Drowsiness Metrics and Thresholds: The Search for Valid and Reliable Standards

Abstract

Data for this project was leveraged from an existing 61-driver, on-road instrumented vehicle study (unpublished) designed to induce drowsy driving episodes. The research did not target sleep-deprived individuals but rather normally rested individuals under the assumption that low-workload environments could be used to induce drowsiness. This allowed the progression from alert to drowsy states to be more representative of typical drivers and not sleep-deprived individuals. Outputs were gathered from a range of onboard sensors and devices (including a driver monitoring system) within each of four basic classes of measures (vehicle-based driving performance, behavioral, physiological, and subjective), including percentage of eye closure (PERCLOS), lane tracking, driver Karolinska Sleepiness Scale (KSS) ratings, Observer Rating of Drowsiness (ORD), and extended eye closures, among others. The goal of this project was to develop evidence-based measures, criteria, and thresholds to identify the onset and progression of drowsiness. Work sought to establish and validate drowsiness measures (PERCLOS, KSS, ORD, standard deviation of lane position [SDLP], lane deviations, etc.), either individually or in combination, that can reliably detect drowsiness. Driver self-assessments taken during the trip in the form of periodic KSS ratings show that drivers tended to overestimate drowsiness levels in comparison to researcher ratings (taken concurrently with driver ratings) and ORD scores independently analyzed post hoc. Relationships between extended eye closure events (greater than 2 seconds) and KSS ratings from both the driver and researcher suggest these events are predominantly limited to higher KSS levels associated with sleepiness. Objective indicators of driving performance (SDLP and lane deviation rate) were found to reliably discriminate among the three defined driver alertness states: alert, moderately drowsy, and drowsy. Sensitivity analyses using different ground truth indices consistently found that driver KSS yields moderate “hit” rates (ranging from 64% to 74%) but very high “false alarm” rates (ranging from 22% to 32%) across all measures and metrics. Higher overall hit rates were observed for researcher KSS, ORD, PERCLOS, and eye closure measures across all four indices. More importantly, combining individual measures to define driver states (e.g., alert or drowsy) was found to increase overall predictive accuracy.