184

Bioinformatics of the Brain

enable efficient and effective feature extraction that does not require extensive

manual feature engineering. In the temporal domain, CNNs can analyze the

functions of brain signals over time, capturing dynamic neural responses to

stimuli as well as spontaneous brain activity. On the other hand, the spectral

quality of EEG signals is complementary to CNN’s ability to identify patterns

within specific frequency bands, which are crucial for discriminating between

different states of brain activities and cognitive tasks [49]. CNNs, although

somewhat limited, extend their feature workspace to the spatial domain, where

a varied number of EEG electrodes collect data from the different regions

of the human brain. The spatial features are essential qualities for locating

changes on the scalp that occur during different cognitive processes. Recent

research explored the integration of multi-domain strategies for decoding in

BCI systems [50, 51]. The use of CNN to combine features from multiple

domains has been shown to improve interpretability and decoding accuracy

[52, 53].

The other DL method that is Transfer learning has emerged as a key

strategy in overcoming the inherent challenges of cross-subject and cross-

session variability in EEG-based BCI systems, enabling more adaptability.

Early transfer learning efforts in BCI focused on configuring decomposition

algorithms to provide easier adaptation across sessions and subjects. Kraule-

dat et al. [54] focused on stable prototype filters to reduce the need for ex-

tensive calibration at the start of new sessions, which may cause fatigue on

the subject. Similarly, Kang et al. [55] introduced a composite Common Spa-

tial Pattern (CSP) approach with regularized covariance matrices to provide

subject-to-subject transfer. Some other studies focused on statistical aspects

of the EEG signals. One notable approach is the Stationary Subspace Analysis

(SSA) [56], which separates EEG signals into stationary and non-stationary

components, focusing analysis on the more consistent stationary components

to improve the classification tasks. A more recent study [57], applied a sophis-

ticated cross-subject data augmentation-focused deep learning framework for

enhancing the cross-subject capabilities of motor imagery BCI applications.

The architecture of the proposed framework incorporates a multi-domain fea-

ture extractor based on CSP equipped with a sliding window mechanism,

alongside a parallel two-branch CNN.

7.4

Current BCI Systems Applications

This chapter intended to provide a short overview of the application domains

of BCI, encompassing systems facilitating the control of external devices and

the diagnosis of neuropsychiatric diseases.