Multi-level Stress Assessment From Electrocardiogram Data in a Virtual Reality Environment Using Contrastive Self Supervision

Electrocardiogram (ECG) data is an ideal option to assess stress in Virtual Reality (VR) applications due to its non-invasive nature and high correlation to changes in stress levels. Yet, most studies on stress assessment only collect binary measurements. Developing a multi-level assessment is necessary for a biofeedback-based application. Existing studies annotate and classify a single experience (e.g. watching a VR video) to a single stress level, which again prevents design of dynamic experiences where real-time in-game stress assessment can be utilized. Furthermore, existing works heavily rely on supervision information to train stress assessment models. However, obtaining supervised labels for training data is an expensive and labor intensive task that is not scalable on large datasets. Given the advancements in data driven technologies, supervised learning prevents us from leveraging unlabeled data excessively available nowadays. In this thesis, we report our findings on a new study on VR multi-level stress assessment. ECG data was collected from 12 participants experiencing a VR roller coaster. The VR experience was then manually labeled in 10-second segments to three stress levels. We then propose a contrastive self-supervised learning approach for stress assessment that is superior to non-contrastive self-supervised approaches. Experimental results show that when compared to non-contrastive self-supervised learning, we obtained a 9% increase in accuracy on the WESAD dataset and a 3.7% increase on our collected data at the Ryerson Multimedia Lab.