Beyond Supervised Limits: Semi-Supervised Cybersickness Prediction from Physiological Signals with Minimal Labeled Data

Cybersickness, characterized by discomforts such as dizziness, nausea, and eye strain, remains a significant barrier to the widespread adoption of virtual reality (VR). Recent research have proposed supervised machine learning models to predict the onset of cybersickness; however, these approaches depend heavily on labeled datasets. Acquiring labeled datasets typically necessitates time-consuming and resource-intensive user studies, limiting the feasibility of these supervised methods for consumer-level VR applications where obtaining labeled user data during use is impractical. Moreover, due to individual differences, often these datasets are not generalizable in consumer VR use. To address these limitations, we propose a novel semi-supervised learning framework for predicting cybersickness (i.e., Fast Motion Sickness (FMS)) using eye tracking, heart rate, and galvanic skin response data. Our proposed semi-supervised approach uses pseudo-labeling techniques (i.e., Self-training, Label Propagation, and Label Spreading) fused with temporal deep learning models (i.e., DeepTCN, CNN-LSTM, Transformers, LSTM). We evaluated our approach on three public cybersickness datasets (i.e., Bumpy Ride, Simulation 21, Maze) and our proposed semi-supervised approach demonstrates strong cybersickness predictive performance using only $1-5 {\%}$ labeled data (i.e., $95-99 {\%}$ data remains unlabeled). Notably, the self-training approach with a DeepTCN model achieved an accuracy of $\mathbf{7 5. 8 6 \%}$ in FMS prediction, outperforming the other models and pseudolabeling approaches. Our findings establish the viability of semisupervised learning for cybersickness prediction with minimally labeled datasets, paving the way for more practical and potentially generalizable cybersickness prediction systems in consumer VR applications.