A comprehensive exploration of motion sickness process analysis from EEG signal and virtual reality

Virtual reality motion sickness is a significant barrier to the widespread adoption of virtual reality technology. Current virtual reality motion sickness detection methods using EEG signals often fail to identify comprehensive neuro-markers and lack generalizability across multiple subjects.

To address this issue, we analyzed the pre- and post-induction phases of virtual reality motion sickness, as well as the induction process, from multiple domain features. The features were extracted from time domain, frequency domain, spatial domain and Riemann space across delta, theta, beta, and all frequency bands. The neuro-markers selected have a correlation greater than 0.5 with behaviors information and showed significant changes in both phases. Five kinds of traditional machine learning methods were used to classify VR motion sickness states in within-in subjects and cross-subjects by using neuro-markers.

Traditional machine learning methods achieved a maximum accuracy of 92 % for within-subject classification and 68 % for cross-subject classification. Spectral entropy across all frequency bands yielded the highest classification accuracy during the pre- and post-induction phases, while spectral skew showed the most significant changes during the task phase.

These findings suggest that these features hold strong potential for future neurofeedback studies.