Published Paper Details:

Psychological stress is a major, often invisible, contributor to cardiovascular disease, anxiety disorders, burnout and reduced workplace productivity. Conventional stress assessment relies either on subjective self-report questionnaires or on contact-based physiological sensors (electro dermal activity, ECG, respiration belts) that are intrusive, uncomfortable over long durations, and impractical for continuous, unobtrusive monitoring. Thermal infrared imaging offers a non-contact alternative because psychophysiological arousal produces measurable, involuntary changes in facial skin temperature distribution, most notably nasal-tip cooling due to peripheral vasoconstriction and forehead/periorbital warming due to sympathetic activation. However, existing thermal-based stress detection systems are largely opaque "black-box" deep networks, which limits clinical trust, auditability, and adoption in healthcare and human-factors settings. This paper proposes an Explainable Multimodal Deep Learning Framework that fuses spatial thermal facial features with thermally-derived respiratory dynamics through an attention-based fusion layer, and augments the resulting classifier with Gradient-weighted Class Activation Mapping (Grad-CAM) for spatial saliency and Shapley Additive explanations (SHAP) for global and local feature attribution. Because no proprietary clinical thermal-stress corpus was accessible for this study, we validate the framework as a reproducible proof-of-concept using a literature-parameterized synthetic multimodal dataset whose class-conditional statistics reproduce documented directional thermal and respiratory stress effects. A compact feed-forward fusion network (MLP), trained and evaluated with stratified 5-fold cross-validation, is compared against Random Forest and Support Vector Machine baselines. The proposed fusion model achieves 89.5% accuracy, 89.6% F1-score and an AUC of 0.953, comparable to the 74-98% accuracy range reported by related thermal- and physiological-signal stress-detection studies in the literature. Permutation-based feature importance and Grad-CAM-style saliency consistently highlight facial-temperature variability, nasal-tip temperature and periorbital temperature as the dominant, physiologically plausible decision drivers, corroborating the model's internal reasoning against established psychophysiology. The results support the feasibility and design rationale of an explainable, non-contact, multimodal thermal stress-detection pipeline and outline the steps required to move from this proof-of-concept to validation on real clinical thermal video datasets.