AI and Neurological Rare & Neuroimmune Diseases ── Multiple sclerosis ── 2026-10-25
Multiple sclerosis: treatment, trials and AI
- Abstract
Understanding balance in virtual reality (VR) is important not only for measuring whether a user is stable or unstable but also for explaining how stability is maintained during immersive interaction. This is especially important for people with Multiple Sclerosis (MS), because similar levels of overall balance may be accompanied by different motion patterns across individuals. Existing VR-based approaches typically focus on raw sway signals, concurrent validity comparisons, or task-specific classification, and rarely capture balance as a single continuous measure or explain the motion patterns underlying stability maintenance. Using a publicly available multimodal dataset of head-mounted display (HMD), controller, center-of-pressure (CoP), and motion-capture recordings from 37 participants (15 MS, 22 non-MS), we present a three-stage framework for predicting and interpreting balance ability from HMD and controller motion sensors. First, we develop a continuous CoP-derived balance score from standard CoP features, representing balance as a single communicable measure, and examine its consistency across MS and non-MS participants. Second, we predict this score from HMD and controller sensors using machine learning. Third, we analyze VR motion to identify exploratory patterns beyond overall balance-score severity and test whether their participant-level prevalence differs between MS and non-MS groups. The balance score showed moderate between-trial reliability and similar feature-to-score correlation patterns across the MS and non-MS groups. Predictions were evaluated at two levels: the window level, which uses short segments of VR motion-sensor data from the HMD and controllers, and the participant level, which aggregates all segments for a participant into a single participant-level prediction. CatBoost showed the strongest overall predictive performance. At the window level, it achieved R2 = 0.388, MAE = 0.597, and RMSE = 0.782. At the participant level, it achieved R2 = 0.715, Pearson, r = 0.867, and Spearman, ρ = 0.838. The balance score also aligned with motion-capture-derived center-of-mass (CoM) mean velocity, providing additional biomechanical grounding within the available dataset. Motion-pattern analysis identified two distinct exploratory patterns: the Lower head-dominance pattern, characterized by relatively greater hand contribution, and the Head-dominant pattern, characterized by relatively greater head contribution. Although the Head-dominant pattern occurred more frequently at the descriptive window level among participants with MS, the participant-level comparison did not provide statistically significant evidence of a group difference. These findings indicate that VR headset and controller motion alone can support participant-level prediction of balance ability while also providing exploratory insight into the relative organization of head and hand motion during immersive interaction.
Journal IF: unknown (could not be matched)Reference: Pavel MI, Anjum N, Cordova A, Quarles J, Mahmud MR. Predicting Balance Ability and Explaining Control Strategies from VR Motion Sensors: An Interpretable CoP-Derived Balance Score Approach for Multiple Sclerosis. Proceedings of the 2026 28th International ACM SIGACCESS Conference on Computers and Accessibility. 2026 Oct 24 [Epub ahead of print]. doi:10.1145/3797867.3828977.Checked: Abstract only