Submitted to: submitted by:

Dr. Sandeep Kumar Nikita

Department of ECE 212220005
NIT Delhi M. Tech.-ECE
2nd semester
 A review on EEG based multi-class motor imagery classification

Abstract - Classification of Electroencephalogram(EEG) plays an important role in signal acquisition equipment

of brain-computer interface research and application. motor imagery brain-computer interface (BCI) became a
promissory system that helps help people with disabilities or paralysis, who can't move their body parts like
hands, feet on their own. Brain-computer interface enables the human brain to communicate with external
devices which enable humans to move their body parts without the actual involvement of muscles and their
associated nerves. These systems use brain activity patterns that we obtained from motor imagery without
actual movement of limbs. Preprocessing techniques to increase the signal-to-noise ratio and techniques for
processing EEG data have been used. In this paper, a review on EEG-based multiclass motor imagery
classification has been presented like the combination of kullback-Leiber regularized Riemann mean (KLRRM)
and linear Support Vector Machine, Hybrid one versus one classifier, Linear discriminant analysis, Shallow
convolutional neural network.

Keywords - EEG, Brain-computer interface( BCI), Motor imagery, classification.

Introduction - Motor imagery-based Brain-computer interface is a process in which the human brain thinks or
imagines performing a movement without actual involvement of peripheral nerves and muscles and without
even tensing the muscles. MI-based BCI is an independent system with higher classification and accuracy. Brain-
computer interface translates the brain signals generated by the user's intention into a control signal or
commands and sends them to external devices such as a computer or prosthesis. Among a variety of non-
invasive BCI methods, the electroencephalogram(EEG) method is one of the best methods to record or test brain
activity because this method has short time constants, less environmental limits, and requires less expensive
equipment. Therefore, it is more convenient and practical to use EEG signals as input in BCI systems [6].

The BCI system consists of 3 components- input signal, processing unit, and command control. processing unit
takes EEG signals as an input signal. The EEG equipment work by capturing the bioelectric signals originating
from electrical activity in the brain [3]. Generally, an EEG signal has a frequency range of 0.5–100 Hz, and its
frequency range is sub-divided into delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and
gamma (30–100 Hz). It is well established that motor imagery produces an event-related synchronization(ERS)
and event-related desynchronization (ERD) over the sensorimotor areas within the mu rhythm (8– 13 Hz) and
the beta band (13–30 Hz)[17]. When a mental activity such as imagining of movement of hands or feet or tongue
are performed mu (8– 13 Hz) rhythm occurs and these rhythms are named event-related
desynchronization(ERD) and event-related synchronization(ERS). For example, when someone intends to move
or imagine the movement of the right or left hand, a short period of mu and beta ERD rhythms occurred in the
opposite hemisphere of the brain. Then a period of mu and beta rhythms ERS occurs in the opposite hemisphere
of the part of the body's imagined movement [13].

The BCI system has witnessed recent technological advancements which aim at improving the quality of life in
metro cities, elderly persons with disabilities, neuroprosthetics, and gaming.

The feature extraction technique plays an important role in obtaining better classification accuracy.
Feature extraction techniques play an important role to enhance MI signals to obtain better classification rates.
In this paper, we have presented a review on EEG-based multi-class motor imagery classification.

Methods for classification - Some of the basic methods of classification are discussed below

1.1 Linear discriminant analysis (LDA)

The Regularizing Multi-bands Common Spatial Patterns (RMCSP) method utilizes data in spectral, temporal, and
spatial domains to increase the accuracy of the classification of signals in BCI. The limitation of the BCI system is
the use of the large number of channels used as recording devices. RMCSP is designed to handle EEG signals
with a smaller number of channels.

The EEG data is decomposed by five FIR filters into five different frequency ranges. The RMCSP technique has
two phases of operation.

• In the first phase, five FIR filters have been employed that cover five different frequency ranges, using
these filters spectral features that characterize ERD/ERS events from the brain are obtained.

• The second phase involves regularizing common spatial patterns is done to learn the spatial features to
discriminate the spectral features. In this technique, the one versus rest (OVR) CSP method is employed.

Fig. 1. The architecture of Regularizing Multi-bands Common Spatial Patterns (RMCSP)

Linear discriminant analysis (LDA) is used as the classifier in which the log variances of the extracted features
from RMCSP filters were used as input to LDA. The output of the four-class classifiers LDAs are then combined by
using the method majority voting strategy i.e., the classifier that has the largest probability value has assigned a
class label.[12]
1.2 Hybrid One versus one/ one versus other (OVO)

• One versus one also called pairwise classifiers are used frequently in multi-class motor imagery
classification.

