| Summary: | This thesis is an investigation of the Outlier Processing Method, which was developed
to eliminate the dependence of conventional techniques on a priori information for the
detection of event-related potentials (ERPs) from EEG signals. Instead of attempting to
model the ERP, the OPM assumes the ERP to be an outlier signal which is imbedded
in a background EEG signal. The background EEG is modelled as an auto-regressive
signal and approximated using robust statistical parameter and signal estimation. The
ERP estimate is obtained by subtracting the estimate of the background EEG signal from
the observed EEG signal.
The basis of the robust statistical parameter estimation used in OPM was the
generalized maximum likelihood (GM) estimate. It required the generation of initial
parameter estimates to start the process. MEM, a least squares estimator used in the
original OPM, was found to be a more consistent estimator when compared to median
regression estimation. Also, a method of automatically generating the tuning constants
used by the influence functions in the GM estimate was proposed.
, The robust signal estimation was a robust variation of a Kalman filter which required
a cleaner function to minimize the effect of outliers such as ERPs. Of the three different
cleaners that were evaluated, the fixed-lag cleaner had the best overall performance.
A method of detecting the location of the outlier was developed and was found to be
effective for signals with low SNRs and a large amount of spectral overlap between the
outlier and the background EEG signal.
|
|---|
