StressVision: Non-Invasive Stress Detection from Thermal Videos

Abstract

Timely and accurate stress detection is crucial for effective healthcare monitoring and intervention. Existing methods for stress detection often rely on invasive or subjective measures, limiting their use. Here, we propose StressVision; a non-invasive and automated transformer-based deep learning approach that uses thermal video analysis to capture and analyze facial thermal patterns, and enables objective and continuous stress detection. We validate our approach by applying StressVision to two datasets comprised of healthy human adult participants who were exposed to an acute stressor (ice-cold water) while thermal video of their faces and electrocardiography were recorded. One of these datasets was collected specifically for the purpose of this work (n=36) and the other dataset was acquired from a previous study (n=42). With StressVision we were able to achieve state-of-the-art stress detection performance, such that stress state could be classified (i.e., stress, no-stress) with accuracy = 0.8748. We make the StressVision source code available on GitHub along with our new dataset, which will serve as a valuable resource for stress-detection research and allow for bench-marking against other methods.