Brain–Computer Interfaces - Index of

A Simple, Spectral-Change Based, Electrocorticographic Brain–Computer Interface 249
respect to the other target). What this metric tells us is how much of the variation
in the joint data set σ 2 m∪r can be accounted for by the fact that sub-distributions of
movement (m) and rest (r) periods might have different means, ¯m and ¯r (Nm, Nr,
and are the number of samples of type m, r, respectively, and Nm∪r = Nm + Nr).
“Feature maps” (Fig. 4f) can tell us about which electrode-frequency combinations
discriminate between cues. We can calculate Amr for each electrode, frequency band
combination, to create feature maps of discriminative potential. When performed on
a screening task with actual movements (overt) or imagined movements (covert),
we can identify specific electrode-frequency power features as candidates for
feedback.
Several previous findings [6, 18–20] have shown a consistent decrease in power
at low frequencies, and a characteristic increase in power at high frequencies during
movement when compared with rest (Fig. 3). These changes in the cortical spectrum
may be decoupled into distinct phenomena [13]. At low frequencies, there is a band
limited spectral peak which decreases with activity, consistent with event-related
desynchronization (ERD). At high frequencies, a broad, power-law like increase in
power may be observed, which is highly correlated with very local cortical activity.
This functional change has been denoted the “χ-band” or “χ-index” when explicitly
targeting this broad spectral feature [13, 25], and “high-γ ” when seen as a bandspecific
increase in power at high frequencies [19, 26, 27].AsshowninFig.4, itis
often convenient to choose a low frequency band to capture ERD changes (α/β) in
the classic EEG range, and a high frequency band to capture broad spectral changes
[6]. Because this high frequency change is more specific for local cortical activity,
they are often the best choice for BCI control signals [28].
As with EEG, one feature-driven approach is to look across channels and frequency
bands to obtain reliable features for feedback [2, 8–10]. This can be done
manually, selecting spectral features from intuitive cortical areas, such as sensorimotor
cortex. It can also be done using naïve, blind-source deconvolution and machine
learning techniques [21, 29–34]. Sophisticated recombination techniques, optimized
for an offline screening task, face the potential confound that the resulting mapping
is not intuitive for subject control, particularly if the distribution of cortical spectral
change is different for screening than it is for feedback studies (which it can be).
Simple features, in contrast, may be employing only a fraction of the potential signal
for the feedback task, and also may suffer because the simple feature chosen is
not the best out of a family of potential simple features. Our approach, which we
describe in detail here, has been to keep feature selection as simple and straightforward
as possible, and then examine how brain dynamics change, with feedback,
over time. Future strategies will have to take both approaches into account.
Adaptive feature techniques, which dynamically change the parameterization
between brain signal and feedback, represent another approach, where the machine
iteratively “learns” the signals that the subject is attempting to use during the BCI
task. The potential advantage of such adaptive techniques is that they might be
robust against non-stationarity in the distribution of cortical spectral change and
compensate for shifts in the signal. The disadvantage is that a subject may be trying
to adapt the signal at least as fast as the machine algorithm, and we have at times