Logo image
Back
Making highways and workplaces safer: An interpretable machine learning approach to predicting recent cannabis use and impairment
Journal article   Peer reviewed

Making highways and workplaces safer: An interpretable machine learning approach to predicting recent cannabis use and impairment

Sarah M. Bird, Ryan Peterson, Michael Kosnett, Ashley Brooks-Russell and Julia Wrobel
Journal of safety research, Vol.97, pp.142-151
06/2026
DOI: 10.1016/j.jsr.2026.02.004

View Online

Abstract

Introduction: Due to a rise in fatal crashes where cannabis use is detected, there is substantial need for objective measures of cannabis-induced impairment. Research has identified promising measures to detect recent cannabis use and impairment, but they are often considered in isolation. Our study aimed to jointly evaluate the predictive potential of various physiological and cognitive measures to determine which are most effective at detecting recent cannabis use and impairment. Methods: Adult participants (N = 125) were recruited for an observational study. Users were asked to consume inhaled cannabis products ad libitum over a 15-minute interval and non-users were used as controls. Participants were assessed pre and post smoking with: (1) a tablet-based psychomotor test battery, (2) handheld pupillometry, and (3) blood cannabinoids at baseline (pre-use), and at 40 and 100 minutes after the start of inhalation. These measures were used to predict recent cannabis inhalation and impairment (defined as change in standard deviation of lateral position >6 cm on a MiniSimTM driving simulator) via penalized logistic regression with the Minimax Concave Penalty (MCP). Separate models were fit using data collectable in different contexts (roadside and occupational settings). Results: Models including covariates from each data source performed best at both timepoints. At 40 minutes post-use, recent inhalation and impairment models had areas under the curve (AUCs) of 0.996 and 0.886, sensitivities of 100% and 83%, and specificities of 99% and 82%, respectively. At 100 minutes post-use, models for recent use and impairment performed worse with AUCs of 0.988 and 0.866, sensitivities of 97.8% and 79.2%, and specificities of 96.7% and 82.3%. Conclusion: Including multiple sources of data can improve prediction of recent cannabis use and impairment, over single sources. Additionally, predictive performance was stronger for measures collected at 40 minutes relative to those collected at 100 minutes post-use.
 Drug impairment Penalized regression Predictive model Cannabis impaired driving Interpretable machine learning

Details

Metrics

1 Record Views
Logo image