Data-Driven Virtual Reality Environment Recommendation for Long-Term Care Residents With Major Neurocognitive Conditions: Feasibility and Proof-of-Concept Study
JMIR XR and Spatial Computing · 2026-11-12 · 3 · e101192-e101192
DOI: 10.2196/101192 ↗Astratt
Abstract Background Virtual reality (VR) comfort sessions are increasingly used in residential dementia care, but environment selection remains ad hoc—staff have no systematic way to identify which virtual experience best suits a given resident. Objective This feasibility and proof-of-concept study aimed to determine whether a pipeline integrating wearable physiological signals (heart rate variability and electrodermal activity), behavioral observations, and semantic environment characteristics can detect a person-specific, environment-level signal usable for between-session VR environment recommendation in long-term care and whether the resulting recommendations can be communicated to care staff and families through plain-language reports. Methods Twenty-one residents with major neurocognitive conditions received 420 immersive VR sessions (5 environments, 2 sessions per week, 10 weeks) in a long-term care facility. We defined an operational hybrid calm target requiring behavioral calm (Pittsburgh Agitation Scale vocal intensity score of 0) and root mean square of successive differences at or above the participant’s personal median to address the ceiling effect of behavioral measures alone (sessions already rated calm: 234/358, 65%). A compact 23-feature model (10 physiological, 3 behavioral, and 10 environment) was selected from more than 2100 configurations on the basis of cross-validation error and interpretability; the environment-level signal was tested with an environment permutation test on held-out sessions (weeks 8‐10) and in sensitivity analyses. Results Environment features carried a statistically significant but borderline signal for predicting hybrid calm (environment permutation test, P =.045 with 200 permutations; P =.04 with 2000 permutations; Cohen d =1.75); 4 alternative models using 344 to 368 features ignored environment entirely (all P >.79). Person-level effects dominated prediction variance (partial η²=0.42; P
EUVIMED