Intelligent detection of alcohol-impaired driving via variational autoencoders and SHAP-based interpretation

Drunk driving poses a serious risk to road safety. This study introduces a semi-supervised approach to detect intoxicated drivers by combining Variational Autoencoders (VAEs) for feature extraction with a One-Class Support Vector Machine (OCSVM) for anomaly detection. The method is trained only on data from sober drivers, making it suitable for limited-label scenarios. Model predictions are explained using SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations), which consistently identify alcohol concentration as the most influential feature. The proposed framework is evaluated using a publicly available dataset from IEEE Dataport, which includes sensor readings from alcohol gas sensors, facial temperature data, and pupil measurements captured by a Raspberry Pi camera. The VAE-OCSVM model achieves strong results, with an F1-score of 98%, outperforming standard clustering-based methods as well as classical semi-supervised statistical monitoring approaches, including Principal Component Analysis (PCA)-based Hotelling \(T^2\) , PCA-based Squared Prediction Error (SPE), and Independent Component Analysis (ICA)-based SPE. Bootstrap-based confidence intervals are computed to assess the robustness of performance metrics.