Proposal of a Method to Predict Subjective Rating on Drowsiness Using Physiological and Behavioral Measures

Abstract

OCCUPATIONAL APPLICATIONS Subjective drowsiness was predicted during a simulated driving task with an accuracy of more than 90%. This was done using a multinomial logistic regression model, using physiological and behavioral measures as predictors. The actual and/or potential applications of these results include the development of a system for predicting drowsiness and presenting drivers a warning. These results can contribute to the enhancement of transportation safety by decreasing the risk of crashes or traffic accidents caused by drowsy driving. TECHNICAL ABSTRACT Background: From the viewpoint of automotive safety, it is useful to detect a decrease in arousal level and to warn drivers of the risk of a traffic accident. Although many measures of drowsy states have been developed, effective methods for predicting drowsy driving states and to warn drivers of these states have not been established. Purpose: The aim of this study was to explore the effectiveness of physiological and behavioral evaluation measures for predicting a drivers' subjective drowsiness using a regression model. Methods: Eight participants completed the study, which involved simulated driving. They were required to steer and maintain their vehicle at the centerline and to maintain the distance between their own car and a preceding car. Physiological measures were obtained (electroencephalography, heart rate variability and blink frequency), along with behavioral measures (neck bending angle, back pressure, foot pressure, and tracking error), and participants reported subjective drowsiness once every minute. Drowsy states were predicted via three multinomial logistic regression models consisting of different independent variables—Model A: both physiological and behavioral measures, Model B: only behavioral measures, and Model C: only physiological measures. For each model, prediction accuracies were examined, and the length of the data window used for predicting drowsiness was explored. Results: When both physiological and behavioral measures were used, prediction accuracy was 96.8%. The interval used for attaining the highest prediction accuracy was 100 seconds (from 120 to 20 seconds before the prediction). When only physiological measures were used, prediction accuracy was 90.2%, and accuracy was 94.9% using only behavioral measures. Conclusions: The proposed multinomial model could attain higher prediction accuracy when both physiological and behavioral measures are used and is potentially useful for the development of drowsiness warning systems.