Benchmarking the MQ-3 Gas Sensor for Alcohol Impairment Detection: Insights from a Driving Simulator Study

Accurate detection of driving under the influence (DUI) of alcohol is essential for improving road safety and fostering public trust in vehicle safety technologies. While MQ-3 gas sensors are widely used for ethanol detection, their effectiveness in realistic driving scenarios with human participants remains underexplored. This study aims to fill this gap by systematically evaluating the performance of various machine learning classifiers in detecting alcohol impairment based on sensor data. A high-fidelity, dynamic driving simulator study was conducted with N = 120 participants, each completing two driving sessions: one sober and one intoxicated. Of these, n = 80 began sober, while n = started alcohol, 40 under the influence of resulting in residual blood alcohol concentrations during their second session. Voltage output from the MQ-3 sensor was continuously recorded, visualized, and analyzed to extract key features for classification. Multiple algorithms were trained and tested, with the XGBoost classifier achieving a maximum accuracy of 94% in differentiating between sober and intoxicated states. However, the model showed reduced performance (69% accuracy) in distinguishing between fully intoxicated individuals and those with residual alcohol levels, highlighting a critical limitation of single-modality sensing. These results underscore the necessity of future research into multimodal detection frameworks that integrate ethanol sensing with additional in-cabin contextual and behavioral data to enhance robustness and real-world applicability.