Advances in epileptic seizure onset prediction in the EEG with ICA and phase synchronization
University of Southampton, Institute of Sound and Vibration Research,
Seizure onset prediction in epilepsy is a challenge which is under investigation using
many and varied signal processing techniques, across the world. This research thesis
contributes to the advancement of digital signal analysis of neurophysiological signals
of epileptic patients. It has been studied especially in the context of epileptic seizure
onset prediction, with a motivation to help epileptic patients by advancing the
knowledge on the possibilities of seizure prediction and inching towards a clinically
viable seizure predictor.
In this work, a synchrony based multi-stage system is analyzed that brings to bear
the advantages of many techniques in each substage. The 1st stage of the system
unmixes and de-noises continuous long-term (2-4 days) multichannel scalp
Electroencephalograms using spatially constrained Independent Component Analysis.
The 2d stage estimates the long term significant phase synchrony dynamics of
narrowband (2-8 Hz and 8-14 Hz) seizure components. The synchrony dynamics are
assessed with a novel statistic, the PLV-d, analyzing the joint synchrony in two
frequency bands of interest.
The 3rd stage creates multidimensional features of these synchrony dynamics for two
classes (‘seizure free’ and ‘seizure predictive’) which are then projected onto a
2-dimensional map using a supervised Neuroscale, a topographic projection scheme
based on a Radial Basis Neural Network. The 4th stage evaluates the probability of
occurrence of predictive events using Gaussian Mixture Models used in supervised
and semi-supervised forms.
Preliminary analysis is performed on shorter data segments and the final system is
based on nine patient’s long term (2-4 days each) continuous data. The training and
testing for feature extraction analysis is performed on five patient datasets. The
features extracted and the parameters ascertained with this analysis are then applied
on the remaining four long-term datasets as a test of performance. The analysis is
tested against random predictors as well. We show the possibility of seizure onset
prediction (performing better than a random predictor) within a prediction window
of 35-65 minutes with a sensitivity of 65-100% and specificity of 60-100% across the
Actions (login required)