Hybrid ensemble learning framework for epileptic seizure detection using electroencephalograph signals

Abstract

An automated method for accurate prediction of seizures is critical to enhance the quality of epileptic patients While numerous existing studies develop models and methods to identify an efficient feature selection and classification of electroencephalograph (EEG) data, recent studies emphasize on the development of ensemble learning methods to efficiently classify EEG signals in effective detection of epileptic seizures. Since EEG signals are non-stationary, traditional machine learning approaches may not suffice in effective identification of epileptic seizures. The paper proposes a hybrid ensemble learning framework that systematically combines pre-processing methods with ensemble machine learning algorithms. Specifically, principal component analysis (PCA) or t-distributed stochastic neighbor embedding (t-SNE) combined along k-means clustering followed by ensemble learning such as extreme gradient boosting algorithms (XGBoost) or random forest is considered. Selection of ensemble learning methods is justified by comparing the mean average precision score with well known methodologies in epileptic seizure detection domain when applied to real data set. The proposed hybrid framework is also compared with other simple supervised machine learning algorithms with training set of varying size. Results suggested that the proposed approach achieves significant improvement in accuracy compared with other algorithms and suggests stability in classification accuracy even with small sized data.