Virtual reality (VR) is increasingly used in residential dementia care to offer structured, immersive experiences (eg, nature scenes, familiar places, and guided activities) within short sessions that fit the operational constraints of long-term care settings [1-3]. The rationale is not clinical treatment in the conventional sense but day-to-day support: facilitating engagement, offering moments of calm or interest, and providing structured activities during periods when behavioral and psychological symptoms of dementia (BPSD) might otherwise escalate [4]. A growing body of evidence suggests that such sessions can reduce agitation-related behaviors, improve momentary affect, and are generally well tolerated, although findings remain heterogeneous across protocols, populations, and measurement approaches [5,6].

2,3

A persistent challenge, however, is that responses to VR are heterogeneous both between and within individuals. A resident may appear comfortable during one session and restless during another, even when the content seems similar. The same virtual environment can be experienced differently depending on fatigue, momentary stressors, sensory sensitivity, or cognitive fluctuation [7]. This variability complicates routine practice: care staff typically rely on memory, informal notes, or rotating schedules to select which environment to present, with no systematic mechanism for detecting or acting on individual preferences. The broader nonpharmacological intervention literature has repeatedly identified personalization—tailoring activities to the individual’s history, preferences, and current state—as a key determinant of effectiveness [8,9].

Personalized adaptive interventions are an established class in behavioral health: just-in-time adaptive interventions adjust the type, timing, or amount of support to an individual’s changing internal and contextual state, drawing on continuous data streams to tailor delivery [10]. Within dementia care specifically, algorithmic individualization has empirical support—Cohen-Mansfield et al [11] found that nonpharmacological interventions matched to each resident’s unmet needs reduced agitation in nursing home residents in a placebo-controlled trial, and the Tailored Activity Program (250-dyad randomized controlled trial [RCT]) showed that occupational therapist–led activity tailoring improved functional outcomes and caregiver well-being [12]. These precedents rely on human-administered tailoring at the authoring or delivery stage. This work contributes a data-driven environment recommendation mechanism in that lineage, supporting between-session decisions by care staff rather than executing within-session content adaptation.

Within VR specifically, Li et al [13] provide a recent systematic review of 132 studies on personalized interaction in VR. They describe a 5-stage closed-loop framework for real-time adaptive systems (input collection, data processing, adaptive logic, system update, and feedback loop) and distinguish real-time adaptation, in which the system modifies the experience dynamically during a session, from offline adaptation, in which models use previously collected data to inform subsequent interactions. They further categorize adaptive objectives into 4 types: difficulty adjustment, display adaptation, environment manipulation, and feedback personalization. Across the 132 reviewed studies, most systems operate in real time, rely on machine learning adaptation of physiological and behavioral signals (notably electroencephalography [EEG], galvanic skin response [GSR] or electrodermal activity [EDA], heart rate [HR] or heart rate variability [HRV], and ocular data), and concentrate in mental health, education and training, gaming, and rehabilitation domains, targeting cognitively intact users. Residential long-term care for residents with major neurocognitive conditions is not a represented application domain in their review, and between-session environment selection from a curated library—as distinct from within-session environment manipulation—is not an explicit category within their taxonomy.

Within dementia care specifically, personalization has developed along 2 largely separate tracks. The first track creates memory-specific custom content through qualitative elicitation with residents and families, either by filming personally meaningful locations [14,15] or curating scene libraries from family-supplied materials [16]; this personalizes content at the authoring stage but does not use a resident’s own session-to-session response data to choose between preauthored options. The second applies person-specific machine learning to wearable physiological and behavioral data for BPSD detection—Iaboni et al [7] showed that person-specific multimodal models classify agitation episodes substantially better than population-level models, with the most predictive features varying across residents. These detection models answer “Is this resident agitated now?” rather than “Which environment is most likely to keep them calm next session?”—they do not select interventions. HRV in particular has been independently validated as an indicator of autonomic dysregulation and agitation propensity in Alzheimer disease [17,18]. The gap is specific: no system combines person-specific physiological modeling with semantic environment characteristics for data-driven between-session environment recommendation from a curated VR library in residents with major neurocognitive conditions.

