Toward Postural State Classification in Immersive VR with Multimodal Data and Explainability Analysis

High accuracy is achievable but requires rigorous testing methods: The study achieved 96.76% accuracy in classifying balanced versus imbalanced postures in VR, but this success relied on Leave-One-Participant-Out cross-validation—a stringent testing approach that ensures the system can generalize to new users rather than simply memorizing patterns from the training data.

Explainability reveals which data matters most: Using SHAP analysis, researchers identified that body-motion patterns (kinematic features) were the most important factors for detecting imbalance, rather than other physiological signals like muscle activity or skin conductance. This transparency helps build trust in the system and guides future development.

Simpler models can perform nearly as well: The system maintained 95.7% accuracy while using only two-thirds of the original features, suggesting that VR safety systems don't need to collect excessive data. This has practical benefits for reducing computational demands and privacy concerns by limiting data collection.

Multimodal data collection provides robustness: The study combined three different types of signals (body motion, muscle activity, and skin response), which likely contributed to the high performance and may help the system work reliably across different users and VR environments.

Safety-critical VR applications need participant-specific validation: The emphasis on testing across unseen participants highlights that VR balance-detection systems must be validated with diverse users before deployment, as individual differences in physiology and movement patterns could affect real-world performance.