Study of Single-Channel EEG Signal Analysis for Drowsiness Detection using Machine Learning

Abstract

Electroencephalogram (EEG) is an essential tool used to analyze the activities effectively and different states of the brain. Drowsiness is a short period state of the brain that is also called an inattentiveness state. Drowsiness can be observed during the transition from being awake state to a sleepy state. Drowsiness reduces a person’s attention that increases accidental risks when involved in their personal and professional activities like vehicle driving, operating a crane, working with heavy machineries such as mine blasts. Drowsiness Detection (DD) has a significant role in preventing the problems mentioned above. So many traditional algorithms are proposed to detect drowsiness, but among these, the combination of neuroscience with artificial intelligence can effectively diagnose the state of drowsiness. Neuroscience with artificial intelligence algorithms used to detect drowsiness is also popularly known as brain-computer interface (BCI) systems.
Single-channel EEG BCIs are highly preferred for convenient use in real-time applications, even though there are many challenges in the actual experimental process. They are feature extraction, feature selection and choosing the best channel. These challenges have badly affected the performance of the BCI in the detection of drowsiness. In this work, a novel channel selection approach is proposed for a single-channel EEG BCI system by integrating the statistical characteristics of the available channels EEG signal to detect drowsiness state successfully. This thesis addresses some EEG sub-band extraction methods and their limitations. These limitations and practical issues are overcome by proposing a time-domain sub-band based feature extraction procedure using the wavelet packet transformation method. This thesis also addresses the asymmetric feature interference between the subjects. These limitations are overcome by proposing a novel feature selection technique using a nonparametric statistical test. In addition to this, a novel single-channel EEG signal analysis approach and single feature computation are also offered to deploy most quickly on low computing capacity systems. Apart from the machine learning methods, this thesis also discusses a novel deep learning architecture based on a convolutional neural network (CNN) for automated single-channel EEG signal classification to detect drowsiness
Subject wise, cross-subject wise, and combined subject’s wise validations are also employed to improve the generalization capability of the proposed techniques in this thesis. The whole work is carried out over prerecorded EEG databases such as Physionet real-time sleep-analysis-EEG and simulated-virtual-driving-driver-EEG.