IEEE Computational Intelligence Magazine - May 2021 - 38

conducted the experiments with shorter length data.6 We preprocessed the SSVEP signals by applying band-pass filtering
between 4 and 15 Hz and selected eight channels in the occipital region, 'PO3, POz, PO4, PO9, O1, Oz, O2, and PO10,'
because this region is widely used for SSVEP classification [22].
4) Drowsiness
With respect to passive BCI [6], we considered two different
paradigms, seizure EEG signals [12] and vigilance EEG signals
[36]. Owing to its theoretical and practical benefits, in this
study, we conducted experiments identifying drivers' mental
fatigue. We also used a publicly available SEED-VIG EEG
dataset [36]7 for the drowsy driving task data. This dataset
consists of 23 experiments, i.e., trials, and each trial is recorded for approximately 2 hours while simulated driving occurs.
The EEG signals are acquired from 17 electrode channels
according to the 10-20 system and sampled at 200 Hz. For
this dataset, we band-pass filtered EEG signals in the range
between 0.5 and 40 Hz, each epoch was 8 sec in length. Since
the dataset was originally labeled using PERCLOS levels [36],
we categorized the label vectors into three classes, awake, tired,
and drowsy, with two threshold values (0.35 and 0.7). Then, for
the 23 experiments, a five-fold cross-validation was used for
performance estimations.
5) Seizure
Finally, we conducted seizure onset detection experiments with
the widely used and publicly available CHB-MIT dataset
[12]8. The CHB-MIT dataset contains EEG data from 24
subjects sampled at 256 Hz acquired from 23 electrode channels (24 or 26 in a few cases) according to the 10-20 system.
In this work, we selected EEG trials that have the same 23
channels montage and removed some trials acquired from the
different montage. By following [14], we used a leave-onerecord-out cross-validation. More precisely, we trained the proposed method using all non-seizure records and all seizure
records but one, and tested the model on the remaining seizure record. Then, we repeated this process for the number of
seizure records in the dataset, thus, each seizure record was
tested. For training, the test trial epochs were 10 sec in length.
During validation and testing sessions, a 10 sec length EEG
signal was fed into our proposed network using a 1/256 sliding. In other words, we used epochs in units of 10 seconds in
the training session, but a sliding-window of 10 sec length in
the testing session. Then, we observed how the probability of
ictal or normal for each EEG signal timepoint was continuously changing over time.
For all datasets, we exploited early stopping strategy. More
specifically, we randomly selected and split again training samples at a 9:1 ratio for each experiment and used them for training and model selection respectively.
6
Experimental results of the KU-SSVEP dataset with shorter length are reported in
Supplementary B.
7
Available at: http://bcmi.sjtu.edu.cn/seed/download.html
8
Avaliable at: https://physionet.org/content/chbmit/1.0.0/

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2021

B. Experimental Settings

In our work, we compared our method with paradigm-specific
linear model-based and deep learning-based methods for each
EEG paradigm.
1) Linear Models-Motor Imagery
First, we built a CSP with a linear discriminant analysis (CSP +
LDA) [1] and an FBCSP with an LDA (FBCSP + LDA) [11]
for MI decoding. We used four filters and regularized covariance for the CSP and FBCSP. Additionally, we also used nine
non-overlapped filter banks in the 4 + 40 Hz range, i.e.,
4 + 8, 8 + 12, f, 36 + 40 Hz, and finally selected 10 features
using the mutual information-based feature selection method
FBCSP. Finally, for the BSSFO [13], we set 30 particles on the
4 + 40 Hz range MI-EEG to achieve robust classification
results and classified the extracted features with a SVM
(BSSFO + SVM).
2) Linear Models-Steady-State Visually Evoked Potentials
We also built a standard CCA [9] for SSVEP classification. We
set reference signals for each stimulus including second harmonics. Furthermore, the standard CCA does not require
training samples for the optimization, thus we only estimated
each session in its entirety from the KU-SSVEP dataset [8] for
the CCA performance estimation.
3) Linear Models-Drowsiness
For the drowsy state detection experiment, we estimated the
filter-banked input EEG data PSD in a channel-wise manner
for extracting spatio-spectral features and classified the learned
features using an SVM with a radial basis function (RBF) kernel (c = 1/d input where d input denotes the input feature dimension) [36].
4) Linear Models-Seizure
In addition, we also reimplemented Shoeb and Guttag [14]'s
method for the seizure onset detection experiment. We applied
the PSD to the EEG data in a channel-wise manner. Then, the
3 sec time window time evolution method was used for capturing temporal information. Finally, the represented spatiospectral-temporal features were fed into an SVM using an RBF
kernel (c = 1/d input).
5) Deep Neural Networks-Motor Imagery
We also implemented deep learning-based BCI models9 for
MI. Basically, most of the existing deep learning models [3],
[17], [18], [25], [26] have focused on a paradigm-specific BCI
task. However, we conducted experiments over all types of
datasets for each deep learning model to demonstrate the validity of the proposed method. We built a Shallow ConvNet and a
Deep ConvNet as proposed by Schirrmeister et al. [3]. The
Shallow ConvNet consists of two convolutions, temporal and
9
See 'Supplementary A: Architectural Details of Deep Models for BCIs' for more
detail architectures and learning schedules.

http://bcmi.sjtu.edu.cn/~seed/seed-vig.html https://www.physionet.org/content/chbmit/1.0.0/

IEEE Computational Intelligence Magazine - May 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - May 2021