Automatic Psychophysiological Stress Detection Through Infrared Thermography and Explainable Artificial Intelligence

Stress is the body's response to an unusual situation. While it can be beneficial to a certain extent, in large amounts, it can be adverse. This paper proposes a methodology for detecting psychophysiological stress using infrared thermography, deep learning and explainable artificial intelligence. The Trier Social Stress Test protocol was used to induce psychophysiological stress in 58 university students, and thermographic images were collected during this time. Subsequently, the facial and hand areas were automatically segmented. This information was used to train a convolutional neural network to detect psychophysiological stress. Similarly, the Grad-CAM method was implemented to identify which areas of the face or hands contributed to the classification. Finally, the computational complexity of the proposal was evaluated. The results demonstrate that stress detection can be achieved with 99.5% accuracy, 98.3% precision, 98% recall and 98% F1-Score. Regarding the Grad-CAM results, it was observed that both areas, the hand and facial areas, tend to contribute more to classification in women than in men. The computational complexity analysis of the model indicates that it used 480.371MFLOPS, 11.569 MPARAMS and an average training time of 150.82 seconds. Based on this, it is concluded that infrared thermography is a technology that can aid in the detection of stress, and when combined with explainable artificial intelligence techniques, it enables users to understand the basis for their decisions, specifically which areas of the face and hands are affected.