The absence of systematic environment selection carries a direct clinical cost. In long-term care, BPSD, including agitation, vocal distress, and social withdrawal, are among the primary drivers of pharmacological intervention, despite guidelines that prioritize nonpharmacological approaches as first-line responses. A poorly matched activity is not merely neutral: it can actively increase distress or accelerate behavioral escalation. In the context of VR, assigning environments by rotating schedule or staff memory risks delivering sessions that produce no benefit or cause discomfort, with no mechanism to detect this mismatch or correct it over time. Two constraints specific to this population sharpen the problem. First, residents with major neurocognitive conditions cannot reliably self-report preferences. Second, behavioral observation alone is insufficient to distinguish comfort from passive withdrawal, a ceiling effect that limits what staff can infer from watching a session. The data needed to improve selection (physiological signals from wearable sensors, behavioral ratings from trained observers, and the semantic characteristics of the VR content itself) are either already collected or collectable at low additional cost during routine sessions.

Physiological wearables offer a complementary, continuous trace of the body’s response during a session. Photoplethysmography (PPG)–derived HRV and EDA can provide time-resolved indicators of autonomic state and arousal regulation without requiring the resident to articulate a subjective experience. HRV, in particular, has been linked to emotion regulation capacity and agitation propensity in Alzheimer disease [17,18], and wearable multimodal sensors have shown promise for detecting behavioral symptoms in dementia care units [7]. However, these signals are notoriously difficult to use in real-world care environments: devices are worn inconsistently, motion artifacts are frequent, recordings contain gaps, and exports are often incomplete [19]. Moreover, observer-rated behavioral measures such as the Pittsburgh Agitation Scale (PAS) remain subjective and coarse; in our own data, nearly two-thirds of sessions were already rated as behaviorally calm, leaving limited room for between-environment differentiation using behavioral scores alone. Combining physiological and behavioral information into a stronger, convergent indicator of comfort is both a methodological opportunity and a necessity.

At the same time, even a well-performing predictive model is of limited clinical value if its outputs cannot be communicated to the people who make care decisions. Families want to know how their relative experienced a session; staff need structured summaries that support documentation and planning. Large language models (LLMs) have demonstrated the ability to generate fluent, structured text from numerical inputs, but their application to dementia care reporting, where nonmedical framing, evidence tracing, and avoidance of clinical overclaiming are essential, has not been explored.

Immersive VR is being explored in long-term care as a way to deliver structured sensory experiences (nature exposure, cultural scenes, reminiscence cues, and light interactive activities) to residents who can no longer access such experiences in the physical world [6,20]. Sessions are typically brief (5‐20 min), supervised by staff, and delivered through head-mounted displays. The therapeutic rationale is not cognitive rehabilitation but moment-to-moment well-being: facilitating engagement, eliciting positive affect, and reducing the frequency or severity of BPSD during and after exposure [5,21].

Evidence on outcomes remains mixed. A recent systematic review and meta-analysis found that VR may reduce depressive symptoms and aggression but reported no significant effect on agitation as measured by standard instruments, with high heterogeneity across included studies [5]. An RCT in acute care found that VR significantly reduced aggressive behavior and loud vocalizations but no other BPSD domains [1]. Qualitative studies consistently report that residents enjoy the experiences and that sessions can prompt social interaction with staff and family members [22,23]. Feasibility is generally supported: headsets are tolerated by most residents, including those with moderate-to-severe dementia, although comfort, fit, and cybersickness remain concerns for subsets of users [20,24].

A construct central to VR research but rarely examined in dementia care is the sense of presence—the subjective experience of being in the virtual environment rather than in the room where the headset is worn. Presence is conventionally assessed with postexposure questionnaires, which residents with major neurocognitive conditions cannot reliably complete. Neurophysiological approaches that characterize presence from combined electroencephalographic and autonomic recordings [25] and that compare neural and autonomic responses to virtual and physical versions of the same stimulus [26] offer a possible route to objective assessment, although they have so far been validated only in cognitively intact adults. This study did not measure presence; the implications of this omission for interpreting between-environment differences in physiological response are discussed in the Limitations section.

