Resilience-Gated Multimodal Attention for Physiological Stress Detection

Automatic stress detection from physiological signals is challenging due to substantial inter-individual variability in stress reactivity. This paper proposes a multimodal deep learning framework that incorporates psychological resilience, measured using the Connor–Davidson Resilience Scale (CD-RISC), into the feature fusion process. Five physiological signals (EEG, BVP, EDA, HR, and HRV) are processed through modality-specific encoders and combined using a resilience-conditioned gating mechanism (RGMA). The model is evaluated on a PhysioNet dataset using leave-one-subject-out cross-validation. The proposed method achieves a mean AUC of 0.791 and outperforms baseline models including EEGNet, naive fusion, and SVM. The results demonstrate that incorporating resilience information improves cross-subject stress detection, particularly for individuals with suppressed physiological responses. This work is currently submitted to an IEEE conference and is available as a preprint.