Feature extraction method for classification of alertness and drowsiness states EEG signals

Drowsy driving is one of the major causes of road accidents. The road accidents can be avoided by the discrimination of alertness and drowsiness states of the drives. The neurological changes in alertness and drowsiness states can be asses by electroencephalogram (EEG) signals. In this paper, the non stationary characteristic of the EEG signal is explored by tunable Q-factor wavelet transform (TQWT). TQWT decomposes the EEG signal into sub-bands, which further used for the extraction of features. Statistical features of the Hjorth mobility such as minimum value, maximum value, mean and standard deviation (SD) are used for characterization of the alertness and drowsiness states. Various classifiers such as decision tree, logistic regression, fine Gaussian support vector machine, weighted KNN, ensemble boosted trees and extreme learning machine (ELM) are considered. The alertness and drowsiness EEG signals discriminative performance of TQWT-based features are assessed by the Kruskal-Wallis (KW) test. The results of KW-test show that the proposed features are effectively discriminative of the alertness and drowsiness states. According to the obtained results, the best accuracy score of 91.842% is produced by the ELM classifier. (C) 2020 Elsevier Ltd. All rights reserved.