Information Technology Reference
In-Depth Information
the P3a with such devices using the average technique. Although the P3a is less
useful for creating speller-type applications, it can be used with certain approaches
where only the oddball response is required. In addition, we have had some success
in using machine learning to build a P300 classi
er using the NeuroSky, and these
results are presented at the end of this chapter.
Our system is based on our generalised C++ ERP detection algorithm previously
described in [x]. This algorithm is agnostic to hardware, providing the same code
interface for any device for which raw EEG data is available. There are some
important aspects to the design of the algorithm which are worth noting if you are
considering engineering your own P300 solution.
When producing custom signal processing techniques for ERP detection, it is
vital that the signal is conditioned in the correct manner. Failure to do this will
result in false and/or confusing detection rates. Providing that your hardware is
correctly con
gured and operating normally, the derivation of a proper baseline for
the raw EEG signal is the next priority. A good baseline is essential before any
signal processing can begin
fundamentally if there is any positive or negative
going offset in the signal, this will cause biased results, especially when averaging
signals.
For example, if the baseline signal is positively biased, this will introduce a
higher average peak in results that have a smaller number of signal blocks from
which to derive an average. The offset will be reduced as more signal blocks are
used to derive an average, but in the case of an oddball test, where the less common
signal should contain a higher average than the more common signal, the positive
offset biases the entire result.
In order to avoid these sorts of statistical anomalies, one can either use a high
pass IIR
—
filter, or subtract a continuous average signal made up of the last n
samples. Neither approach is without
aws. An IIR
filter may well introduce phase
shift in signi
cant areas of the signal, whereas subtracting the average will remove
more than just the offset. This may well help rather than hinder your ERP detection
—
but either way it is a choice the reader must make for themselves.
3.8.2 The P300 Moving Average Method
Building on our baseline method, we created a modi
cation of the standard ERP
paradigm to allow control of directional movement within a 3D virtual world
through continuous control. In order to achieve this, we created a windowed
moving average algorithm that reliably detects ERP signals from raw EEG data in a
continuous fashion, eliminating the need to stop the test in order to reach a decision.
This has many applications in BCMI, for example, where continuous control is
required to adjust parameter values.
In cases where the hardware or signal quality is poor, for example when using a
low-cost consumer EEG system such as the NeuroSky MindSet, this approach can
help to improve signal conditioning in a way that is more satisfying and less time
