Abstract
Mental stress poses a growing threat to public health, yet video based stress detection remains challenging because of substantial inter individual variability in physiological and expressive responses. To address this issue, we propose a novel two level learning framework grounded in allostasis theory, where stress is modeled as a personalized deviation from an individual's physiological baseline rather than as an absolute state. We introduce CalmScore, a metric based on resting heart rate variability, to robustly identify each subject's most relaxed reference state. Building on this reference, an intra subject Physiological Discrepancy based Representation Adaptive Modulation module computes multimodal deviations between the current state and the resting state, and further modulates them using resting heart rate variability as a proxy for regulation capacity. In addition, an inter subject analogical reasoning mechanism based on In Context Instruction Tuning retrieves physiologically similar peers and provides stressed, unstressed, and resting examples for contextual calibration. Extensive experiments on the UVSD, RSL, and MUSE datasets demonstrate the effectiveness of the proposed framework. Our method achieves state of the art F1 scores of 96.85% on UVSD and 88.67% on RSL, surpassing the strongest baseline. Ablation studies further verify the necessity of each component. The results show that robust video based stress detection benefits from modeling individualized deviations and interpreting them through group level physiological analogy.