Epileptic Seizure Detection and Source Localization Based on Stockwell Transform

No Thumbnail Available



Journal Title

Journal ISSN

Volume Title


Addis Ababa University


Neurologists often have to scan long term electroencephalogram (EEG) recordings in order to diagnose epilepsy. Detection of seizures from recorded EEG signal is crucial for diagnosis of epilepsy. Localization of the seizure origin is also important for treatment and surgery of focal epilepsy cases. Visual scanning of EEG is time consuming and suffers from issues of subjectivity due to imprecise definition of abnormal seizure EEG patterns. EEG recordings between seizures or inter-ictal EEG findings also offer evidence of epilepsy though not decisive as observed epileptic seizures. Although the main task is detecting seizures, the accuracy of a seizure detection scheme is based on clear characterization of inter-ictal, seizure and normal EEG’s. The challenge here is that abnormal patterns of EEG signals from epilepsy patients are case specific, especially for inter-ictal EEG. Additionally, EEG signals are composed of multiple frequencies and hence non-stationary. This thesis majorly considers temporal lobe epilepsy. An automated EEG signal classification scheme has been proposed for use in efficient detection and source localization of epileptic seizures based on the Stockwell (S) transform. Important features were extracted from the S-transform plane of EEG segments to categorize them into seizure, inter-ictal and normal signals. Classification of the features was done using support vector machine. For classification problem between seizure and normal EEG (recorded with closed eyes and/or open eyes), 100% sensitivity, 100% specificity and 100% accuracy were obtained. For classification between seizure and inter-ictal EEGs recorded from the epileptogenic zone, the proposed scheme achieved 99 % sensitivity, 99% specificity and 99 % accuracy. For seizure and inter-ictal signals recorded from non-epileptogenic zone, the classification scheme resulted in 99% sensitivity, 98% specificity and 98.5% accuracy. Empirical mode decomposition was also employed to improve the performance of the classification between seizure and non-seizure dataset. The methodology used for seizure detection was also employed to automated epileptic focus localization. The features extracted for source localization were intended to characterize focal and non-focal signals. A scatter plot was generated using the features and simple thresholding was able to classify focal and non-focal EEGs with 84% sensitivity, 90.21% specificity and 88% accuracy. The proposed method uses fewer number of features resulting in smaller feature space which in turn makes it simple and robust compared to other schemes proposed in the literature.



Epileptic Seizure, Focus Localization, Stockwell Transform, Empirical Mode Decomposition, Vector Support Machine