Comparing The Effect of Under-Sampling and Over-Sampling on Traditional Machine Learning Algorithms for Epileptic Seizure Detection

Abstract

Epilepsy disease, a neurological disorder that causes recurrent and sudden crises, occurs at unforeseen times. This study presents the classification of electroencephalogram signals for epileptic seizure prediction. The performances of the machine learning algorithms were evaluated on the dataset extracted from electroencephalogram signals. The dataset consists of brain activities for 23.5 seconds of 500 individuals with each has 178 data points for one second, and totally of 11500 pieces of information. In this study, since the aim was to develop a model to predict epileptic seizure, the problem was transformed into a two-class problem by combining target categories except than epileptic seizure. Since combined target categories made the dataset unbalanced, Random Under Sampling and Random Over Sampling methods were applied to prevent the machine learning algorithms from overfitting the dominant class. Thus, each of the three datasets was divided into training and test sets by ratios of 60/40, 70/30, 80/20. The performance of the several machine learning algorithms were evaluated and discussed through three different scenarios. Overall results showed us that Random Forest algorithm offered superior performance than others for all scenarios in terms of accuracy, sensitivity and specificity metrics.