
augmentation. The overall average improvement
across all augmentation methods was 0.29±0.08%,
with the best improvement using the Noise Injection
(NI) method (with the overall average improvement
of 0.36%), followed by the sliding window method
(average improvement of 0.33%), and GAN (average
improvement of 0.28%). The most commonly (62%)
used classifier was CNN. The second most popular
(16%) classifier was a hybrid one, the combination of
Long Short-Term Memory (LSTM) and CNN. In 8%
of studies, MultiLayer Perceptron (MLP) was used,
and in 6% of studies, only LSTM was used.
(Al-Saegh et al., 2021) discussed studies dealing
with EEG signal processing in MI tasks; specifically,
they went through 40 studies written between 2015
and 2020. The most significant proportion of studies
(45.6%) focused only on detecting two MI classes,
left-hand and right-hand movement. The second most
frequent (31.6%) classification task dealt with de-
tecting left-hand, right-hand, tongue, and foot move-
ments. The most used classifier (73% of the studies)
was CNN. Another 14% of studies used a form of hy-
brid architecture, typically a combination of CNN and
LSTM. The most frequently (66%) used classifier ac-
tivation function was ReLU, and the most frequently
(47%) used learning optimizer was Adam.
Experiments aimed at detecting MI in the EEG
signal are also described in (Mochura, 2021) and
(Saleh, 2022). (Mochura, 2021) work involved the
construction of various feature vectors as an input to
a MultiLayer Perceptron (MLP). The most success-
ful, in terms of classification accuracy, feature vec-
tor was constructed by computing ERD and ERS for
each measurement. (Mochura, 2021) produced an
inter-subject model, and the best average classifica-
tion accuracy (90.05%) was achieved using a feature
vector consisting of the calculated ERD followed by
the calculated ERS. Only one class of MI was consid-
ered (movement of any limb); movement vs. resting
state was classified. ERD was calculated in the alpha
band, whereas the ERS was calculated in the lower
beta band.
Saleh (Saleh, 2022) focused on detecting Senso-
riMotor Rhytm (SMR), where a band-pass type filter
with cutoff frequencies of 8-13Hz was applied to each
signal epoch. Either the CSP method was then applied
to the filtered epochs, or the filtered epochs created di-
rectly the input to the SVM and Linear Discriminant
Analysis (LDA) classifiers. The EEG signals from
the C3, C4, and Cz electrodes were used. (Saleh,
2022) formed intra-subject models, i.e., a personal-
ized model for each subject, and performed a multi-
class classification where classes represented the left
motion, right motion, and resting state, respectively.
When summarizing the literature review, we can
state that various methods and techniques are used for
processing EEG signals and detecting MI patterns. In
the preprocessing phase, a band-pass filter is used for
the alpha and beta bands. Then, the channels (elec-
trodes) to be used are selected; the relevant channels
are C3, Cz, and C4. The parts of the EEG signal for
which an event has occurred (epochs, e.g., when the
subject has been instructed to move their hand) are
selected. The duration of an epoch varies, typically
ranging from 2-4 seconds. Removing signal artifacts
is also crucial, but this step can be quite complex, and
most studies have not mentioned it.
The last step in the preprocessing phase is the se-
lection of features. Most reviewed studies have not
constructed feature vectors and used the time series
of each epoch as inputs to classifiers. Other stud-
ies performed feature extraction by calculating the
signal properties or converting the spectrogram into
an image. (Mochura, 2021) used directly calculated
ERD and ERS, and each epoch’s average power de-
crease/increase as input features. However, other
studies did not use averaging; single trials (individual
epochs) were used as classifier inputs.
The time series is the most commonly used repre-
sentation of the EEG signal due to the popularity of
CNN. This representation is also the easiest to imple-
ment, as no feature extraction is necessary. However,
the sampled signal is not usually used directly; it is
first preprocessed: electrode selection, filtering, and
artifact removal are generally performed as described
above. To detect MI, it is necessary to use a synchro-
nization label with the measured EEG signal. For the
analysis, we are only interested in selected windows
of the EEG signal around the synchronization mark-
ers; these time intervals are called epochs. Typically,
MI tasks are repeated during experiments, i.e., more
epochs are obtained from the EEG signal.
In addition to analyzing the signal in the time do-
main, it is also reasonable to analyze the signal in the
frequency domain. Since MI is associated with the
desynchronization (ERD) and subsequent synchro-
nization (ERS) of the alpha and beta frequency bands,
detecting motion from the EEG signal could be done
by analyzing its frequency spectrum.
By extracting features from the time domain only,
we do not consider the frequency spectrum features.
Similarly, we lose information from the time domain
by extracting features only from the frequency do-
main. For this reason, these characteristics are some-
times considered weak for extracting significant fea-
tures (Al-Saegh et al., 2021). The short-term Fourier
transform (STFT), wavelet transform, and Hilbert fil-
ter, in particular, convert the input signal into the time-
HEALTHINF 2024 - 17th International Conference on Health Informatics
814