Page 56 - IJEEE-2023-Vol19-ISSUE-1

52 |
IV. RESULTS AND DISCUSSION
It was mentioned that this study uses CWT to transform
the given MI EEG signals into scalograms (time-frequency
images). Different mother wavelet functions produce
different time-frequency characteristics. To examine those
differences, three types of mother wavelets namely Morlet,
Bump, and Amor were used in this study.
Also, the hybrid feature extraction and classification
model has been achieved by using the VGG-16 as a features
extractor with one classifier. Six classification algorithms
have been examined namely neural network (NN), K-nearest
neighbors (KNN), Naïve Bayes (NB), logistic regression
(LR), SVM, and decision tree (DT).
The six classifiers are experimented with each of the
mother wavelet functions to find the best hybrid model over
the combined subject dataset of (BCI Competition dataset 2b
, the training part). The results of this experiment are
presented as confusion matrices in Fig. 4. A confusion matrix
is presented for each classifier and mother wavelet function.
Fig. 5 shows the region of convergence (ROC) for each one
can see that NN based hybrid model outperformed the
performance of other classification methods with the three
mother wavelet functions.
It can be noticed that CNN+NN has achieved the best
results in comparison to the other experimented hybrid
models. The confusion matrix shows 100% classification
accuracy with Amor and Morlet wavelet mother functions.
For this optimal model, the training time, testing time, and
log loss are computed for the three mother functions. As
shown in Fig. 6 and overall, the WTNN model with Amor
mother functions has delivered the best classification results.
The second experiment is depicted for evaluating the
proposed model with another different MI EEG dataset
which is (dataset II training part, but for individual subject) .
This helps overcome the inter-subject classification problem.
To evaluate the optimal model (the WTNN) over different
brain signal complexities to overcome the problem of inter-
subjects, the model was tested in experiment-2 with another
dataset (dataset II evaluation part, individual subjects) that
consists of nine subjects. The result showed that the
developed model attained 99% of mean accuracy over the
nine subjects as presented in TABLE II. Additionally, to
evaluate the performance of the WTNN in the ability to
overcome intra- subjects' brain signal challenges of sessions
(in this study we consider the problem of with feedback and
without feedback recording protocols of two sessions). The
result showed that the WTNN model attained 99% of mean
accuracy over the nine subjects as presented in TABLE III.
Comparing the result of this study with academic literature
over dataset-I and dataset-II, it is clear that our WTNN model
outperformed the accuracy of the literature as presented in
TABLE IV and TABLE IV. This comparative result
appraises the efficiency of the proposed WTNN model due
to the capability of VGG-16 and Amor wavelet in extracting
MI signal features. And also, the efficiency of the hybrid
model in decoding the right and left commands. This model
will contribute to the BCI community by facilitating the
Al-Qaysi1, Al-Saegh, Hussein & Ahmed
deployment of the proposed framework model in MI-based
BCI applications.
V. CONCLUSION
Deep learning and wavelet transformation are useful
techniques for dealing with the high-dimensional and
nonstationary MI EEG signals. This paper studied the use of
deep learning, three different wavelet mother functions, and
six different classifiers for the analysis of MI EEG signals.
The performance of different combinations of mother
functions and classifiers are compared on two MI EEG
datasets. Several evaluation metrics show that a model of
VGG-16 feature extractor with a neural network classifier
using the Amor mother wavelet function has outperformed
the results of state-of-the-art studies by achieving 99% of
classification accuracy on dataset-II and 100% accuracy on
dataset I. This result will facilitate the deployment of an
accurate model based on our technique to help the
community of the BCI users.
CONFLICT OF INTEREST
The authors have no conflict of relevant interest to this
article.
REFERENCES
[1] A. Al-Saegh, et al., "Deep learning for motor imagery
EEG-based classification: A review," Biomedical Signal
Processing and Control, vol. 63, p. 102172, 2021.
[2] Z. Al-Qaysi, et al., "A review of disability EEG based
wheelchair control system: Coherent taxonomy, open
challenges and recommendations," Computer methods and
programs in biomedicine, vol. 164, pp. 221-237, 2018.
[3] S. Sreeja and D. Samanta, "Classification of multiclass
motor imagery EEG signal using sparsity approach,"
Neurocomputing, vol. 368, pp. 133-145, 2019.
[4] S. U. Amin, et al., "Deep Learning for EEG motor
imagery classification based on multi-layer CNNs feature
fusion," Future Generation computer systems, vol. 101,
pp. 542-554, 2019.
[5] S. Aggarwal and N. Chugh, "Signal processing
techniques for motor imagery brain computer interface: A
review," Array, vol. 1, p. 100003, 2019.
[6] X. Tang, et al., "Motor imagery EEG recognition with
KNN-based smooth auto-encoder," Artificial intelligence
in medicine, vol. 101, p. 101747, 2019.
[7] A. Al-Saegh, et al., "CutCat: An augmentation method
for EEG classification," Neural Networks, vol. 141, pp.
433-443, 2021.
[8] C. H. Nguyen and P. Artemiadis, "EEG feature
descriptors and discriminant analysis under Riemannian
Manifold perspective," Neurocomputing, vol. 275, pp.
1871-1883, 2018.
[9] W. Qiao and X. Bi, "Ternary-task convolutional
bidirectional neural turing machine for assessment of
EEG-based cognitive workload," Biomedical Signal
Processing and Control, vol. 57, p. 101745, 2020.