A Nonlinear Penalty Driven Adaptive Thresholding Algorithm for Drowsiness Detection using EEG

Abstract

Drowsiness, leading to traffic and workplace accidents has been a persistent safety concern over years. Most of the electroencephalogram (EEG)-based drowsiness detection methods in literature use pre-trained classifier models. However, due to the non-stationarity of EEG signals, the patterns associated with drowsiness vary from subject to subject (inter-subject variability) and from session to session for each individual subject (intra-subject variability), necessitating an adaptive drowsiness detection algorithm. In this paper, an electroencephalogram (EEG) based drowsiness detection algorithm, that can adapt to the inter-subject and intra-subject variabilities is proposed. Drowsiness detection is performed based on a simple thresholding algorithm in which, session dependent thresholds are predicted adaptively using a regression model. The proposed drowsiness detection is done using a consumer grade wearable headband ensuring user comfort and the algorithm yields a better detection accuracy of 85.01 % compared to conventional classifier-based approach (83.15%). The proposed adaptive thresholding algorithm can effectively be used for drowsiness detection and is suitable for real time drowsiness detection since the thresholds are determined adaptively.