Logo image
Back
How machine learning has been used to detect alcohol-induced driver impairment using in-vehicle sensors: A systematic review
Journal article   Open access   Peer reviewed

How machine learning has been used to detect alcohol-induced driver impairment using in-vehicle sensors: A systematic review

Brent G. Devcich, Li-minn Ang, Mingzhong Wang and Grégoire S. Larue
Journal of Safety Research, Vol.97, pp.52-66
2026
DOI: https://doi.org/10.1016/j.jsr.2026.01.021
pdf
1-s2.0-S0022437526000216-main1.80 MBDownloadView
Published VersionCC BY-NC-ND V4.0,  Open Access

Abstract

Advanced driver assistance systems Blood alcohol concentration Classification Intoxication Neural network Road safety
Introduction: Alcohol-impaired driving is a persistent global public health concern, with limited recent progress in reducing its impact on roads, highlighting the need for innovative approaches. The increasing adoption of machine learning (ML) has led to its application in detecting alcohol-induced driving impairment. This systematic review examines and synthesizes existing research that leverages ML utilizing in-vehicle sensors to detect alcohol-impaired driving, with a focus on the ML models employed and the input variables analyzed. Method: Studies were included if they applied ML techniques to detect alcohol-induced driving impairment using in-vehicle sensor data. The literature search was conducted in various academic databases and was supplemented by citation-based article retrieval. Primary outcomes of interest were the ML models employed and input variables. A risk of bias assessment was performed to evaluate study reliability and validity. Results: There were 26 relevant studies identified. The reported classification accuracy was consistently high, with median accuracy of 89%. Although the majority of studies were performed using driving simulators, there was significant heterogeneity with respect to other important study characteristics. The most common input variables used related to vehicle dynamics and control inputs, and the most common ML models implemented were neural networks. Conclusions and Practical Applications: This systematic review highlights limitations in the current literature related to significant heterogeneity in study characteristics and methodological issues in many identified studies. While some promising results were observed, further research is required to determine the optimal approach, particularly with respect to finding the most compatible and practical ML models and input variables for reliable detection.

Details

Metrics

1 Record Views
Logo image