206
G. D. Johnson and D. J. Krusienski
The actual EEG patterns corresponding to the two mental states can be visualized
by inverting the filtering matrix W . For standard CSP analysis of EEG, the features
of X are simply the instantaneous bandpass filtered voltages at each electrode. The
incoming data is projected onto the CSP subspace and the variance for each projection is computed. Thresholds or classifiers can be implemented on the resulting
variances to identify the appropriate class. Extensions of the CSP algorithm have
been developed to further exploit the temporal and spectral characteristics of the
underlying signals [9, 16].
9.3 Reactive Paradigms
9.3.1 Transient Evoked Potentials
Transient responses in reactive paradigms, known as evoked potentials (EPs), are
time-locked to an external sensory stimulus. Thus, segments of EEG are analyzed
over a predefined temporal window with respect to the onset of the stimulus. Because
the signal-to-noise ratio (SNR) of EEG is low, multiple transient response observations are typically averaged to attenuate the background noise and produce a more
reliable detection. While EPs can exhibit transient frequency bursts, they are most
commonly analyzed using time-domain techniques. Thus, forms of spatio-temporal
template matching are typically implemented to detect the relevant combination of
amplitude deflections (e.g., voltages) at various channels that best characterize the
response to the stimulus. The following describes a fundamental methodology for
detecting and classifying EPs such as the P300 response [11].
9.3.1.1 Preprocessing
The appropriate channels must be identified for the EP to be analyzed, which are
well-defined in the literature [2, 15]. Alternately, subspace decomposition approaches
such as principle component analysis (PCA) [8] or independent component analysis
(ICA) [12, 19] can be used to create a spatial filter that enhances the representation of
the response for detection. Because EPs are generally comprised of low-frequency
oscillations, it is common to bandpass filter the signals in the range of 0.1–0.5 Hz
for the highpass cutoff and 10–30 Hz for the lowpass cutoff. It is also common to
decimate the resulting signals based on the lowpass cutoff according to the Nyquist
rate [28] to further reduce the dimensionality of the data.
G. D. Johnson and D. J. Krusienski
The actual EEG patterns corresponding to the two mental states can be visualized
by inverting the filtering matrix W . For standard CSP analysis of EEG, the features
of X are simply the instantaneous bandpass filtered voltages at each electrode. The
incoming data is projected onto the CSP subspace and the variance for each projection is computed. Thresholds or classifiers can be implemented on the resulting
variances to identify the appropriate class. Extensions of the CSP algorithm have
been developed to further exploit the temporal and spectral characteristics of the
underlying signals [9, 16].
9.3 Reactive Paradigms
9.3.1 Transient Evoked Potentials
Transient responses in reactive paradigms, known as evoked potentials (EPs), are
time-locked to an external sensory stimulus. Thus, segments of EEG are analyzed
over a predefined temporal window with respect to the onset of the stimulus. Because
the signal-to-noise ratio (SNR) of EEG is low, multiple transient response observations are typically averaged to attenuate the background noise and produce a more
reliable detection. While EPs can exhibit transient frequency bursts, they are most
commonly analyzed using time-domain techniques. Thus, forms of spatio-temporal
template matching are typically implemented to detect the relevant combination of
amplitude deflections (e.g., voltages) at various channels that best characterize the
response to the stimulus. The following describes a fundamental methodology for
detecting and classifying EPs such as the P300 response [11].
9.3.1.1 Preprocessing
The appropriate channels must be identified for the EP to be analyzed, which are
well-defined in the literature [2, 15]. Alternately, subspace decomposition approaches
such as principle component analysis (PCA) [8] or independent component analysis
(ICA) [12, 19] can be used to create a spatial filter that enhances the representation of
the response for detection. Because EPs are generally comprised of low-frequency
oscillations, it is common to bandpass filter the signals in the range of 0.1–0.5 Hz
for the highpass cutoff and 10–30 Hz for the lowpass cutoff. It is also common to
decimate the resulting signals based on the lowpass cutoff according to the Nyquist
rate [28] to further reduce the dimensionality of the data.