Exploring machine learning techniques in epileptic seizure detection and prediction
Abstract
Epilepsy is the most common neurological disorder, affecting between 0.6% and 0.8%
of the global population. Among those affected by epilepsy whose primary method of
seizure management is Anti Epileptic Drug therapy (AED), 30% go on to develop
resistance to drugs which ultimately leads to poor seizure management. Currently,
alternative therapeutic methods with successful outcome and wide applicability to
various types of epilepsy are limited. During an epileptic seizure, the onset of which
tends to be sudden and without prior warning, sufferers are highly vulnerable to injury,
and methods that might accurately predict seizure episodes in advance are clearly of
value, particularly to those who are resistant to other forms of therapy.
In this thesis, we draw from the body of work behind automatic seizure prediction
obtained from digitised Electroencephalography (EEG) data and use a selection of
machine learning and data mining algorithms and techniques in an attempt to explore
potential directions of improvement for automatic prediction of epileptic seizures. We
start by adopting a set of EEG features from previous work in the field (Costa et al.
2008) and exploring these via seizure classification and feature selection studies on a
large dataset. Guided by the results of these feature selection studies, we then build on
Costa et al's work by presenting an expanded feature-set for EEG studies in this area.
Next, we study the predictability of epileptic seizures several minutes (up to 25
minutes) in advance of the physiological onset. Furthermore, we look at the role of the
various feature compositions on predicting epileptic seizures well in advance of their
occurring. We focus on how predictability varies as a function of how far in advance
we are trying to predict the seizure episode and whether the predictive patterns are
translated across the entire dataset.
Finally, we study epileptic seizure detection from a multiple-patient perspective.
This entails conducting a comprehensive analysis of machine learning models trained
on multiple patients and then observing how generalisation is affected by the number of
patients and the underlying learning algorithm. Moreover, we improve multiple-patient
performance by applying two state of the art machine learning algorithms.