Detecting Impaired Driving: Vehicle Inputs with the Greatest Influence on Accurate Identification

Abstract

Alcohol and drug impairment are major causes of crashes and fatalities. Despite evidence of significant differences in lateral control, speed, and steering across impaired and sober states, leveraging this information for the automated detection of impairment using vehicle inputs remains a challenge. Using data from two driving simulator studies involving sources of impairment (cannabis, alcohol, and combined cannabis and alcohol) and a range of driving environments, we develop machine learning models to identify acute drug use using vehicle data with a focus on understanding which features exert the greatest influence on predictions across a variety of scenarios. Under a robust validation framework, we used SHapley Additive exPlanations to determine that features derived from lateral position were consistently among the most predictive, whereas vehicle speed, steering wheel rate, and brake force also made meaningful contributions. Features derived from lateral position were more predictive when calculated over longer periods (60 s) relative to shorter time spans (0.33 to 5 s). Leveraging subject-specific information tended to improve classification performance. For detecting recent cannabis use (~30-min postdose), receiver operating characteristic—area under the curve (AUC) improved from 0.0625 to 0.706 in a straight, two-lane rural road scenario, and for classifying combined cannabis and alcohol impairment ROC-AUC improved from 0.643 to 0.689 on a straight segment of a four-lane divided expressway. These findings highlight the potential for carefully selected features derived from vehicle inputs to contribute to automated impairment detection, and the need to consider driving environments and sources of impairment when using these features.