Logo image
Back
Detection of alcohol-involved driving via machine learning using eye-tracking and vehicle data
Journal article   Open access   Peer reviewed

Detection of alcohol-involved driving via machine learning using eye-tracking and vehicle data

Ryan Miller, Andy Chestovich, Anshul Satav, Angela Eichelberger, Jonny Kuo, Michael Lenne and Timothy Brown
Journal of safety research, Vol.99, pp.147-156
10/05/2026
DOI: 10.1016/j.jsr.2026.09.014
url
https://doi.org/10.1016/j.jsr.2026.09.014View
Published (Version of record) Open Access

Abstract

•Convolution-based vehicle and eye-tracking features distinguished sober and alcohol-dosed driving (AUC = 0.955).•Using both eye-tracking and vehicle inputs outperforms eye-tracking only (AUC = 0.916) and vehicle only (AUC = 0.880).•Model accuracy is higher using longer measurement periods (90s) than short periods (30-60s).•Features capturing variation in pupil and blinking behavior are most influential in the detection of alcohol-presence.•Models trained using eye-tracking inputs maintain predictive accuracy in new driving scenarios that they were not trained on. Introduction: Alcohol impaired driving remains a significant public safety problem, prompting the United States and the EU to move towards the inclusion of impaired driving prevention technology in future vehicle production. A large body of literature suggests acute alcohol use produces changes in ocular behavior and vehicle control, suggesting driver monitoring system (DMS) data can be used to detect alcohol-impaired driving. Method: We use DMS data from a simulated driving study to assess the role of different monitoring lengths, data inputs, and machine learning methods on the accurate identification of alcohol presence using a baseline-controlled study with no alcohol at baseline, followed by four drives after alcohol dosing with BrACs of 0.10, 0.085, 0.07, and 0.055 g/210 L. Results: Using data from an 8-minute segment of rural driving, we find convolution-based feature extraction methods using vehicle and ocular inputs can differentiate between samples of sober and alcohol-dosed driving with a high degree of accuracy (AUC = 0.955), outperforming models using only ocular inputs (AUC = 0.916) and only vehicle inputs (AUC = 0.880). When applied to new driving environments, models involving ocular features continue to perform well (AUC > 0.95), while models using only vehicle-based features are less predictive. We find pupil and blink behavior are the most influential predictors of alcohol presence. Future research should consider a broader range of BrACs and fatigue combinations, particularly if a DMS aims to differentiate between alcohol-impaired and fatigued states. Practical Application: As policymakers explore the feasibility of DMS to prevent alcohol-involved accidents, we find these systems should focus on longer measurement durations and utilize both vehicle and eye data (especially blink and pupil behavior) to achieve optimal performance. Furthermore, we find that machine learning models trained to detect patterns in eye data can be successfully applied in new driving environments without retraining.
 Alcohol Driver Monitoring Systems Driving under the Influence Eye-tracking

Details

Metrics

2 Record Views
Logo image