• one versus one use number of classifiers for N number class classification, each classifier
discriminates one class from the other class.

fig 2. The hybrid OVO classifier

• A Naive Bayesian classifier is designed by combined OVR extracted features because if three classes win
equally, the existence classifier can't specify the trial's label.[13]

• For four classes classification, we have to design four Naive Bayesian and six linear support vector
machine (SVM) to specify classes labels without any ambiguity.

1.3 Shallow Convolutional Neural Network (CNN) Architecture

• The proposed novel CNN architecture consists of three convolutional layers and four fully connected
layers.

• The first layer performs the convolution phase along the time axis, which is equivalent to a linear pre-
filtering of the EEG signals for each channel.

• The second layer performs convolution along the EEG channel axis, which can reduce the effect of areas
not related to motion movement.

• The next layer uses the largest pooling layer to provide a more robust architecture.

• After the convolutional layers, three fully connected(FC) layers are applied from 6300 neurons in the first
FC layer to 32. The last one is a SoftMax layer with the number of neurons in the data to classify for.[8]
 Fig 3. Novel CNN architecture

1.4 Kullback-Leibler Regularized Riemannian Mean and Linear Support Vector Machine (KLRRM and LSVM)

• Determining the Riemannianmean for all the four MI classes, we perform regularization which makes the
feature extraction robust against the outliers.

fig 4 Architecture for MI classification using KLRRM and LSVM

• LSVM is trained using the one-vs-all mechanism for the multi-class classification. Using this trained LSVM
and regularized Riemannian mean matrices, performance using the sub-testing set is analyzed.
• A similar process is repeated for all the values of β (β is the regularization factor) and the β with the
highest accuracy is considered as the optimal for the target subject. After achieving the optimal β, the
validation is performed on the validation set.[2]

DISCUSSION AND CONCLUSION

In this paper, we have reviewed various EEG-based multi-class motor imagery classifiers like LDA, LSVM, shallow
CNN, hybrid OVO.
• The limitation of conventional OVO has unfolded when more than one class is specified as the trials
label. It means that more than one class has the most and equally won in comparison to other classes.
The hybrid OVO classification system was proposed to overcome this limitation.
• The novel shallow CNN architecture was proposed to overcome the limitation of conventional CNN
architecture as the conventional architecture is suitable for 2 class classification only but the Novel CNN
architecture is suitable for four-class classification and more suitable for real-life BCI systems and better
than some conventional machine learning-based approaches.
• The limitation of the BCI system is the use of a large number of channels used as recording devices.
RMCSP is designed to handle EEG signals with a smaller number of channels. LDA shows classification
accuracy better than conventional CSP by 10%.
 • KLRRM framework provides good accuracy for poor or noisy channels which shows its robustness
towards noise and outliers while maintaining the accuracy of good subjects. The KLRRM+ LSVM
mechanism provides better performance in terms of classification accuracy for both good and poor
subjects.

REFERENCES

[1] M. E. Sharbaf, A. Fallah and S. Rashidi, "EEG-based multi-class motor imagery classification using variable-
sized filter bank and enhanced One Versus One classifier," 2017 2nd Conference on Swarm Intelligence and
Evolutionary Computation (CSIEC), 2017, pp. 135-140, doi: 10.1109/CSIEC.2017.7940174.

[2] P. K. Mishra, B. Jagadish, M. P. R. S. Kiran, P. Rajalakshmi, and D. S. Reddy, "A Novel Classification for
EEG Based Four Class Motor Imagery Using Kullback-Leibler Regularized Riemannian Manifold," 2018
IEEE 20th International Conference on e-Health Networking, Applications, and Services (Healthcom),
2018, pp. 1-5, doi:10.1109/HealthCom.2018.8531086.
[3] Safitri, E. C. Djamal and F. Nugraha, "Brain-Computer Interface of Motor Imagery Using ICA and Recurrent
Neural Networks," 2020 3rd International Conference on Computer and Informatics Engineering (IC2IE), 2020,
pp. 118-122,doi:10.1109/IC2IE50715.2020.9274681.
[4] E. Dong, G. Zhu and C. Chen, "Classification of four categories of EEG signals based on relevance vector
machine," 2017 IEEE International Conference on Mechatronics and Automation (ICMA), 2017, pp. 1024-1029,
doi: 10.1109/ICMA.2017.8015957.

