Combined utilization of objective and subjective data as determinants of an oculometric index for the non-invasive detection and prediction of operator fatigue and drowsiness

Sleep-deprivation affects people in their jobs, which can create hazardous environment for the public. Alternative actions must be taken to prevent the devastating effects of operators falling asleep during critical procedures.;This dissertation attempted to overcome most of the technological and methodological limitations of systems detecting real-time operator fatigue and drowsiness. It used a newly developed, non-invasive, wearable oculometric technology to determine whether or not a set of oculometric measures can exhibit sufficient sensitivity and specificity in predicting the onset of drowsiness. It also attempted to track an operators' impaired performance related to fatigue.;The IRB-approved study involved20 healthy volunteers. They were evaluated during a 24-hour sleep-deprivation cycle and performed three 5-hour blocks of sequenced tasks. Test blocks included the Stanford Sleepiness Scales (SSS), Epworth Sleepiness Scales (ESS) and the Maintenance of Wakefulness Test (MWT), which was performed in synchronization with Oculometrics and Pulse Oximetry. A driving simulator session was conducted, synchronized with Electroencephalographic (EEG) and Pulse Oximetry data. At the end of each block, a Simulator Sickness Questionnaire (SSQ) was given before a 3-hour break.;More than half of the volunteers did not report any SSQ symptoms. Severe symptoms were reported by fewer than 4%. There was a significant increase in moderate and severe discomfort associated with fatigue due to drowsiness rather than any visual-vestibular conflict produced by the simulator. Both subjective and objective tests showed significant trends, confirming the sleep deprived states of participants. Non-parametric tests revealed significant differences in volunteers' own perception of sleepiness (SSS) and in sleep propensity (ESS) with p < .0001 for both tests.;Repeated measure ANOVA showed that mean MWT latency differed statistically significantly (p = .002), with significant linear decrease; driving performance variables showed significant within-subject effects (p = 0.01), with significant univariate effects for Tailway Distance, Velocity, and Collision. Similar analysis indicated that mean Heart Rate (HR) had significant differences ( p = .015), with a significant quadratic effect due to a slightly higher HR in Session 2 compared to Sessions 1 and 3. SPO2 did not show a statistically significant difference.;Percentage of time the eyes were closed (PERCLOS) and eye blink frequency (EBF) and duration (EBD) were sensitive to sleep deprivation effect with highly significant results (p < .0001), showing most predominant changes at session 3. EEG did not show clear significant trends during the driving sessions.;Averaged oculometric data as regressors for the driving sessions proved to be inefficient for predicting collisions and drowsy events. Accident data for these models did not necessarily represented crashes due to a drowsy event (categorized by microsleep). While all the EEG-related microsleep events were detected only during Session 3, not all of those microsleeps were associated to a crash accident or closed eyes.;To assess whether a set of oculometric measures can exhibit sufficient sensitivity and specificity in predicting the onset of operator drowsiness, volunteer crash accidents directly related to falling asleep at the wheel were identified. Datasets with different time-segments before a crash accident and a dichotomous dependent variable were created to allow the use of binary logistic regression to model the probability that a volunteer driver will fall asleep at the wheel and, as a consequence, have a crash accident. This approach yielded some interesting outcomes related to the use of PERCLOS and real-time EBD, suggesting that 1) EBF may have large individual variability that confound the prediction of drowsiness related fatigue, 2) PERCLOS, as an averaged function, may be more useful in predicting a "tendency" rather than an "imminent" impel to fall asleep at the wheel, and 3) real-time EBD could be a more powerful indicator of an immediate drowsy event that most likely will end in a crash accident.;Results had also determined that, as the time prior to falling asleep at the wheel gets closer, the models tend to reduce its predictive probability when PERCLOS is used as predictor. Conversely, when real-time EBD is used, the models tend to increase its predictive probability.;Based on these results, an initial portable drowsiness detection system can be designed as a two-level alarm system based on a combination of spontaneous-blink free PERCLOS calculation to determine drowsiness propensity, and a real-time EBD calculation to determine an imminent drowsy event. To this end, some recommendations were developed as part of the conclusion for this dissertation.;Although logistic regression models can provide a relatively good predictive algorithm, individual variability among the users can be detrimental in the specificity and sensitivity levels required for implementation. As group means can be proven insufficient, the use of a known distribution will allow the application of statistical decision theory to construct optimal decision rules minimizing risks.