Brain–Computer Interfaces - Index of

A Simple, Spectral-Change Based, Electrocorticographic Brain–Computer Interface 255
each cue. There were 30 cues each for hand and tongue, and they were interleaved
in random order to reduce the influence of anticipation. Following the 6 min of overt
movement, an identical task was performed, except that rather than actually move
the hand and tongue, she imagined moving her hand and tongue. She was explicitly
instructed to imagine the kinetics of the movement, not just the sensation (this
kinesthetic imagery was found to produce the most robust signal change in a study
by Neuper et al. [24]). A comparison of movement and rest periods (as shown in
Fig. 4) produced feature maps for movement and imagery, shown in the top row
of (a). These reveal that changes during imagery mimic those of overt movement,
but are less intense (note that the color bar scales are different for movement than
imagery). Cortical projections of the imagined and real changes for both high (76–
100 Hz) and low (8–32 Hz) frequency changes reveal that the spatial distributions
are similar, but the imagery associated change is more focal [36]. The number to
the top right of each cortical projection indicates the maximum absolute value of
the activation, since each brain is scaled to the maximum. Electrodes’ weights with
statistically insignificant change were not projected. These types of projections can
be a very useful sanity check to verify that the locations of features identified on a
feature map make sense anatomically. A similar screening for repetition of the word
“move” was performed, to select a feature for the imagery-based feedback shown in
part (b).
Figure 7b, One-dimensional cursor feedback: In the second stage, we provided
the subject with feedback based upon movement and speech imagery. The imagery
was coupled to cursor movement, as detailed in Fig. 5. Figure inset (b) demonstrates
the feature chosen, an 80–95 Hz frequency range from an electrode in primary
mouth motor area. The mean power, P0, lay between the mean power during speech
imagery and rest. If the mean power, during the task, was above this level (obtained
by speech imagery) then the cursor would move up, and if it was below, the cursor
would move down, according to the relation ˙y = g(P(t) − P0). The speed
parameter, g, was determined online prior to the first experimental run, such that
it was “reasonable” i.e. the cursor velocity changes on roughly the same timescale
as the electrode-frequency range feature. In theory, this could be estimated offline
prior to the first run by examining the variation in the power of the feature during
the screening task, but in practice, the rapid adjustment parameter built into the
BCI2000 program [15] is a much easier way to get a comfortable speed parameter.
The right-most portion of (b) demonstrates the activation during the 4 (~2 min)
trials of imagery based feedback. Each cortical projection represents the activation
between upper targets (speech imagery) and lower targets (rest); because the feature
chosen was at a high frequency, above the intersection in the power spectrum (shown
in Fig. 4), the power in the feature increases with activity. She rapidly learned to control
the cursor, with the signal becoming both more pronounced and more focused
in the feedback area. After the third run, she reported having gone from the coupled
imagery (imagined speech) to just thinking about the cursor moving up and down.
Again, only significant changes were plotted to the template cortex. The activity in
the most ventral posterior electrode (bottom right) hints at activity in Wernicke’s
area, and persists throughout (it appears to be less, but only because the primary