JMIR Publications

ABSTRACT

Background:

Virtual reality 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:

In this feasibility and proof-of-concept study, we present a decision-support pipeline that integrates wearable physiological signals (heart rate variability, electrodermal activity), behavioral observations, and semantic environment characteristics to provide care staff with data-driven VR environment recommendations on a per-resident basis.

Methods:

Twenty-one residents with major neurocognitive conditions received 420 immersive VR sessions (5 environments, 2 sessions/week, 10 weeks) in a long-term care facility. We defined a hybrid wellbeing target requiring convergence of behavioral calm (PAS-2 vocal intensity = 0) and elevated parasympathetic tone (RMSSD above the participant’s personal median), addressing the ceiling effect of behavioral measures alone (65% of sessions already rated calm). A compact model using 23 features (10 physiological, 3 behavioral, 10 environment) was selected from over 2,100 configurations.

Results:

Environment features carried a statistically significant signal for predicting hybrid calm (permutation test p = 0.045, Cohen’s d = 1.75); four alternative models using 344–370 features ignored environment entirely (all p > 0.79). Person-level effects dominated prediction variance (partial η² = 0.42, p < 0.001), but the environment signal enabled above-chance environment-level discrimination (pairwise ranking accuracy 57.9% vs. 50% chance). Water presence was the most influential environment characteristic. A physiology-only ablation achieved comparable prediction accuracy but produced undifferentiated recommendations — predicting who is calm rather than which environment calms them. The pipeline generated 60 plain-language reports (30 family, 30 staff) via a large language model with evidence tracing; all passed safety checks (100% forbidden-term compliance), though staff reports showed higher data fidelity while family reports better maintained non-medical framing.

Conclusions:

The environment-level signal is subtle but real, person-specific, and clinically interpretable — consistent with the expectation that VR environment choice produces modest rather than pharmacological-magnitude effects in this population.