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

In real-world driving scenarios, reliable monitoring of drivers' stress and cognitive workload is critical for driving safety. Both tasks can be characterized using multimodal physiological signals; however, due to pronounced inter-subject variability, decision instability becomes particularly prominent under limited calibration data. This paper proposes a subject-adaptive graph-based modeling with decision-level calibration (SAGE) framework, which supports both driver stress and cognitive workload estimation under a unified modeling and decision paradigm. SAGE learns shared physiological representations with cross-subject generalization capability and extends subject adaptation from the representation level to the decision level. By incorporating sample-quality-aware calibration and threshold adaptation, SAGE enables reliable personalized adaptation using only a small amount of labeled calibration data. Under a strict leave-one-subjectout (LOSO) evaluation protocol, we validate the proposed method on two datasets with distinct feature distribution characteristics. Experimental results show that SAGE achieves a classification accuracy of 80.6% for stress recognition on the AffectiveROAD and 80.5% for cognitive workload recognition on the hciLab, significantly outperforming representative existing methods. Under randomsplit settings, SAGE further achieves classification accuracies of 91.8% on AffectiveROAD and 89.6% on hciLab Driving. These results suggest that, under conditions of subject heterogeneity and unconstrained signal quality, combining multimodal physiological representations with decision-level adaptive mechanisms provides an effective approach for stable driver state modeling. Furthermore, validation on the SWELL Knowledge Work (SWELL-KW) dataset suggests that the proposed framework captures transferable physiological representations that generalize across different tasks and real-world scenarios.