Page 54 - IJEEE-2023-Vol19-ISSUE-1
P. 54
50 | Al-Qaysi1, Al-Saegh, Hussein & Ahmed
• BCIC IV 2b dataset 0 1 2 3 4 5 6 7 8 9 sec
In this dataset, three EEG channels namely C3, Cz, and
C4 were used to acquire the signals of two motor imagery Feedback period with Cue
tasks (left hand and right hand). The dataset was collected
from nine subjects at a 250 Hz sampling frequency. EEG data Trigger
from 160 trials were collected while a subject watching a Beep
flat-screen and sitting in an armchair. Two types of recording
sessions were conducted, namely training without feedback Fig. 3: Trials recording time scheme of BCIC II dataset.
and evaluation with smiley feedback. During the first two
sessions, subjects were given a short warning tone to do four B. Preprocessing
seconds of a required motor imagery movement based on a
pointing arrow presented on a blank screen. However, during Inevitably, the EEG-MI signal is contaminated by noise
the other three sessions, subjects were instructed to move from various sources, such as body movements, eye blink,
grey smiley feedback centered on the monitor into either the facial muscle movements as well as artifacts from the
right or left direction after they are given a short warning surrounding environment, such as electromagnetic fields
beep. The smiley feedback was presented in four seconds and generated by electrical devices [1]. Since the framework
its color changes to red when it moved in the wrong direction relies on deep learning, the least preprocessing is used.
and green when it moved in the right direction. Fig. 2 shows Frequency filtering is carried out for enhancing the signal-
the timing scheme of the two types of sessions. to-noise ratio of the raw brainwaves and to enhance relevant
information of the signals. Specifically, the fourth-order
(a) Butterworth filter is applied with the range (8-30 Hz) given
that the MI EEG signals rely on the alpha (8-13 Hz) and beta
(b) (14-30 Hz) rhythms.
Fig. 2: Trials recording time scheme of BCIC IV 2b dataset
C.Time-Frequency Analysis
(a) without feedback, (b) with smiley feedback.
The time-frequency domain is a hybrid representation of
• BCIC II dataset a time-series signal. In principle, this representation
This dataset was recorded from a normal subject considers the signal properties in both temporal and
(female, 25y). The experiment consisted of 280 trials of 9 frequency domains. This representation yields images that
seconds total duration for each trial with a 128 Hz sampling highlight the contained frequencies in the time-series signal
frequency. The subject was quiet for the ?rst 2s, at t=2s an with the time slot those frequencies have been occurred.
acoustic stimulus indicating the beginning of the trial and Various methods of time-frequency analysis exist such as
across ‘+’ were displayed for 1s. Then at t=3s, an arrow (left autoregressive model (ARM), short-time Fourier transform
or right) was presented as a cue. Simultaneously, the subject (STFT), wavelet transform (WT), and discrete wavelet
was requested to perform the required motor imagery task transform (DWT). The DWT method has been proven to be
(left hand or right hand). The EEG data was ?ltered between more useful in characterizing non-stationary signals
0.5 and 30 Hz. The imaginary task was to move a block based effectively. Hence, DWT is adopted in this study for
on the given cue in the left or right direction. The used three representing the MI EEG signals in 2D images called
EEG channels were C3, Cz, and C4. Each session contains scalograms.
40 trials, half of them are for left-hand and half for right-hand
which are placed randomly. Seven sessions of such 40 trials DWT relies upon dilating and translating a particular
have been recorded with their labels [30]. Fig. 3 shows the function, called a mother wavelet, for representing a signal
timing scheme of the recording technique. as a linear combination of a set of wavelet functions. The
mother wavelet gives rise to these wavelets as a part of
resulting functions through shifting (dilation) and stretching
(translation) operations along the time axis, respectively. To
date, great attention is given to the WT in the field of
biomedical signal processing because of its efficiency in the
diagnostic as well as in the pattern recognition [32]. WT is
classified into two types; namely continuous wavelet
transform (CWT) and discrete wavelet transform (DWT):
%& (1)
!"#(%, ') = + ,(-).!*,# (-)/-
'&
