Liberty Mutual Research Institute for Safety, 71 Frankland Rd., Hopkinton, MA 01748, USA.
Liberty Mutual Research Institute for Safety, 71 Frankland Rd., Hopkinton, MA 01748, USA.
Accid Anal Prev. 2019 May;126:105-114. doi: 10.1016/j.aap.2017.11.004. Epub 2017 Nov 7.
The morning commute home is an especially vulnerable time for workers engaged in night shift work due to the heightened risk of experiencing drowsy driving. One strategy to manage this risk is to monitor the driver's state in real time using an in vehicle monitoring system and to alert drivers when they are becoming sleepy. The primary objective of this study is to build and evaluate predictive models for drowsiness events occurring in morning drives using a variety of physiological and performance data gathered under a real driving scenario. We used data collected from 16 night shift workers who drove an instrumented vehicle for approximately two hours on a test track on two occasions: after a night shift and after a night of rest. Drowsiness was defined by two outcome events: performance degradation (Lane-Crossing models) and electroencephalogram (EEG) characterized sleep episodes (Microsleep Models). For each outcome, we assessed the accuracy of sets of predictors, including or not including a driver factor, eyelid measures, and driving performance measures. We also compared the predictions using different time intervals relative to the events (e.g., 1-min prior to the event through 10-min prior). By examining the Area Under the receiver operating characteristic Curve (AUC), accuracy, sensitivity, and specificity of the predictive models, the results showed that the inclusion of an individual driver factor improved AUC and prediction accuracy for both outcomes. Eyelid measures improved the prediction for the Lane-Crossing models, but not for Microsleep models. Prediction performance was not changed by adding driving performance predictors or by increasing the time to the event for either outcome. The best models for both measures of drowsiness were those considering driver individual differences and eyelid measures, suggesting that these indicators should be strongly considered when predicting drowsiness events. The results of this paper can benefit the development of real-time drowsiness detection and help to manage drowsiness to avoid related motor-vehicle crashes and loss.
早晨的通勤回家路对于从事夜班工作的工人来说是一个特别脆弱的时间段，因为他们在这段时间内有更高的风险经历昏昏欲睡的驾驶。管理这种风险的一种策略是使用车载监控系统实时监测驾驶员的状态，并在驾驶员困倦时发出警报。本研究的主要目的是使用在实际驾驶场景下收集的各种生理和性能数据，构建和评估早晨驾驶时发生困倦事件的预测模型。我们使用从 16 名夜班工人收集的数据进行了研究，这些工人在两次情况下在测试轨道上驾驶了大约两个小时的仪器化车辆：一次是在夜班后，一次是在一夜休息后。困倦通过两个结果事件定义：性能下降（车道偏离模型）和脑电图（EEG）特征睡眠事件（微睡眠模型）。对于每个结果，我们评估了包括或不包括驾驶员因素、眼睑测量和驾驶性能测量的预测因子集的准确性。我们还比较了使用与事件相关的不同时间间隔的预测（例如，事件前 1 分钟至 10 分钟）。通过检查接收者操作特性曲线（AUC）的面积、预测模型的准确性、灵敏度和特异性，结果表明，包含个体驾驶员因素可以提高两个结果的 AUC 和预测准确性。眼睑测量可以提高车道偏离模型的预测准确性，但不能提高微睡眠模型的预测准确性。增加驾驶性能预测因子或增加到事件的时间，都不会改变预测性能。对于两种困倦测量方法，最好的模型都是考虑驾驶员个体差异和眼睑测量的模型，这表明在预测困倦事件时应强烈考虑这些指标。本文的研究结果可以有助于实时困倦检测的开发，并有助于管理困倦以避免相关的机动车事故和损失。
