Page 144 - Handbook of Biomechatronics

Biomechatronic Applications of Brain-Computer Interfaces 141

responses to specific cortical regions (Naseer and Hong, 2015; Lloyd-Fox
et al., 2010). When measured properly, the fNIRS signal closely correlates
with the blood oxygen level dependent (BOLD) signal from functional
magnetic resonance imaging (Huppert et al., 2006), but can be measured
with relatively simple, portable hardware.

1.3.1 fNIRS Paradigms
The most common fNIRS paradigm is to measure mental workload using
methods similar to EEG: fNIRS of the prefrontal cortex is recorded over
1–5min, different features are extracted from it, and classification algorithms
are used to translate the features into different levels of workload (Naseer and
Hong, 2015; Girouard et al., 2013). Less commonly, it is also possible to use
fNIRS to measure motor imagery—using multiple fNIRS channels over the
human motor cortex allows observation of distinctly different hemodynamic
responses to, for example, imagery of the left hand and the right hand
(Naseer and Hong, 2015; Sitaram et al., 2007).

1.3.2 Signal Processing and Pattern Recognition
Regardless of the paradigm, fNIRS signals still contain various types of noise
that are not related to brain activity. These are commonly reduced by
preprocessing the optical density signals before converting them into oxygen
saturation signals, and can be roughly divided into instrumental noise (e.g.,
instrumental degradation), experimental error (e.g., sudden head motions),
and physiological noise (e.g., effects of heartbeat and respiration on blood
pressure fluctuations) (Naseer and Hong, 2015). Some of these (e.g.,
high-frequency instrumental noise) can be removed using simple bandpass
filters while others require more advanced methods such as principal/inde-
pendent component analysis or adaptive filtering (Naseer and Hong, 2015).
After noise removal, it is common to convert the optical density signals
into oxygen saturation signals via the modified Beer-Lambert law, then
extract different features from the oxygen saturation signals as a basis for pat-
tern recognition (Naseer and Hong, 2015). The most frequently used fea-
tures are those related to the signal shape (signal mean, signal slope, signal
variance, skewness, kurtosis, zero crossing rate, etc.) though more advanced
feature extraction methods such as wavelet transforms have been used with
some success (Naseer and Hong, 2015). These features are then input into
standard classification algorithms such as linear discriminant analysis, support
vector machines, and artificial neural networks (Naseer and Hong, 2015).