Abstract
BACKGROUND: Anxiety prediction can be used for enhancing Virtual Reality applications. We aimed to assess the evidence on whether anxiety can be accurately classified in Virtual Reality. METHODS: We conducted a scoping review using Scopus, Web of Science, IEEE Xplore, and ACM Digital Library as data sources. Our search included studies from 2010 to 2022. Our inclusion criteria were peer-reviewed studies which take place in a Virtual Reality environment and assess the user's anxiety using machine learning classification models and biosensors. RESULTS: 1749 records were identified and out of these, 11 (n = 237) studies were selected. Studies had varying numbers of outputs, from two outputs to eleven. Accuracy of anxiety classification for two-output models ranged from 75% to 96.4%; accuracy for three-output models ranged from 67.5% to 96.3%; accuracy for four-output models ranged from 38.8% to 86.3%. The most commonly used measures were electrodermal activity and heart rate. CONCLUSION: Results show that it is possible to create high-accuracy models to determine anxiety in real time. However, it should be noted that there is a lack of standardisation when it comes to defining ground truth for anxiety, making these results difficult to interpret. Additionally, many of these studies included small samples consisting of mostly students, which may bias the results. Future studies should be very careful in defining anxiety and aim for a more inclusive and larger sample. It is also important to research the application of the classification by conducting longitudinal studies.