Page 141 - Handbook of Biomechatronics

138 Domen Novak

is constantly analyzed for the presence of the EEG waveform of interest (e.g.,
motor imagery). Asynchronous operation is thus significantly more com-
plex, as it must account for the fact that the system is likely in a “no
command” state the majority of the time. This is acknowledged to be a sig-
nificant challenge in BCIs, and was the subject of a BCI signal-processing
competition in 2008 (Tangermann et al., 2012). At the same time, the asyn-
chronous mode is more realistic and commonly used in, for example, assis-
tive devices: the user may require assistance at any point in time, but will
likely spend long periods of time not needing it (Ortner et al., 2011;
Pfurtscheller et al., 2005).
In both synchronous and asynchronous modes, the pattern-recognition
method depends on the paradigm being used:
• For SSVEPs, the goal is to measure the dominant frequency in the EEG,
which can be done using any established power spectral density (PSD)
calculation method (Rangayyan, 2015). The dominant frequency in
the EEG can then be matched to the closest frequency shown on the
screen: for example, if symbol A flashes with 6Hz and B flashes with
12Hz, a measured dominant frequency of 6.5Hz is interpreted as the
user choosing symbol A.
• For the P300 wave and ERPs, the goal is to detect a specific waveform,
which can be done with any standard event detection and classification
method (Rangayyan, 2015). Once the event has been detected and iden-
tified as a P300 or ERP, its cause can be determined. For example, to find
the cause of the P300, we look for a stimulus that was presented to the
person 300ms prior to the P300.
• Motor or mental imagery causes EEG power to decrease in some fre-
quency bands and at some electrode locations while increasing in other
bands and at other electrode sites. Thus, to recognize imagery, several
features are extracted from PSD estimates and input into classification
algorithms such as linear discriminant analysis (Horkietal.,2011)or
support vector machines (Xu et al., 2011). Among such “classic” algo-
rithms, particularly support vector machines have been recommended
for the synchronous mode of operation (Lotte et al., 2007). However,
recent years have seen extensive development of new types of classifi-
cation algorithms for motor and mental imagery, including adaptive
classifiers, matrix and tensor classifiers, transfer learning, and deep
learning (Lotteetal.,2018). Among these, particularly adaptive
classifiers have been shown to outperform most other algorithms
(Lotte et al., 2018).