2.1. Dataset and Preprocessing

Experimental tests of the proposed model was performed using the PhysioNet EEG Motor Movement/Imagery dataset (Available online: https://physionet.org/content/eegmmidb/1.0.0/, accessed date: 22 September 2022) which contains about 1500 one- and two-minute EEG recordings from 109 subjects. The BCI2000 [39] system was used to record EEG data while the experiment’s subjects engaged in various motor imagining activities. Each participant completed 14 experimental runs, which included two baseline runs with their eyes opened and closed as well as three two-minute runs with the tasks of executing or imagining the opening and closing of the left or right hand, both fists or both feet. For the validation of the model, two subgroups of the dataset were used. Left-hand or right-hand movement tasks are included in the first subgroup. Imaginary left-hand or right-hand movement tasks are included in the second subgroup. The motor movement or imagery tasks were recorded as EEG signals on 64 channels placed on subject’s scalps. Each channel is annotated with three codes: T0, T1, and T2. T0 refers to the rest period, while T1 refers to the motion of the left hand in some tasks and both fists in others. T2 denotes the movement of the right hand for some tasks and both feet for others. Each experimental run was partitioned based on these annotations. But according to the literature [35,40], the EEG channels of six subjects (subject 38, 88, 89, 82, 100, and 104) were not annotated as specified in the experiment. As a result, partitioning each experimental run based on these annotations carries the potential risk of making incorrect decisions. Due to incorrect annotations, these six subjects were eliminated. Hence, 103 subjects out of 109 were used.

Each trial’s input data were divided into (C, W) dimensions, where C stands for the number of channels and W is the temporal dimension. All trials contained 4 to 4.1 s sustained and continuous movements for executed and imagined tasks. Therefore, to keep the dataset consistent, 4 s of data were clipped on each trial, sampled at 160 Hz, for a total of 640 samples. The sliding window approach was used to divide the 640 samples into eight non-overlapping windows of 80 samples each. The target label of each window was identical to the initial trial. The eight-time windows can provide more discriminatory information on the motor imagery data. After that, the signal processing module in the Gumpy BCI library [41] was applied to process the EEG signals. To eliminate the alternating current (AC) noise at the 60 Hz frequency, a notch filter was applied to the data. Then, Butterworth band-pass filtering was performed on the data in the 2 Hz to 60 Hz range with an order value of 5.

BCI Competition IV-2a dataset (BCI Competition IV dataset. Available online: https://www.bbci.de/competition/iv/, accessed date: 15 October 2022) was used to evaluate the proposed model [42]. The Graz University of Technology generated famous public MI-EEG dataset in 2008 known as BCI-2a. The dataset’s small number of samples taken in uncontrolled conditions with many artifacts makes decoding MI tasks difficult. The dataset contains 5184 trials (samples) of MI-EEG data collected from 9 participants applying 22 EEG electrodes (576 trials per participant). Moreover, 3 extra electrooculography (EOG) channels give information about eye movements. MI trials are 4 s long, captured at 250 Hz, and filtered between 0.5 and 100 Hz. Two sessions were captured for each subject on different days. There were 288 trials per subject. Four MI tasks corresponds to each trial: imagined movement of the left hand, right hand, both feet, and tongue.

The time frame length selected for this dataset is 4.5 s (from 1.5 s to 6 s), producing 1125 samples [33]. Standardization was applied in the pre-processing step [43].

This publicly available dataset [44] was collected from 9 subjects using 3 bipolar electrodes at a sampling rate of 250 Hz. A bandpass filter between 0.5 and 100 Hz was then used for filtering. Moreover, 3 extra EOG channels were employed to collect data on eye movement. The dataset consisted of two classes, called the motor imagery of left hand and right hand. Each subject participated in 2 screening sessions without feedback and 3 screening sessions with feedback. Each session consisted of six runs with ten trials each and two classes of imagery. This resulted in 20 trials per run and 120 trials per session. The pre-processing step was similar to the BCI IV-2a dataset.