Analysis of stress condition through EEG in Virtual Reality

Recognizing and monitoring emotions has become more important in different research fields, such as Human-Computer Interaction (HCI), pattern recognition and computer vision, with the aim of enhancing user experience by making interactions with digital systems more effective, personalized, and empathetic. It has numerous practical applications across var- ious sectors, including medicine and psychology. In fact emotion recognition is extremely helpful for therapeutic purposes and for improving the social interpretation of emotions. For emotion recognition and monitoring many techniques can be used, for instance the EEG interpretation, the analysis of facial expressions and the study of Heart Rate Variabil- ity (HRV), depending on which emotion is considered. In particular, the present thesis is fo- cused on one specific emotion: stress. Analysis of the previous literature revealed that stress can have various negative effects on human body, both in short and long term, including skin disorders (i.e. acne, eczema, psoriasis), muscle pain and tension, gastrointestinal disorders, increased risk of heart diseases due to the increased Heart Rate (HR) and blood pressure, weakening of immune system due to cortisol release, cognitive issues and increased risk of anxiety and depression disorders. Hence it is crucial to monitor stress responses. VR was used as eliciting stimuli. For this purpose two 3D VR environments were developed with Unity software. During the experiment, participants were asked to deal with 8 different tasks, each of which included a stress triggering stimulus, called stressor. Each environment had 8 stressors, that have been chosen on the basis of the previous literature research. The aim of the first experience was to raise stress in participants while the second environment was designed with softer stressors in order to provide a less stressful experience. The ex- periment was carried out in the Polytechnic of Turin’s 3D Lab and the 39 participants wore an EEG headset (EMOTIV Epoc X) and a sensor on the wrist for skin conductance (Shim- mer3 GSR). Facial expressions were also acquired through a depth camera (Intel RealSense), which was positioned in front of the subject. At the end of the experiences, participants com- pleted two questionnaires, where they had to evaluate the felt emotions in terms of valence, arousal and dominance with a score from 1 to 5. Although different biological parameters were observed, the present work is focused on EEG analysis. From EEG waves, 4 indicators, namely stress indicator, arousal, valence and dominance, were calculated. These indicators were subsequently labelled by using ques- tionnaire answers in order to perform the EEG indicators classification through the use of 4 different Machine Learning (ML) algorithms. The same steps were repeated changing the percentage size of test set with the aim of finding the most efficient one. The different results were compared by evaluating and confronting some performance metrics, such as precision and recall. The purpose of the classification was to research a correspondence be- tween questionnaire answers and EEG indicators, which was confirmed by the high obtained performances.