Deep Neural Network for Drowsiness Detection from EEG

Abstract

This study aimed to detect drowsiness and find optimal electrode set by collecting and classifying the EEG dataset labeled with three classes: awakeness, drowsiness, and sleep. Blindfolded subjects were presented short audio stimulus in random duration and instructed to push button according to audio stimulus. For classification of 3 classes, EEG signal was segmented and labeled according to the sequence of button response. Then, segmented data were directly fed into classifier without further transformation. Training was accomplished by the proposed deep learning including four LSTM layers. The proposed drowsiness detection deep learning network resulted 82.8% accuracy with 18 channels, and 79.8% accuracy with 3 channels located at premotor area of right hemisphere. © 2021 IEEE.