Detection of Alcohol-Induced Driving Impairment Using Heart Signal Monitoring and Machine Learning

Abstract

Alcohol-related driving impairment remains a significant safety challenge, contributing to approximately 25% of traffic-related fatalities in Europe. This master’s thesis investigates whether heart signal monitoring, in the form of heart rate (HR) and heart rate variability (HRV), together with machine learning, can detect alcoholinduced driver impairment. A within-subject experimental study was conducted with 37 participants in a controlled driving simulator under three conditions: sober, low-to-moderate blood alcohol concentration (BAC), and high-BAC (≥0.7‰). Additionally, an external real-world driving dataset from 17 participants (sober and high-BAC) was used to evaluate model generalizability. Although a contactless 24 GHz radar-based heart signal monitoring system was initially evaluated, the radarderived signals lacked the reliability needed for robust HRV analysis in this thesis. Therefore, the main classification pipeline utilized reference data from a Polar H10 chest-worn sensor. Several machine learning classifiers, including Support Vector Machines (SVM) and Extreme Gradient Boosting (XGBoost), were evaluated using leave-one-subject-out cross-validation. Tested on the real-world dataset, the models achieved a session-level area under the receiver operating characteristic curve (AUC) of 0.75. The best-performing model evaluated on simulator data reached a sessionlevel AUC of 0.81 in the high-BAC setting. An exploratory late-fusion approach that combines simulator-based heart signals and camera-based driver monitoring system (DMS) outputs further improved performance, achieving an AUC of 0.93 in the high-BAC setting. These findings highlight the potential of heart signal monitoring to complement camera-based DMS, thereby supporting multimodal impairmentdetection approaches aligned with Euro NCAP 2026 safety requirements.