Abstract
Two different problems of reflecting brain functioning are addressed.
This involves human performance monitoring during the signal detection task
and depth of anaesthesia monitoring. The common aspect of both problems is to monitor
brain activity through the electroencephalogram recordings on the scalp. Although these
two problems create only a fractional part of the tasks associated with physiological data
analysis the results and the methodology proposed have wider applicability.
- A theoretical and practical investigation of the different forms of kernel-based nonlinear regression models
and efficient kernel-based algorithms for appropriate features extraction is undertaken. The main focus is on
solving the problem of
providing reduced variance estimates of the regression coefficients when a linear regression
in some kernel function defined feature space is assumed. To
that end Kernel Principal Component Regression and Kernel Partial Least Squares Regression techniques
are proposed. These kernel-based techniques were found to be very efficient when observed data are mapped
to a high dimensional feature space where usually algorithms as simple as their linear counterparts in
input space are used. The methods are used and compared with existing kernel-based regression techniques
in measuring the human signal detection performance from the associated Event Related Potentials.
-
The depth of anesthesia (DOA) problem was addressed by assuming different complexity measures.
These measures were inspired by nonlinear dynamical systems and information theories. Data from
patients undergoing general anesthesia were used and the results were compared with traditional spectral
indices. The promising results of this pilot study suggest the possibility to include these measures
into the existing family of DOA descriptors. This opens a new area of more detailed and extensive research
into this very important medical problem.
Go back