It matters who is behind the wheel: driver monitoring feature analysis using explainable AI

This work-in-progress examines how gaze-based features and individual driver characteristics influence takeover performance prediction in partially automated vehicles. We present preliminary findings from a driving simulator study (N=33) that used a decision-tree (XGBoost) machine learning model and explainable AI techniques (permutation feature importance and SHAP analysis). Results show that driver profile features—particularly professional training, experience, and age—emerged as highly predictive of takeover readiness alongside traditional gaze metrics like fatigue indicators. While current Driver Monitoring Systems (DMS) approaches and regulatory recommendations focus on universal gaze thresholds, our preliminary analysis reveals that individual driver characteristics may be more important for predicting takeover performance. These findings suggest potential for developing adaptive automotive interfaces that adjust based on driver profiles rather than one-size-fits-all approaches. The preliminary results highlight the need for careful consideration when designing driver monitoring systems and automotive interfaces for partially automated vehicles.