A Comparison among Different Supervised Parameters-Tuned Machine Learning Algorithms with Application to Epileptic EEG Signal Classification

Abstract

This study investigates the use of advanced machine learning algorithms to classify different classes of epilepsy EEG signals. Using the TSFRESH package, a large set of features was reduced to the most important ones by extracting them from the University of Bonn EEG dataset. Following that, seven distinct machine learning classifiers were trained using these features. Three different classification tasks are used to evaluate the classifiers using cross validation. The evaluation results of each algorithm after testing them on each of the three tasks show that the accuracy of the Support Vector Machine (SVM) reaches 99.60% for binary classification, the accuracy of Histogram-based Gradient Boosting Trees (HGBT) reaches 99.20% for the three-class task, and for the five-class classification, the Random Forest achieves 94.80% accuracy, which is the highest among the others. These results show that there is no dominant algorithm that works for all classification tasks and it is necessary to always run more than one algorithm to get a better assessment.