19] C. Chen, K. Li, M. Duan and K. Li, "Extreme learning machine and its applications in big data processing," in Big
Data Analytics for Sensor-Network Collected Intelligence: Intelligent Data-Centric Systems, Elsevier, 2017.
[20] P. Tan, W. Sa and L. Yu, "Applying Extreme Learning Machine to classification of EEG BCI," 2016 IEEE
International Conference on Cyber Technology in Automation, Control, and Intelligent Systems (CYBER),
Chengdu, pp. 228-232, 2016.
[21] D. Xiao, B. Li and Y. Mao, "A multiple hidden layers extreme learning machine method and its application," Math
Probl Eng, pp. 1-10, 2017.
[22] A. R. Mane, S. D. Biradar and R. K. Shastri, "Review paper on Feature Extraction Methods for EEG Signal
Analysis," Int J Emerg Trend Eng Basic Sci, vol. 2, no. 1, pp. 545-552, 2015.
[23] R. M. Parkavi, M. Shanthi and M. C. Bhuvaneshwari, "Recent trends in ELM and MLELM: a review," Adv Sci
Technol Eng Syst J, vol. 2, no. 1, pp. 69-75, 2017.
[24] Kasun LLC, Zhou H, Huang G-B et al, "Representational learning with ELMs for big data. IEEE Intell Syst, vol. 4,
pp. 1–4, 2013.
[25] H. Cho, M. Ahn, S. Ahn, M. Kwon and S. C. Jun, "EEG datasets for motor imagery brain computer interface,"
GigaScience, vol. 6, no. 7, pp. 1-7, 2017,
[26] S. Sarraf, "EEG-Based Movement Imagery Classification Using Machine Learning Techniques and Welch’s Power Spectral
Density Estimation," Am Sci. Res. J Eng Technol Sci, vol. 33, pp. no. 1, 124-145, 2017.
[27] M. Peker, "An efficient sleep scoring system based on EEG signal using complex-valued machine learning
algorithms," Neurocomputing, vol. 207, pp. 165-177, 2016.
[28] A. Adda and H. Benoudnine, "Detrended fluctuation analysis of EEG recordings for epileptic seizure detection,"
2016 International Conference on Bio-engineering for Smart Technologies (BioSMART), Dubai, pp. 1-4, 2016.
[29] A. Khorshidtalab, M. J. E. Salami and M. Hamedi Khorshidtalab A, Salami MJE, Hamedi M, "Robust classification
of motor imagery EEG signals using statistical time-domain features," Physiol Meas, 34, no. 11, pp. 1563-1579, 2013.
[30] P. Ghorbanian, D. M. Devilbiss, A. J. Simon, A. Bernstein, T. Hess and H. Ashrafiuon, "Discrete wavelet transform EEG
features of Alzheimer'S disease in activated states," 2012 Annual International Conference of the IEEE Engineering in
Medicine and Biology Society, San Diego, CA, 2012, pp. 2937-2940.

Multilayer extreme learning machine for hand movement prediction based on electroencephalography

More Related Content

What's hot (18)

Similar to Multilayer extreme learning machine for hand movement prediction based on electroencephalography (20)

More from journalBEEI (20)

Recently uploaded (20)

Multilayer extreme learning machine for hand movement prediction based on electroencephalography