Abstract
We present an information-theoretic framework for identifying which physiological and eye-gaze-based indicators, and at what temporal scales, are most informative of cognition-related reliance behavior during conditionally automated (SAE Level 3) driving. We analyze data collected from an in-person driving simulator study in which participants interact with a conditionally automated vehicle in a single continuous drive. By evaluating physiological signals (heart rate, galvanic skin response) and eye-gaze fixations through mutual information and conditional mutual information, the approach enables assessment of both overall and uniquely contributed informational relevance without assuming linear relationships or predefined model structures, making it suitable for continuous, non-trial-based settings. Results show substantial inter-participant variability in both informative features and optimal time scales, suggesting that no single physiological indicator is universally optimal. The proposed framework highlights the importance of personalized sensing strategies and provides a scalable methodology for future larger-scale studies on cognitive-state inference in automated driving.
