182
Bioinformatics of the Brain
occurrence of a stimulus and the corresponding neural response) could be eval-
uated in both analyses. Additionally, Hjorth parameters that include three
statistical parameters (activity, mobility, and complexity) [39] are another
useful analysis, especially for real-time systems due to low computation cost.
In BCI systems, the other prevailing and standard analyses could be cat-
egorized in the frequency domain. Fourier decomposition is the most widely
used method to convert the time domain to the frequency domain, result-
ing in a complex spectrum in which each frequency is represented by power
and phase. Wavelet analysis and Hilbert transformation after band-pass fil-
tering are the other common methods for the transformation to the frequency
domain. By utilizing the outputs of the analysis, multiple features could be
calculated such as mean and/or peak frequencies, mean and/or maximum
power in specific frequency bands, etc. Subsequently, frequency domain anal-
yses contain various functional connectivity metrics that are estimated to
evaluate neuronal interactions, including coherence, phase synchronization,
phase-slope index, and Granger causality [40].
On the other hand, if there is a need to assess changes in amplitude or
power spectrum over time, it is recommended to choose the time-frequency
domain analysis. This domain can include techniques like event-related spec-
tral perturbations (ERSPs) that could be obtained via different mathematical
methods (e.g., wavelet transform) focus on changes in spectral properties over
time, and inter-trial phase coherence (ITPC) that represent phase consistency
over trials (typically within the range of zero to one) [41]. The percentage of
using the time-frequency domain features is higher than the other domains
in the EEG MI-BCI systems [42]. In recent motor imagery BCI applications,
spatial domain analyses are also employed for the feature extraction step.
One of the popular algorithms in this domain is Common Spatial Patterns
(CSP) which involves converting EEG signals into a different space through
spatial filtering techniques, aiming to optimize the variance of one group while
minimizing the variance of the second group [43].
7.3.5
Artificial Intelligence Techniques in EEG-based BCI
Systems for Neural Decoding
The mid-20th century saw the birth of Artificial Intelligence (AI) as a field
of study, with early applications focusing on problem-solving and symbolic
methods. With today’s technological advancement in AI, from industries to
home appliances, it has become a ubiquitous reality that is integral to our
daily experiences.
Once the feature-extraction process has been established in a BCI system,
proper neural decoding models should be selected to recognize feature patterns
(e.g., ERP features when a paralyzed user’s intent to move a cursor using their
thoughts or oscillatory activities when an individual with cognitive decline
pay attention to external stimuli). To date, various AI techniques, including
conventional ML and DL approaches, have been developed and introduced to