EEG-based BCI Systems in Neuropsychiatric Diseases
183
detect task-specific patterns in the BCI field (please see the comprehensive
review [44]).
In the BCI context, conventional decoding techniques have been pioneer-
ing the field for understanding how machines and humans can work together,
as they require extensive feature extraction and selection mechanisms to de-
code effectively. This was necessary to identify meaningful patterns within the
complex data generated by the brain, facilitating the development of effective
BCI systems. Some widely employed ML approaches used in BCI applications
are Linear Discriminant Analysis (LDA), Support Vector Machines (SVM),
k-nearest neighbors (k-NN), Naive Bayes classifier, Decision Trees (DT), and
Random Forests [45]. These techniques have allowed researchers to decode
complex brain data, enabling the development of BCIs that can perform vari-
ous tasks, from translating motor imagery into commands to recognizing P300
signals by attention to stimuli, and even interpreting SSVEPs for control in-
terfaces.
In particular, LDA and SVM are some of the most efficient classification
approaches among the conventional ML techniques. However, due to LDA’s
strict reliance on the linear separability of the data, its performance heavily
deteriorates when dealing with multiclass classification tasks or complex data
where there is significant overlap among classes [46]. Despite the limitations,
SVM is considered a more robust classifier that can work with a limited num-
ber of training samples in a high-dimensional setting [47]. In some cases, DT
and Artificial Neural Networks, which are algorithms that do not assume lin-
ear separability, are capable of modeling overlapping class distributions more
effectively [46]. Among the ML approaches, along with DT, LDA, SVM, and
k-NN, non-linear approaches have also proven effective in BCIs for classifying
more complex tasks and handling multiclass problems, often outperforming
other classifiers in accuracy [48].
Additionally, conventional ML approaches have paved the way for more
sophisticated DL approaches that are an extension of artificial neural net-
works (ANN). Over the last decade, the integration of DL with EEG-based
BCIs has significantly advanced the field in terms of more intricate and prac-
tical applications. The adoption of DL and transfer learning techniques within
EEG-based BCI systems has revolutionized the way we decode and interpret
brain signals, showing the invaluable insights and necessary groundwork laid
by conventional ML approaches. Convolutional Neural Networks (CNN), deep
neural networks (DNN), Long Short-Term Memory Networks (LSTMs), and
Generative Adversarial Networks (GANs) are some of the popular DL meth-
ods in the literature.
CNN has widespread use and the ability to work and extract features
across temporal, spatial, and spectral domains, which overall characterizes
the nature of EEG data in a broad sense. As mentioned in Section 7.3.4, EEG
signal encapsulates a variety of information across its temporal, spatial, and
spectral domains, each holding great clues into brain function and cognitive
states. CNN architectures, when compared to conventional ML techniques,