Automated Stress Recognition Using Supervised Learning Classifiers by Interactive Virtual Reality Scenes

Virtual reality (VR) technology offers a great opportunity to explore stress disorder therapies. We created a VR stress training system, which incorporates three highly interactive stressful scenes to elicit stress, and demonstrate the concurrent variations between physiological data (heart rate, electrodermal activity and eye-blink rate) and self-reported stress ratings through a self-designed customized perceived stress questionnaire(SSAI) and wearable devices. Several supervised learning models were rigorously applied to automate stress recognition. Our findings include the evaluations of the VR system by computing Cronbach's alpha (alpha=0.72) and Kaiser-Meyer-Olkin (KMO) coefficient (eta=0.78) through a retrospective survey, which were subsequently confirmed as reliable on four aspects (sense of presence, sense of space, sense of immersion and sense of reality) via factor analysis. Additionally, we demonstrate the effectiveness of physiology-based stress level classification (no stress, low stress and high stress) and continuous SSAI score prediction, with accuracy reaching 0.742 by bagging ensemble learning model and goodness-of- fit reaching 0.44 via multivariate stepwise regression. This study provides detailed insight regarding the effect of objective physiological measures on the validation of subjective self-ratings under a novel complex VR stress training system, which stimulates the further investigations of stress disorder recognition and treatment.