[5] L. Wang and X. Wu, "Classification of Four-Class Motor Imagery EEG Data Using Spatial Filtering," 2008 2nd
International Conference on Bioinformatics and Biomedical Engineering, 2008, pp. 2153-2156, doi:
10.1109/ICBBE.2008.868.

[6] Y. Wang, S. Gao and X. Gao, "Common Spatial Pattern Method for Channel Selelction in Motor Imagery Based
Brain-computer Interface," 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference, 2005, pp.
5392-5395, doi: 10.1109/IEMBS.2005.1615701.

[7] Z. Y. Chin, K. K. Ang, C. Guan, C. Wang and H. Zhang, "Filter Bank Feature Combination (FBFC) approach for
brain-computer interface," The 2011 International Joint Conference on Neural Networks, 2011, pp. 1352-1357,
doi: 10.1109/IJCNN.2011.6033381.

[8] B. Du, Y. Liu and G. Tian, "Improving Motor Imagery EEG Classification by CNN with Data Augmentation," 2020
IEEE 19th International Conference on Cognitive Informatics & Cognitive Computing (ICCI*CC), 2020, pp. 111-118,
doi: 10.1109/ICCICC50026.2020.9450227.

[9] Le Quoc Thang and C. Temiyasathit, "Increase performance of four-class classification for motor-imagery based
brain-computer interface," 2014 International Conference on Computer, Information and Telecommunication
Systems (CITS), 2014, pp. 1-5, doi: 10.1109/CITS.2014.6878959.
[10] Z. Yang, S. Guo and Y. Hong, "Motor Imagery Multi-classification based on Canonical Correlation Analysis
Feature Fusion," 2021 40th Chinese Control Conference (CCC), 2021, pp. 3334-3339, doi:
10.23919/CCC52363.2021.9549799.

[11] A. Barachant, S. Bonnet, M. Congedo and C. Jutten, "Multiclass Brain–Computer Interface Classification by
Riemannian Geometry," in IEEE Transactions on Biomedical Engineering, vol. 59, no. 4, pp. 920-928, April 2012,
doi: 10.1109/TBME.2011.2172210.

[12] L. Q. Thang and C. Temiyasathit, "Regularizing multi-bands Common Spatial Patterns (RMCSP): A data
processing method for brain-computer interface," 2015 9th International Symposium on Medical Information and
Communication Technology (ISMICT), 2015, pp. 180-184, doi: 10.1109/ISMICT.2015.7107524.

[13] M. E. Sharbaf, A. Fallah and S. Rashidi, "Shrinkage estimator based common spatial pattern for multi-class
motor imagery classification by hybrid classifier," 2017 3rd International Conference on Pattern Recognition and
Image Analysis (IPRIA), 2017, pp. 26-31, doi: 10.1109/PRIA.2017.7983059.

[14] L. Wang and X. Wu, "Classification of Four-Class Motor Imagery EEG Data Using Spatial Filtering," 2008 2nd
International Conference on Bioinformatics and Biomedical Engineering, 2008, pp. 2153-2156, doi:
10.1109/ICBBE.2008.868.

[15] J. Mourino, J. del R Millan, F. Cincotti, S. Chiappa, R. Jane and F. Babiloni, "Spatial filtering in the training
process of a brain computer interface," 2001 Conference Proceedings of the 23rd Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, 2001, pp. 639-642 vol.1, doi:
10.1109/IEMBS.2001.1019016.

[16] X. Bai, X. Wang, S. Zheng and M. Yu, "The offline feature extraction of four-class motor imagery EEG based on
ICA and Wavelet-CSP," Proceedings of the 33rd Chinese Control Conference, 2014, pp. 7189-7194, doi:
10.1109/ChiCC.2014.6896188.

[17] Mahamune R, Laskar SH. Classification of the four-class motor imagery signals using continuous wavelet
transform filter bank-based two-dimensional images. Int J Imaging Syst Technol. 2021;1–12.
https://doi.org/10.1002/ima.22593

[18] M. E. Sharbaf, A. Fallah and S. Rashidi, "Shrinkage estimator based common spatial pattern for multi-class
motor imagery classification by hybrid classifier," 2017 3rd International Conference on Pattern Recognition and
Image Analysis (IPRIA), 2017, pp. 26-31, doi: 10.1109/PRIA.2017.7983059.