Brain–Computer Interfaces - Index of

A Simple, Spectral-Change Based, Electrocorticographic Brain–Computer Interface 247
3 Feature Selection
In order to implement a BCI paradigm, a specific signal feature must be chosen.
This will need to be a feature that can be determined in a computationally rapid
fashion. Second, the feature must be translated into a specific output. The choice of
signal feature should be an empiric one. There are two complementary approaches
to choosing a BCI feature. One approach is to start with a strictly defined task, such
as hand movement, and look for a particular feature at the signal change associated
with this task. Then, the most reliable signal is identified and used to run a BCI.
Another approach is to choose a signal that is less well characterized behaviorally
and then, over time, to allow the subject to learn to control the feature by exploiting
feedback, and then control the BCI. In a dramatic example of the latter, it was found
that the spike rate from an arbitrary neuron that grew into a glass cone could be
trained to run BCI [4], without necessary a priori knowledge about the preferred
behavioral tuning of the given neuron.
The most straightforward approach is a motor imagery-based, strictly defined,
task-related change for feature control. In order to identify appropriate simple features
to couple to device control, a set of screening tasks is performed. In these
screening tasks, the subject is cued to move or imagine (kinesthetically) moving
a given body part for several seconds and then cued to rest for several seconds
[6]. Repetitive movement has been found to be useful in generating robust change
because cortical activity during tonic contraction is quickly attenuated [19, 20].
Different movement types should be interleaved, so that the subject does not anticipate
the onset of each movement cue. Of course, there are multiple forms of motor
imagery. One can imagine what the movement looks like, one can imagine what the
movement feels like, and one can imagine the action of making the muscular contractions
which produce the movement (kinesthetic) [24]. It was demonstrated by
Neuper, et al. [24] that kinesthetic imagery produces the most robust cortical spectral
change, and, accordingly, we and others have used kinesthetic motor imagery as the
paired modality for device control. In order to establish that the control signal is truly
imagery, experimenters should exclude, by surface EMG and other methods, subtle
motor movement as the underlying source of the spectral change. In the screening
task, thirty to forty such movement/imagery cues for each movement/imagery type
should be recorded in order to obtain robust statistics (electrode-frequency band
shifts with significance of order p