Contributor:
Ray Smith BE MIEI
Dept. of Electronic and Electrical Engineering
University College Dublin,
Belfield, Dublin 4
Rep. of Ireland
Supervisior:
Dr. Richard Reilly
DSP Research Group, UCD
Description of Methodology
Preprocessing:
Utilizing the 1000Hz trials each epoch was LPF filtered (fc=45Hz) and
then downsampled to 100Hz. Each epoch was then zero-meaned.
Channel Selection:
Cross correlative analysis between L & R trials proved that electrode
positions located above the primary motoral cortex maximally
distinguished between trials, thus supporting physiological
beliefs. Depending on the feature extraction technique employed, an
optimum subset of the optimum electrodes positions found (FC1 FC2 C1
C2 C3 C4 CP1 CP2 etc) were chosen based on classification accuracy of
the training data. (10 shuffle by 10 fold cross validation). Overall
aim here is to reduce Feature Vector dimensionality in the hope of
developing a simple classifier capable of real time implementation
Feature Selection:
A collection of time and frequency domain features associated with
motoral movement were investigated. The overall scheme was a weighted
contribution of the various techniques
ARX
An Autoregressive with Exogenous input (ARX) model for combined
filtering and feature extraction was employed as a natural extension
to an Autoregressive (AR) parametric modelling technique. The
exogenous input is an ensemble average of all trials for a given
channel specific to an upcoming left or right movement. The ensemble
average is represenative of the underlying Bereitschaftspotential
(BP), a gradually rising negative potential focused above the motoral
cortex occurring about 1000ms preceeding the onset of movement, as the
random background EEG activity averages out about the mean. The ARX
case of modelling both the signal, the underlying BP, and the noise,
background EEG assuming no artifacts, is more physiologically
reasonable and yields better results than AR modelling alone. (85%
compared to 60%)
Power Band Analysis (ERD/ ERS)
The mean power in two bands, extensions of the mu and beta bands
(8-16Hz and 17-25Hz) for selected channels above the primary motoral
cortex. Each epoch was windowed (hanning) and then the fft was taken
to estimate the PSD.
Time - Frequency Analysis
Estimation of the power in the extensions of the mu and beta bands
(8-16Hz and 17-25Hz) over overlapping windows 40% the length of the
epochs using a STFT.
Classifier
A simple linear discriminant function (approx. to Bayes rule) was
employed to classify the feature vectors. The computational simplicity
of this approach is a great advantage.
Training Data Accuracy using 10 shuffle by 10 fold cross validation was 84%.