Page 53 - IJEEE-2023-Vol19-ISSUE-1

Al-Qaysi1, Al-Saegh, Hussein & Ahmed
classification accuracy of 81.99% and 77.22% on BCIC IV
2a and BCIC IV 2b datasets respectively.
Xu et. al. [23] proposed a method involving extracting
features based on Hjorth parameter, the power spectrum
estimation, and time-frequency energy. Then sparse
representation was used to acquire lower-dimensional
informative features while keeping the discriminative ability as the best classification accuracy that is retained by the
among the different patterns. They obtained 79% of
classification accuracy on BCIC IV 2b. Oh et. al. [24] used
Hjorth parameter and Fisher ratio to find the dominant
frequency bands and the timing in training EEG signals and
79.1% of classification accuracy was achieved on BCIC IV
2b.
Bagh and Reddy [25] used the Hilbert transform (HT) for
the recognition of Event-related potentials, and the SVM for
decoding the MI signals. They obtained 86.11% and 82.50%
of classification accuracy on BCIC III 3a and BCIC IV 2b
respectively.
Zhu et. al. [26] The multi-channel input is encoded using
a separated channel convolutional network, and the encoded
features are concatenated and fed into a recognition network,
which performs the final MI task recognition. They obtained
83% of classification accuracy on BCIC IV 2b.
Dai et. al. [27] proposed a segmentation data
augmentation method for MI EEG signals and used the new
trails for training a CNN. They obtained 91.57% and 83% of
classification accuracy on BCIC IV 2a and BCIC IV 2b
respectively.
Kim et. al. [28] proposed a method that rely on the power such, it was used in developing and validating the WTNN to
spectral density (PSD) to find the noon-stationarity feature
of each couple of EEG channels by calculating a matrix for
that. Such that, they exploited the time, frequency, and
spatial characteristics of the time-series signals. They
obtained 89.36% of classification accuracy on BCIC IV 2a
for only two classes classification.
Sun et. al. [29] used an EOG channel for retaining the
potential related to MI tasks with the combination of Hjorth
algorithm to train their model. They obtained 76.45% and
Fig. 1: Methodological Framework for WTNN model
| 49
0.79% of classification accuracy on BCIC IV 2a and BCIC
IV 2b respectively.
Kant et. al. [30] combined the continuous wavelet
transform (CWT) with deep learning-based transfer learning
and compared this method to different well-known deep
CNNs using the BCIC III 3a dataset. They obtained 95.71%
VGG-19. Although there are several studies tried to improve
the accuracy of classification, the results of the reviewed
studies show that their still an area for enhancement.
III. METHODOLOGY
The methodological framework of the WTNN for the
two-class MI EEG classification problem is presented in Fig.
1. This framework describes the whole process of pattern
recognition starting from preparing the training samples and
ending with the performance evaluation stage. The following
subsections give more details concerning the methodology
of this research.
A. MI EEG Datasets
A minimum number of channels is normally preferred
by developers in designing BCI-based systems such that they
can be easily employed with minimum cost for real-time
applications [31]. Therefore, two MI EEG datasets recorded
by 3 channels are chosen in this study. The two datasets are
from the BCI competition datasets recorded at Graz
University. More details regarding the two datasets are given
in the following subsections. The datasets consist of two
parts, namely the training part and the validation part. As
deal with the complexity of the nine subjects’ specific brain
signals, such as inter and intra-subject differences. Given the
lack of a large dataset to develop, evaluate, and validate a
WTNN, the datasets of all the nine subjects were combined
(union) to form a large dataset for all the trials involving
different brain complexities.