DriveAware: Multimodal Wearable Sensing–Based Machine Learning Approaches to Detect Driver Distraction

Driver distraction is a major public health and road safety concern, linked to delayed reactions, impaired attention, and increased crash risk. This study introduces a simulation-based framework to detect distraction by systematically collecting multimodal physiological signals, electrodermal activity (EDA) and photoplethysmography (PPG), across three driver states (baseline, undistracted, and distracted) under diverse road conditions. Sessions and secondary-task orders were randomised, and performance was assessed using nine-fold cross-validation by leaving five subjects out each fold for evaluation. Features were extracted from both modalities and refined using Shapley Additive exPlanations (SHAP) to identify the most significant features, enhancing model transparency and offering new insights into physiological markers of distraction. This design frames the method as an enhanced AI decision-support pipeline suitable for driver distraction alerting. We conducted a comprehensive evaluation of multiple machine learning models, comparing unimodal (PPG-only, EDA-only) and multimodal fusion approaches. In particular, middle fusion with Random Forest and XGBoost, and late fusion through soft voting and stacking, were systematically explored. The results show that late fusion through soft voting and stacking, combined with XGBoost and SHAP-based feature selection, significantly outperforms single-modality approaches, achieving accuracy of (73.37 ± 3.52)% for three level of baseline, undistracted, and distracted driving. Multimodal gains were consistent across folds, underscoring the reliability benefits of decision-level fusion over single-modality models. Whereas, we obtained the best performance in XGBoost by (71.19 ± 2.85)% for single modality of EDA only and by (69.63 ± 4.07)% for single modality of PPG only. These findings highlight the potential of interpretable multimodal physiological monitoring, combined with wearable sensing, as a promising direction for real-time driver-state monitoring, subject to further validation.