Published Paper Details:

: Psychological stress is a pervasive determinant of health, productivity, and quality of life, and its unobtrusive, continuous assessment remains an open challenge for affective computing and digital health [1],[3],[7]. Contact-based physiological sensing (electro dermal activity, electrocardiography, and photo plethysmography) offers reliable ground truth but is intrusive, motion-sensitive, and impractical for continuous, real-world deployment [8], [9]. Thermal facial imaging has emerged as a promising non-contact alternative because sympathetic nervous system activation produces measurable, involuntary vascular and thermoregulatory changes in facial regions such as the nose, forehead, and perioral area, which can be captured passively by infrared cameras without requiring skin contact or user cooperation [18], [20], [27],[29]. In parallel, deep learning has progressively replaced handcrafted thermal feature extraction with end-to-end convolutional, recurrent, and transformer-based architectures capable of learning discriminative spatio-temporal stress signatures directly from thermal video [21], [23], [25]. However, the clinical and practical adoption of such systems is constrained by the black-box nature of deep networks, which undermines trust, auditability, and regulatory acceptance in health-adjacent applications [57],[59], [63].