Epilepsy is a disorder that a considerable number of people suffer from. Even though seizures are highly researched there is no known cure for the disorder apart from the resection of the epileptogenic zone which works in a limited number of cases. The aim of modern neuroscience is to automate the procedure of predicting if an EEG signal presents seizure or healthy activity, with the highest accuracy possible. Working in that framework, what this thesis presents, is initially the research on the existing literature by displaying the best classification schemes and trials. Visualizations, plots, and tables were displayed, using R, in order to give better perception of the data. In the next step of the study a group of algorithms and classifiers were presented and tested using extensively Weka 3.8 data mining tool. Tree algorithms were proven to have the best performance on predicting if the signals were ictal or healthy and the average accuracy achieved by the trees group was 86% while the next best group had 81% average accuracy. The best accuracy received by one algorithm was 94.7% given by the Random Committee algorithm, which used as classifiers of the ensemble decision trees. Finally, in the study it is proposed that the Tree algorithmic model fits best the data and presents remarkable predictions especially if we consider that the data had undergone only limited preprocess and no normalization at all.
Collections
Show Collections