A notable gap in this literature is the absence of systematic personalization. Studies typically expose all participants to the same VR content or allow free selection without structured tracking. Residents who expressed preferences for tailored, culturally relevant content showed greater engagement [22], and intervention research in dementia care more broadly has identified personalization as a key determinant of effectiveness across nonpharmacological approaches [8,9]. Yet no study has used physiological or behavioral data to match specific environments to specific residents. This work addresses this gap directly.

HRV, the beat-to-beat variation in cardiac intervals, is widely used as an index of autonomic nervous system function and, by extension, of emotion regulation capacity [27]. In populations with dementia, reduced HRV has been associated with disease severity, with a systematic review and meta-analysis confirming lower parasympathetic activity in people with dementia compared to healthy controls, particularly for time-domain measures such as root mean square of successive differences (RMSSD) and SD of normal-to-normal RR intervals (SDNN) [17]. More directly relevant to this study, Liu et al [18] used data from the Atherosclerosis Risk in Communities cohort to demonstrate that declining HRV across 2 decades was associated with increased risk of agitation in people with clinically diagnosed Alzheimer disease, supporting the plausibility of HRV as a physiological marker of agitation propensity.

Wearable multimodal sensors, combining PPG for cardiac signals with EDA for sympathetic arousal, have been applied in dementia care units to detect agitation episodes. Iaboni et al [7] fitted 17 residents with wristbands for up to 8 weeks and developed person-specific machine learning models that classified behavioral symptoms with a median area under the curve of 0.87, with feature importance varying substantially across individuals. This work demonstrated the value of person-specific modeling and of combining multiple physiological modalities, but it focused on detecting agitation events rather than predicting comfort or recommending activities.

Beyond dementia care, neurophysiological measurement is increasingly used to evaluate immersive experiences objectively rather than through self-report alone. Recent work has proposed EEG-derived indices to track skill acquisition during virtual training [28], evaluated haptic exoskeletons intended to add force feedback to VR training [29], and combined electroencephalographic and autonomic recordings to compare responses to a virtual and a physical cultural artifact [26], in addition to the work cited earlier [25]. These studies concern cognitively intact adults in training or cultural contexts, but they support the premise of this work, that physiological signals recorded during immersion carry interpretable information about how a user is responding to a virtual environment, while also highlighting how little of this literature has reached residential dementia care.

The practical challenges of physiological monitoring in this population are considerable. Motion artifacts from involuntary movements are common, sensor wear is inconsistent, and recording quality varies across sessions. EDA is particularly prone to gaps and artifacts in older adults with thin, dry skin [7,19]. Any system that claims to use physiological data for clinical decision support must therefore include explicit quality control and must distinguish between measured and reconstructed signals, a principle that guides the preprocessing pipeline described in the Methods section.

A final limitation of existing approaches is reliance on behavioral agitation measures alone as outcomes. Instruments such as the PAS capture observable distress but miss internal states: a resident may appear behaviorally calm while showing elevated sympathetic arousal (withdrawn rather than comfortable) or may vocalize without physiological distress markers (engaged rather than agitated). This ambiguity motivates the hybrid calm target used in this study, which requires convergence of behavioral and physiological indicators.

Translating predictive model outputs into actionable guidance for care practitioners remains a largely unsolved problem in applied health technology. In dementia care, this challenge is compounded by the need to communicate with 2 distinct audiences: clinical staff, who need structured summaries to support documentation and session planning; and family members, who want accessible, nonmedical accounts of their relative’s experiences [30].

LLMs have demonstrated capabilities in health care report generation, including transforming pathology reports into patient-friendly language [31], drafting responses to patient portal messages [32], and generating clinical summaries from structured data [33]. Evidence suggests that LLM-generated reports can improve patient comprehension and reduce the time clinicians spend on communication tasks [31]. However, concerns about factual accuracy, hallucination, and the risk of medical overclaiming remain central to responsible deployment [34,35]. Large-scale review evidence indicates that despite rapid growth in clinical LLM evaluation studies, rigorous patient-centered evaluations remain scarce and existing evidence is heavily weighted toward simulated or examination-style tasks, justifying continued caution about clinical deployment [36].

