Epilepsy Detection and Monitoring - Quantitative EEG Analysis Methods and Clinical Applications - page 164

Modeling and Analysis Control and Design

Graph partitioning
Support vector machine
Self-organizing map
K-mean
Piecewise affine map

Markov state machine

Discrete state model

Control
information

Features

Parametric models
Nonparametric models

Feature
extraction

Interface

Simulator

Pattern selection

Simulating
pattern

EEG

Epileptic brain

Figure 6.7 A hybrid system that is composed of four parts of modeling phases: modeling, analysis,
control, and design.

multivariate tools at their disposal. Even with these tools, the richness of the datasets
has meant that these techniques have been met with limited success in predicting sei-
zures. To date, there has been limited amount of research into comparing techniques
on the same datasets. Oftentimes the initial success of a measure has been difficult to
repeat because the first set of trials was the victim of overtraining. No measure has
been able to reliably and repeatedly predict seizures with a high level of specificity
and sensitivity.
While the line between seizure prediction, early detection, and detection can
sometimes blur, it is important to note they do comprise three different questions.
While unable to predict a seizure, many of these measures can detect a seizure. Sei-
zures often present themselves as electrical storms in the brain, which are easily
detectable, by eye, on an EEG trace. Seizure prediction seeks to tease out minute
changes in the EEG signal. Thus far the tools that are able to detect one of these
minor fluctuations often fall short when trying to replicate their success in slightly
altered conditions. Coupled with the proper type of intervention (e.g., chemical
stimulation or directed pharmacological delivery) early detection algorithms could
usher in a new era of epilepsy treatment. The techniques presented in this chapter
need to be continually studied and refined. They should be tested on standard
datasets in order for their results to be accurately compared. Additionally, they need
to be tested on out-of-sample datasets to determine their effectiveness in a clinical
setting.