Epileptic-seizure classification using phase-space representation of FBSE-EWT based EEG sub-band signals and ensemble learners

Electroencephalogram (EEG) signals are non-linear and non-stationary in nature. The phase-space representation (PSR) method is useful for analysing the non-linear characteristics of EEG signals. Hence non-linear features based on a phase-space representation of EEG signals are effective in epileptic-seizure classification. In the past, various machine learning methods are used for classifying seizure EEG signals. However, they might fail to classify accurately due to complex data, so with effective non-linear features, ensemble learning classifiers can be investigated to improve the accuracy of an automated epileptic-seizure detection system. In this paper, the EEG signals are first decomposed in sub-bands using empirical wavelet transform (EWT) based on the Fourier Bessel series expansion (FBSE) which is termed as FBSE-EWT. Then, these sub-bands are reconstructed into threedimensional (3D) PSR. Next, entropy-based features like line-length (LL), log-energy-entropy (LEEnt), and norm-entropy (NEnt) are computed from Euclidean distances of 3D PSR of sub-band signals. The extracted features are ranked based on p-values obtained from the Kruskal-Wallis statistical test for reducing the feature space. Experiments have been conducted using obtained ranked features on different ensemble learning classifiers, and the five best-performing classifiers are reported here, which are random forest (RF), extra tree (ET), extreme gradient boosting tree (xgBT), bagged-SVM (B-SVM), and bagged-k-nearest neighbours (B-k-NN), for classifying epileptic-seizure EEG signals. The performance of the proposed framework is evaluated using a publically accessible Bonn university EEG database for classifying epileptic-seizure EEG signals on well-known classification problems such as C1 (seizure, normal), C2, and C3 (seizure, normal, and seizure-free). This dataset consists recording of 100 single-EEG signals from five seizure, seizure-free, and normal (healthy) subjects each. Model is trained and tested using 10-fold cross-validation to stave off from overfitting. The performance is also compared with other state-of-art methods. Obtained results confirm the superior performance of the proposed framework by achieving maximum classification accuracy of 100 %, 98.3 %, 97.8 % with ET, B-SVM, and ET classifiers from each studied classification problem, respectively. Hence, the proposed framework can assist medical professionals in analysing epileptic-seizure EEG signals more accurately.