Analysis of Skin Conductance in a Stress-Inducing Virtual Reality Experiment

Emotion recognition is an ever-expanding field with applications ranging from medical to psychological to commercial. Currently, innovative methods are being used in emotion detection research, with a focus on stress, a condition that is an integral part of everyday life and can have short and long term consequences. Electrodermal activity (EDA), also known as galvanic skin response (GSR) or skin conductance (SC), is a quantitative method of measuring variations in sweating by detecting changes in electrical conductivity. Two 3D virtual reality environments were developed with the specific aim of eliciting emotions. Each environment comprises seven tasks which incorporating a stress-inducing stimulus, termed "stressor", derived from prior literature research. The primary objective of the first environment is to induce stress, whereas the second environment has been designed to provide a less stressful experience by alleviating stressors. The experiment was conducted at the 3D Lab of the Politecnico di Torino, with 61 subjects undergoing emotional monitoring. Subjects were required to wear EMOTIV EPOC Flex headsets to collect electroencephalography (EEG) data and Shimmer3 GSR+ devices to collect electrodermal activity (EDA) and heart rate (HR) data. At the conclusion of the two experimental sessions, subjects were asked to complete two questionnaires, one for each experience. These questionnaires were based on the Self Assessment Manikin (SAM), which required subjects to rate the emotions they experienced in terms of valence, arousal, dominance, and stress on a scale of 1 to 5. Once the questionnaires had been completed, the signals recorded from the participants were analysed. In particular, the present work focuses on EDA analysis with the aim of assessing skin conductance as an indicator of stress and arousal levels. The EDA signals were processed using a MATLAB toolbox (Ledalab) and five features were extracted from them: the average of the tonic and phasic components, the number of peaks, the maximum peak’s amplitude and the average peak’s amplitude. These features were statistically analysed in two ways: in the first case, the Shapiro-Wilk test was used to check normality and the Friedman test to assess statistically significant differences between the same tasks in the two environments; in the second case, a Paired Samples T-test was used to assess statistically significant differences between the two environments. The non-parametric Friedman's test confirmed a complete significant difference in the tonic component and the maximum peak’s amplitude, indicating that these two characteristics are particularly robust. Regarding the t-test, a complete significant difference was obtained, confirming that the GSR is a suitable physiological signal for the study of arousal and stress. Finally, KNN and SVM machine learning algorithms were applied for binary and multiclass classification by correlating the answers from the questionnaires with the features extracted from the EDA. The aim was to assess the differences between stress and non-stress states and between high and low arousal, which was confirmed by the high performance obtained. In addition, a binary SVM classifier based solely on physiological indicators was developed to assess the differences between the two environments, showing excellent results.