An EEG-based Method for Drowsiness Level Estimation

Abstract

Obtaining accurate estimates of a driver's level of drowsiness to help develop non-invasive methods for drowsiness detection is a challenging and open research problem. Many approaches to drowsiness or sleepiness estimation are supervised machine learning ones that require accurate labels for their sensor data to train a model. In this work, a novel method is presented to annotate time-series data with a driver's estimated level of drowsiness using characteristics from the electroencephalogram (EEG). The proposed scoring algorithm assigns a value between one and ten to segments of EEG data corresponding to a driver's predicted response on the Karolinska Sleepiness Scale (KSS). The parameters of the scoring algorithm are tuned using a metaheuristic optimization algorithm called Late-Acceptance Hill-Climbing and a loss function that utilizes the driver's own KSS ratings. Promising qualitative results have been presented for the proposed method to estimate a person's level of drowsiness on a more granular timescale than traditional survey methods like KSS. Furthermore, the approach could be extended beyond drowsiness estimation to any task involving the need to make use of EEG data between event markers or annotations. In addition, the data acquisition process that was employed in this work is described along with the database created.