Early Seizure Onset Prediction in Clinical Recordings of EEG

Abstract

We apply and evaluate a new variation of Independent Component Analysis (ICA) to the problem of seizure onset analysis in electroencephalographic (EEG) signals in epilepsy. This constrained ICA (CICA) algorithm takes a reference signal as input, along with observed multichannel data, and extracts the independent source which is closest, in some sense, to the reference. We present an implementation of this algorithm whose behaviour has been initially tested using a simple data set. The performance of the algorithm for this application is then assessed using synthetic seizure waveform mixed with real-world EEG epochs, including some common artifacts. The application of the algorithm to both artifact rejection and the extraction of seizure waveform from real EEG recordings of seizures is shown to be possible. For seizure onset analysis, the required a priori reference signal can be based on the frequency and phase of a period of known seizure waveform. The CICA algorithm can then be used to extract seizure waveform from time periods prior to the clinical onset of the seizure. The algorithm implicitly calculates an estimated spatial distribution of the extracted source, which can indicate the focus of a seizure.