Information Technology Reference
In-Depth Information
The speed of the system is to some extent reliant on the number of windows used
to create the moving average (n). Obviously, the greater the value of n, the more
reliable the system becomes. However, as the system uses a
rst out (FIFO)
queue, once the initial averages have been computed, it can take signi
first in
cantly less
time for the change in average amplitude to bring about a change in direction.
We carried out tests with
five participants in controlled conditions. Participants
were tested in a sound insulated environment with low light-levels. We asked
participants to attempt to concentrate on the forward moving arrow. We used this
measure in order to judge how effective the system was, based on the amount of
time that moving-average based EEG epochs relating to the forward arrow stimulus
contained the greatest average area.
Results from controlled experiments demonstrated that this test performs as well
as discrete P300 averaging approaches commonly used to create all BCIs, including
BCMIs. This is not controversial as the main difference with our method is that a
FIFO queue is used, and decisions are made by the system continually as the user
interacts. This approach has some clear advantages for creating real-time continu-
ous controllers for BCI generally, and can be used in BCMI for the control of
mixers and crossfaders. For example, with two
flashing arrows, one leftgoing, one
rightgoing, the user can attend to either arrow in order to improve a continuous
con
dence level being
equal to n averages. This scale can be used to assign a value to the user control, as
opposed to representing a speci
dence level in any particular direction, with a maximum con
c decision.
3.8.3 Template Matching Through Machine Learning of Repetitive
Serial Visual Presentation (RSVP) Derived P300 Features
RSVP (Craston et al.
2006
) is a variant on the oddball test. It offers the potential for
collecting much higher quality datasets for P300 classi
cation than the conven-
tional oddball paradigm. In an RSVP test, the participant is presented with a series
of symbols at high rate (in this case, 10 Hz) (Bowman et al.
2014
). They must try to
spot a symbol of one class in a stream of symbols from another. When the symbol is
recognised by the participant, an ERP will occur. At the end of the test, the par-
ticipant can be asked to identify the oddball symbol they spotted. The nature of this
answer (either correct, incorrect or unsure) tells the experimenter whether the
participant was attending to the task, giving a strong indication of whether a P300 is
present in the EEG signal in the time window following the oddball presentation.
This ground truth gives a signi
cant improvement in data quality compared to data
from the classic paradigm where this distinction cannot con
dently be made.
An online RSVP experiment was run. Participants were asked to spot a single
letter in a series of 30 numbers presented at 10 Hz, in repeated trials, while wearing
a NeuroSky headset. 412 trials were collected from 19 separate user sessions. The
data was preprocessed as follows: trials with bad signal quality (as reported by the
headset) were rejected. The rest were DC
filtered, and then the high frequencies
