Exploratory Multimodal Machine Learning for Hazard Response Detection in Semi-Automated Rural Driving

Partially automated driving systems require drivers to remain cognitively engaged and prepared to respond to unexpected roadway hazards despite reduced direct vehicle control. In rural driving, low visibility and sudden animal crossings pose unique safety challenges to driver readiness and response capability. This study presents an exploratory multimodal machine learning framework for hazard-response classification under semi-automated rural driving. Using a high-fidelity driving simulator, EEG, eye activity, behavioral, and subjective workload data were acquired under manual and Advanced Driver Assistance System (ADAS)-supported conditions. Pre-hazard event-level features were evaluated using logistic regression, random forest, and multilayer perceptron classifiers with leave-one-subject-out cross-validation. Random Forest achieved the highest mean classification accuracy, although performance was constrained by class imbalance and limited sensitivity to brake events. Feature-importance analysis identified SART situation-awareness score as the most influential feature, followed by NASA-TLX workload, pupil-diameter change, and EEG-derived measures. These findings support multimodal hazard-response classification for future driver-readiness modeling.