In dementia care specifically, no prior work has used language models to generate reports from physiological monitoring or predictive model outputs. This study introduces an evidence-tracing framework in which every factual claim in a generated report is linked to a specific data source (a physiological measurement, a behavioral observation, or a model prediction). This design explicitly constrains the language model to nonmedical framing: the system describes patterns and preferences, never diagnoses, prescribes, or makes clinical recommendations. The evaluation strategy assesses both the accuracy of this evidence tracing and the model’s adherence to nonmedical language boundaries, separately for family-audience and staff-audience reports.

The goal of this feasibility and proof-of-concept study was therefore to determine whether passively collected physiological signals, behavioral observations, and semantic environment characteristics can support data-driven, between-session VR environment recommendation for residents with major neurocognitive conditions in long-term care and whether the resulting recommendations can be communicated transparently to families and care staff. Specifically, we tested 3 prospective hypotheses: that a model given environment features would use them to produce environment-specific recommendations whereas a physiology-only model would not (H1); that a hybrid behavioral-physiological target would differentiate between-environment effects better than a behavioral target alone (H2); and that water presence would be the strongest environment-level differentiator (H3). We additionally explored, descriptively, whether an LLM can generate evidence-traced, nonmedical reports from the pipeline’s outputs (RQ3). The study was designed to establish technical feasibility and to obtain a preliminary estimate of the environment-level signal; it was not designed to show that the recommendations improve resident outcomes, and no such claim is made. The report generation component is a prototype, and prospective validation against resident outcomes is identified as necessary future work.

Twenty-one residents of a long-term care facility in Québec, Canada, participated in this study. All had a documented diagnosis of a major neurocognitive condition. Demographic and clinical characteristics are summarized in Table 2. The mean age was 82.6 (SD 6.2; range 68‐97) years; age was missing for one participant. Twelve (57%) participants were female; sex was not documented for 1 (5%) participant. Cognitive diagnoses included Alzheimer disease (n=13, 62%), major neurocognitive condition of mixed or undetermined etiology (n=4, 19%), mixed dementia (n=2, 10%), and vascular dementia (n=1, 5%); the diagnosis was not documented for 1 (5%) participant. Functional status was assessed using the ISO-SMAF (ISO-classified Système de mesure de l’autonomie fonctionnelle [functional autonomy measurement system]) classification system [41] (mean profile 7.2, SD 0.7; range 7‐10; 1 value missing), indicating moderate-to-severe care needs. Most participants (n=18, 86%) could walk without aid; 2 (10%) used a walker; mobility was not documented for 1 participant. Seventeen (81%) participants wore corrective lenses, and 1 (5%) used a hearing aid. The hearing aid count reflects recorded hearing aid use rather than hearing status: hearing acuity was not formally assessed, and hearing aid use among long-term care residents is known to be low relative to the prevalence of hearing loss in this age group [42]. As the environments’ soundscapes were delivered through the headset’s audio, uncorrected hearing loss may have attenuated the audio dimensions of the environment encoding for some participants; this is noted in the Limitations section.

The study was approved by the Research Ethics Board (REB) of the Centre intégré universitaire de santé et de services sociaux (CIUSSS) de la Capitale-Nationale (protocol 2022‐2299, RIS_ForetStress) and conducted in accordance with institutional guidelines for research involving adults with cognitive impairment. As residents with major neurocognitive conditions cannot, in general, provide fully informed prospective consent, consent for participation followed the substitute decision-maker framework prescribed by institutional policy, with families notified through the facility’s standard communication channel. VR sessions were integrated into the facility’s existing leisure programming and delivered under the same care decision framework that governs other activities of daily living, rather than as a separate research-only procedure.

Per-session willingness was verified at each visit by a member of the recreation team together with the researcher, using ordinary, nontechnical language; residents who declined were not approached further, and sessions were terminated immediately at any sign of discomfort, agitation, or fatigue. Trained staff supervised all sessions and could terminate at any time. No monetary or material compensation was offered to participants or families: integration with the existing leisure programming was intended to remove any external incentive that might compete with a resident’s moment-to-moment willingness to participate.

All data were processed under coded participant identifiers (P01-P21); raw physiological recordings and individual-level behavioral ratings were retained within the institutional research infrastructure approved under the REB protocol, and no directly identifying information left that environment. Residual ethical questions arising from continuous physiological monitoring in a population that cannot fully articulate its preferences are discussed in the Limitations section.

