Machine Learning to Identify Neural Correlates of Music and Emotions - Guide to Brain-Computer Music Interfacing - page 88

self-assessment manikin (SAM) (Bradley and Lang 1994 ). SAM is comprised of
descriptive visual images that range across a nine-point rating scale for pleasure,
arousal, and intensity, respectively. This tool for reporting emotional response has
been extensively used to label emotionally loaded stimuli for a range of databases
such as the International Affective Picture System (IAPS) (Lang et al. 2008 ),
International Affective Digital Sounds (IADS) (Bradley et al. 2007 ), and the dataset
for emotional analysis using physiological signals (DEAP) (Soleymani et al. 2012 ).
However, one downside to the use of SAM and other self-assessment tools when
exploring the in

uence of music is that they only provide information of the

individuals affective state at a discrete time, post-stimulus in most experimental

paradigms. As such, there is a growing interest in emotional research for a tool
which provides the participant the ability to continuously report how they feel in an
easy, quick, and accurate manner. Current tools available to researchers include
FEELTRACE (Cowie and Douglas-Cowie 2000 ) and EMuJoy (Hevner 2007 ), both
of which allow the subject to position and navigate their affective state through two-
dimensional emotional models.

'

5.3 Measuring Neurological Activity

There are three broad groups of feature types that may be extracted from neuro-
logical data: those based upon the amplitude of the recorded signals, those based in
the frequency content, and those based upon the phase content (Lotte et al. 2007 ;
Hwang et al. 2013 ). Additionally, combinatorial features, which combine two or
more feature types, may also be considered. For example, time-frequency activity
maps may be used to describe changes in the amplitude and frequency distribution
of the EEG or magnetoencephalogram (MEG) over time.
Amplitude-based features may include measures such as event-related potential
(ERP) amplitude, peak latency, statistical measures of the distribution of the data,
measures of relationships within the data (e.g. correlation), voxel strength in speci c

regions of interest (ROIs) in a magnetic resonance imaging (MRI) map, etc. Fre-
quency-domain features are used to describe how the frequency content of the data
changes with speci

c control tasks or over time. This can include measures of the

magnitude of speci c frequencies, the distribution of frequencies across the power

spectra (power spectral density; PSD), coherence measures, or the relative power of
speci

c frequencies (Wang et al. 2011 ). Phase-domain features are used much less
frequently in BCI research (Hwang et al. 2013 ), but nonetheless show some promise,
particularly when used in the investigation of relationships between different spatial
regions during speci c tasks (Daly et al. 2012 ). Combinatorial features are being used

in an increasing proportion of BCI research (Hwang et al. 2013 ). This is most likely
due to the increasing interest in hybrid BCIs (hBCIs), in which paradigms or signal
types are combined (Pfurtscheller et al. 2010 ;M ü ller-Putz et al. 2011 ).
Figure 5.1 provides a taxonomy of feature types which may be used to describe
neurological signals, including examples of each type.