Localizing the focus of epileptic seizures using modern measures from multivariate time series analysis

Abstract

In this diploma thesis, we present different methods to realize the visualization of the spatio-temporal evolution of an epileptic seizure's focus based on multichannel ECoG data. Furthermore, our methods allow for a precise localization of the seizure's initial focus, which is possible, as that focus exists because all our data stems from patients suffering from temporal lobe epilepsy.<br />Two main methods are presented, starting with a frequency domain approach. Based on the coefficients of a fitted AR-model, modern measures like partial directed coherence (PDC) are derived and discussed. Exhaustive analysis of PDC's problems further leads to a generalized and far better performing version of PDC. In the second part, a Recursive Least Squares (RLS) Algorithm is performed instead of the ordinary least squares approach in frequency domain. This RLS-Algorithm helps us to cope with the instationarities of the ECoG signals far better.<br />Based on the time-dependent AR-model, Granger causality is used to indicate interactions between channels of the multivariate signals.<br />Starting with Granger's basic idea and the analysis of coupling effects between two signals, stepwise partialization leads to improved results.<br />Generalized PDC as well as the Granger causality finally lead to a visualization of the seizure's evolution. The results presented are in excellent accordance with descriptions from clinical experts.