EEG signal analysis for epilepsy diagnosis

Emil Darío Vega-Gualán, Yachay Tech University, Urcuqui - Ecuador, 2019

The epilepsy disorder occurs when the localized electrical activity of neurons suffers from an imbalance. Epilepsy has become the third most common neurological disorder after stroke and dementia -it is believed that affects 0.5 - 1.5% of the world population. It mainly affects children under 10 and people over 65, being more common in developing countries and in disadvantaged socioeconomic classes. Its possible diagnose is via the analysis of electroencephalographic (EEG) signals. Nowadays, since both its appropriate diagnosis and the accurate epileptic source localization must be fulfilled, computational systems are used to support the diagnosis procedure. Broadly, such systems perform the automatic diagnostic-assistance into four main stages, namely: EEG signal acquisition, preprocessing, characterization and classification. Once acquired and preprocessed, EEG signals must be properly represented to be subsequently classified into diagnostic categories (absence or any level of presence of seizure activity). Despite there exists a wide range of alternatives to characterize and classify EEG signals for epilepsy analysis purposes, many key aspects related to the accuracy, computational cost, and physiological interpretation are still considered as open issues. In this connection, in this work, an exploratory study of EEG signal processing techniques is proposed, aimed at identifying the most adequate state-of-the-art techniques for characterizing and classifying epileptic seizures. To do so, a comparative study is designed and developed on several subsets of features (namely, statistical measures on both the original signals and the spectral transformation thereof), as well as some representative classifiers (linear discriminant analysis classifier (LDC), quadratic discriminant analysis classifier (QDC), k-nearest neighbor (kNN) and support vector machine (SVM)). Proposed system validation is carried out by means of an exhaustive experimental setup over a gold standard database from the UCI Machine Learning Repository, so-named: "Epileptic Seizure Recognition Data Set". As remarkable results, it is experimentally proved that a characterization process based on statistical indices from wavelet-transform-driven decompositions, and the support vector machines as classifiers are the most suitable approaches for designing an automatic system to identify epilepsy-diagnosed EEG signals. As well, the overall performance of the obtained pattern recognition system (for the bi-class scenario) -in terms of confusion matrix-based measurements- amounts 96%, 85% and 98% of classification performance, sensitivity, and specificity, respectively.

Graphic Interface

The interface application allows replicating the results in an interactive way. The application is very intuitive and practical. It has four main sections: features, classification, plotting, and tables. And four external buttons: Load data, exit, about and manual. The figure above shows how this graphic interface looks like.

To replicate the experiments first it is necessary to load the original dataset called "Epileptic seizure recognition data set" from the corresponding path where the file is located. Then, in the feature section, the extraction and selection of features are performed. In the classification section, it is necessary to choose the experiment to replicate. After, with the classification button, the process is started. When the process is finished, in the plotting section you can be able to generate the box plot or ROC curve of any experiment. Finally, in the tables section, you can generate a table of any experiment with the statistical measure that you choose. A video-tutorial is attached below for a better understanding of how it works.


Source Code


 Emil Darío Vega Gualán
Information Technology Engineer
 Curriculum Vitae