Each participant received 2 immersive VR sessions per week over 10 consecutive weeks, yielding 420 total sessions (21 participants×20 sessions). Sessions were distributed across 5 virtual environments in a balanced rotation (84 sessions per environment): beach with kite surfers, Île d’Orléans countryside, forest and stream, Old Québec city walk, and active forest experience. Four environments were delivered using a white headset (Meta Quest 2; Meta Platforms; passive 360° video); one (active forest) used a blue headset (HTC Vive Cosmos; HTC Corporation) that supported a limited interactive element, specifically a trigger for forward advance; reverse movement was obtained by pivoting the resident’s chair. The blue headset is not an independent variable confounded with interactivity; it is the apparatus through which limited interaction was delivered, and the interaction demand dimension of the semantic encoding is designed to capture precisely this design variation. Interpretation of findings along the interaction demand dimension should nonetheless be made cautiously, because active forest is the only environment in the library with a nonzero interaction demand rating: any effect attributed to this feature in the importance analysis (reported in RQ2) is estimated from a single environment of contrast and cannot be distinguished from any other characteristic of active forest that the encoding did not capture. Water presence and social presence, which vary across the 4 white-headset environments, are not subject to this limitation. Sessions were supervised by trained staff who monitored the resident’s comfort and could terminate the session at any time. The mean recorded session duration was 11.4 (SD 3.0; range <1 to 25.7) minutes based on 371 sessions with wearable physiological recordings.

To characterize the perceptual properties of each VR environment, we developed a 10-dimension semantic encoding scheme covering visual complexity, motion (camera and scene), social presence, audio characteristics (nature and music), water presence, visual openness, interaction demand, and altitude. Two raters independently rated each environment on an ordinal 0 to 3 scale per dimension [43], normalized to [0, 1] for model input. Interrater reliability was handled via discussion-based reconciliation, given the small number of items (5 environments×10 dimensions). Full dimension definitions and rating rubrics are provided in Multimedia Appendix 1.

Independent prereconciliation ratings were not retained in machine-readable form, which precludes formal interrater reliability calculation (see Limitations section).

This encoding provides continuous, interpretable features that can mediate between-environment differences in physiological response and supports principled predictions for new environments not present during training (rate the new environment on the 10 dimensions and concatenate with physiological or behavioral inputs at inference time).

Physiological signals were recorded during each VR session using an EmotiBit wearable sensor (EmotiBit; Connected Future Labs) [44] attached to the participant’s forearm. The device captured PPG for cardiac signal derivation and EDA for sympathetic arousal monitoring. Of the 420 scheduled sessions, 374 (89%) yielded a matched physiological recording. Matching was performed automatically by aligning device time stamps with session scheduling logs, with manual verification for ambiguous cases.

Signals were processed through a dedicated quality control pipeline. Raw PPG recordings were segmented into analysis windows, and each window was evaluated for motion artifact, signal saturation, and interbeat interval plausibility. Windows failing quality criteria were excluded rather than reconstructed. From the usable windows, the pipeline extracted session-level summary features: mean HR, HRV (RMSSD, SDNN, and percentage of successive RR intervals differing by more than 50 ms [pNN50]), respiratory sinus arrhythmia (RSA) via Welch periodogram and Lomb-Scargle methods (RSA-HF), and window-level statistics (mean, median, SD, minimum, and maximum) for each base measure.

For PPG-derived features, 371 (99.2%) of 374 recorded sessions passed quality thresholds, with a mean usable window proportion of 87.7% (SD 22.6%). EDA processing used the configuration selected by the preprocessing search, with interpolation limited to gaps of ≤1 second: 218 (52%) of 420 sessions produced usable EDA, with a mean usable window proportion of 82.2% (SD 25.8%). From usable EDA segments, the pipeline extracted tonic skin conductance level (mean and SD) and phasic component features (root mean square amplitude and skin conductance response count).

