Single-Lead EEG-based virtual reality motion sickness detection using interpretable feature selection

Electroencephalography (EEG) is a key method for detecting virtual reality motion sickness (VRMS), but multi-lead recordings and high-dimensional features increase processing complexity, device cost, and user burden.

This study proposes a SHapley Additive exPlanations (SHAP)-based single-lead modeling and feature-screening framework to reduce the number of EEG leads and feature dimensions for lightweight VRMS detection.

Six classical deep-learning EEG decoding models were used to assess the contribution of each lead via SHAP values and identify the optimal single lead. Handcrafted features were then extracted and selected using a correlation-SHAP strategy, and XGBoost was adopted for final classification.

FT11 showed competitive and relatively consistent performance across the evaluated models and was selected as the target lead after integrating SHAP-based lead importance, single-lead classification performance, frequency-band differences, and physiological relevance. Under the window-level tenfold cross-validation protocol used in this study, the proposed model achieved an accuracy of 87.53% ± 0.73% using seven selected features, which was 9.55 percentage points lower than the full-lead full-feature model and 0.69 percentage points lower than the single-lead full-feature model.

These findings demonstrate an interpretable framework for identifying a compact single-lead and feature configuration for VRMS-related EEG discrimination. An additional post hoc LOSO test of the globally selected configuration showed limited performance on held-out participants. Because the lead, classifier, and feature configuration had been selected globally before LOSO testing, these results should not be interpreted as a fully independent estimate of cross-subject generalization.