9 Computational EEG Analysis for Brain-Computer Interfaces
207
9.3.1.2 Feature Extraction and Classification
The responses are collected based on the onset of each stimulus. Temporal windows
for responses are typically around 1 s in length, but can vary depending on the application. Additionally, the window can begin prior to the onset of the stimulus to provide
information about the baseline prior to the stimulus. Figure 9.3 illustrates the timing
and associated EEG alignment of a typical synchronous stimulus presentation. The
Stimulus Label in the upper panel (solid trace) indicates the onset and duration of
each sensory stimulus. The lower panel shows the time-aligned EEG corresponding
to the stimulus labels. The three shaded regions are example 800 ms windows corresponding to the onset of the first three stimuli in the upper panel. Similar response
windows would be collected for all subsequent stimulus labels.
Assuming a binary detection scenario of predicting if the response was generated
by a target or non-target stimulus, the collected EPs can be labeled and used to train a
classifier [14]. For instance, in Fig. 9.3, all responses corresponding to Target Trial
1 and all responses corresponding to Target Trial 0 would be parsed for training
the binary classifier. Figure 9.8 shows an example of the averaged P300 responses
for target and non-target stimuli for the commonly-used electrode locations. Simple
yet effective classifiers select the individual spatio-temporal features (circled) that
optimize a regression model shown at the bottom. Using this approach, all features
with high univariate correlation with the task are not necessarily selected for the
model since they might have high covariance. Additionally, features with low univariate correlation may be included in the model to reduce noise or compensate for
other selected features. This can also be generalized to multi-class problems. For
classifying independent data, the resulting classifier scores are averaged over each
stimulus label and the stimulus associated with the largest average score is classified
as the selected target.
9.3.2 Steady-State Evoked Potentials
Steady-state responses such as steady-state visual evoked potentials (SSVEP) and
steady-state somatosensory evoked potentials (SSSEP) in reactive paradigms generally present multiple, spatially-distinct stimuli, each at a unique frequency. Because
the user focuses attention on a single stimulus (target) at a time in the presence of
the other stimuli (distractors), the objective is to detect features of the EEG that are
associated with the current target frequency. This forms a multiclass detection problem. Ideally, the EEG signal power at the target frequency and its harmonics will
dominate compared to the distractor frequencies (and their respective harmonics) but
this is not always the case for various reasons and more sophisticated techniques are
employed to improve the detection.
207
9.3.1.2 Feature Extraction and Classification
The responses are collected based on the onset of each stimulus. Temporal windows
for responses are typically around 1 s in length, but can vary depending on the application. Additionally, the window can begin prior to the onset of the stimulus to provide
information about the baseline prior to the stimulus. Figure 9.3 illustrates the timing
and associated EEG alignment of a typical synchronous stimulus presentation. The
Stimulus Label in the upper panel (solid trace) indicates the onset and duration of
each sensory stimulus. The lower panel shows the time-aligned EEG corresponding
to the stimulus labels. The three shaded regions are example 800 ms windows corresponding to the onset of the first three stimuli in the upper panel. Similar response
windows would be collected for all subsequent stimulus labels.
Assuming a binary detection scenario of predicting if the response was generated
by a target or non-target stimulus, the collected EPs can be labeled and used to train a
classifier [14]. For instance, in Fig. 9.3, all responses corresponding to Target Trial
1 and all responses corresponding to Target Trial 0 would be parsed for training
the binary classifier. Figure 9.8 shows an example of the averaged P300 responses
for target and non-target stimuli for the commonly-used electrode locations. Simple
yet effective classifiers select the individual spatio-temporal features (circled) that
optimize a regression model shown at the bottom. Using this approach, all features
with high univariate correlation with the task are not necessarily selected for the
model since they might have high covariance. Additionally, features with low univariate correlation may be included in the model to reduce noise or compensate for
other selected features. This can also be generalized to multi-class problems. For
classifying independent data, the resulting classifier scores are averaged over each
stimulus label and the stimulus associated with the largest average score is classified
as the selected target.
9.3.2 Steady-State Evoked Potentials
Steady-state responses such as steady-state visual evoked potentials (SSVEP) and
steady-state somatosensory evoked potentials (SSSEP) in reactive paradigms generally present multiple, spatially-distinct stimuli, each at a unique frequency. Because
the user focuses attention on a single stimulus (target) at a time in the presence of
the other stimuli (distractors), the objective is to detect features of the EEG that are
associated with the current target frequency. This forms a multiclass detection problem. Ideally, the EEG signal power at the target frequency and its harmonics will
dominate compared to the distractor frequencies (and their respective harmonics) but
this is not always the case for various reasons and more sophisticated techniques are
employed to improve the detection.