SAGE: Subject-Adaptive Graph-Based Modeling With Decision-Level Calibration for Stress and Cognitive Workload...

Area of Science:

• Physiological computing
• Machine learning for driver monitoring

Background:

• Driver safety relies on monitoring stress and cognitive workload.
• Inter-subject variability in physiological signals complicates accurate estimation, especially with limited calibration data.

Purpose of the Study:

• Propose a subject-adaptive graph-based modeling with decision-level calibration (SAGE) framework.
• Enable unified modeling and decision-making for both driver stress and cognitive workload estimation.
• Achieve reliable personalized adaptation with minimal labeled calibration data.

Main Methods:

• SAGE learns shared physiological representations with cross-subject generalization.
• Adaptation extends from representation to decision levels.
• Incorporates sample-quality-aware calibration and threshold adaptation.

Main Results:

• Achieved 80.6% stress recognition (AffectiveROAD) and 80.5% workload recognition (hciLab) under Leave-One-Subject-Out (LOSO).
• Reached 91.8% (AffectiveROAD) and 89.6% (hciLab) accuracies in random-split settings.
• Demonstrated transferable physiological representations across tasks and scenarios on the SWELL-KW dataset.

Conclusions:

• SAGE offers an effective approach for stable driver state modeling amidst subject heterogeneity and varying signal quality.
• Combining multimodal physiological data with adaptive decision mechanisms enhances driver monitoring.
• The framework's ability to generalize across tasks suggests broad applicability in real-world scenarios.