Evaluation of techniques for drowsiness detection

In recent years, an increasing number of fatigue-tracking technologies have become available with the widespread hope that they will be a key component in the prevention of fatigue-related accidents. To obtain estimates of the current scientific validity of six promising fatigue detection technologies, a double-blind, controlled laboratory validation experiment was undertaken. The six technologies tested included a video-based scoring of eye closure by trained observers; two EEG algorithms; a head tracking device; and two eye-blink monitors. Psychomotor vigilance task (PVT) performance lapses were selected as the validation criterion variable. Fourteen healthy adult males remained awake in the laboratory for 42-hr and performed a computerized test battery every 2-hr that included a 20-min PVT task, the criterion variable. Each technology was time-locked to PVT performance to test coherence between vigilance lapses and each technology's specific drowsiness metric, over both 1-min and 20-min intervals. All technologies performed better when predicting lapses over a 20-min period than over a 1-min period and all of the technologies showed potential for the detection of drowsiness-induced hypovigilance by accurately tracking the relative profile of PVT lapses in at least one subject or a subset of subjects. However, only one technology, PERCLOS-the video-based scoring of slow eye closure by trained observers (Wierwille et al., 1994; Wierwille & Ellsworth, 1994)-correlated highly with PVT lapses both within and between subjects (mean r = 0.875, p = 0.00001). PERCLOS also correlated more highly with PVT lapses than subjects' own ratings of their sleepiness (t = -3.9, p = 0.003). The results of this study revealed that although all of the technologies tested had potential to predict hypovigilant performance on a continuous visual vigilance task of sustained attention, only PERCLOS actually performed with a high level of validity and reliability, suggesting that PERCLOS has the greatest potential to be used as a valid fatigue-tracking device since it was able to consistently track the relative profile of PVT lapses of attention for all subjects.