Automatic Seizure Detection in ECoG by Differential Operator and Windowed Variance

Abstract

Differential operator has long been used in image and signal processing with great success to detect significant changes. In this paper we show that differentiation can enhance certain features of brain electrophysiological signals, contaminated with noise, artifacts, and acquisition defects, leading to efficient detection of those changes. Windowed variance method has been very successful in detecting seizure onset in the brain electrophysiological signals. In this paper we have combined these two powerful methods under the name of differential windowed variance (DWV) algorithm to automatically detect seizure onsets in almost real time, in continuous ECoG (depth-EEG) signals of epileptic patients. The main advantages of the method are simplicity of implementation and speed. We have tested the algorithm on 369 h of nonseizure ECoG as well as 59 h of seizure ECoG of 15 epileptic patients. It detected all but six seizures (91.525% accuracy) with an average delay of 9.2 s after the onset with a maximum false detection of three in 24 h of nonseizure data. Eight novel empirical measures have been introduced to avoid false detections. To ascertain the reliability of the detection method a novel methodology, called quasi-ROC (qROC) curve analysis has been introduced. DWV has been compared with a difference filter based sharp transient (ST) detection algorithm.