The pipeline retained quality indicators as audit variables (eg, usable window percentage, motion artifact fraction, and device-algorithm HR disagreement) but did not include these as input features of the focal 23-feature model. Session duration, window counts, and the proportion of usable recording were preserved as session metadata and entered only the physiology-only feature set used for the ablation baseline (see Predictive Modeling section). Supplementary figures on study design, data quality, signal processing, and model diagnostics are provided in Multimedia Appendix 2. Full preprocessing parameters and quality control criteria are provided in Multimedia Appendix 1.

Two standardized instruments provided behavioral data for each participant.

Standard behavioral agitation instruments capture observable distress but miss important ambiguities. A resident who appears behaviorally still may be physiologically aroused—withdrawn rather than comfortable. Conversely, a resident who vocalizes during a VR session may be physiologically relaxed—engaged rather than distressed. In our data, these ambiguities are compounded by a ceiling effect: 65.4% (234/358) of rated sessions already scored as behaviorally calm (PAS-2 vocal intensity score of 0), leaving limited room for between-environment differentiation using behavioral scores alone.

To address this, we defined a hybrid calm target that requires convergence of 2 independent signals: behavioral calm and elevated parasympathetic tone. A session is classified as calm in this operational sense (y=1) only if both conditions are met simultaneously: y_hybrid=1 if (PAS-2 vocal intensity=0) AND (RMSSD≥participant’s personal median RMSSD); y_hybrid=0 otherwise.

The behavioral component (a PAS-2 vocal score of 0) ensures the resident showed no observable vocal agitation. The physiological component (RMSSD at or above the participant’s own median) ensures the resident’s HRV was elevated relative to their personal baseline, indicating parasympathetic activation consistent with a relaxed autonomic state. The within-person median split accounts for the substantial between-person differences in resting autonomic tone that characterize this population [17]: what constitutes “elevated” RMSSD for one resident may be typical for another. Three further considerations motivated this choice over a fixed absolute threshold or a population cut point. First, with 6 to 20 usable sessions per participant, the median is the only within-person cut point that is robust to the skewed distribution of RMSSD without distributional assumptions. Second, a within-person split deliberately removes between-person differences in level from the physiological component of the target: person identity alone cannot predict whether a session falls above or below that person’s median, so any predictability of this component must come from within-person variation—including the environment presented—rather than from stable individual differences. Third, the same construction was applied uniformly to the RSA, RMSSD, and pNN50 variants evaluated in model selection, so that the variants differed only in the physiological index used. The within-person median was computed over all of a participant’s sessions, including those in the test period (weeks 8‐10), on the imputed and scaled RMSSD series; training labels therefore depend slightly on test period data, a point noted in the Limitations section.

This target is a constructed operational criterion, not an independently validated measure of well-being, and the term “calm” is used throughout as shorthand for meeting it. Two limitations follow directly from the median-split construction. By design, approximately half of each participant’s sessions satisfy the physiological component, whether or not the absolute within-person differences in RMSSD are large; and sessions whose RMSSD lies close to the participant’s median receive labels that are sensitive to measurement noise. Sensitivity analyses using alternative within-person thresholds and a continuous physiological component are reported in the Sensitivity of the Environment-Level Result section.

Within-person medians require an adequate number of RMSSD-bearing sessions per participant to be well defined. In this cohort, RMSSD was available for 369 (87.9%) of 420 sessions; per-participant counts ranged from 6 to 20 (mean 17.6, SD 3.3; median 19.0, IQR 16‐19). No participant fell below 5 RMSSD-bearing sessions, so within-person medians are defined for all 21 residents. Precision varies, however: for the most affected participant (P12, with 6 RMSSD-bearing sessions and 4 labeled sessions with observed RMSSD), the within-person median is more sensitive to individual observations than for participants with near-complete data, and this resident’s hybrid-target labels should be interpreted with correspondingly greater caution.

Three physiological variants of this hybrid target were tested during model selection: RSA via Welch periodogram (T5), RMSSD (T6), and pNN50 (T7). All 3 are time-domain or frequency-domain indicators of parasympathetic cardiac modulation. The selected model uses the RMSSD variant (T6). The choice among variants was determined empirically by validation performance; all 3 hybrid targets outperformed purely behavioral targets (PAS-2 vocal alone, summed agitation, and binary calm) in the automated search (see Predictive Modeling section).

Of 420 scheduled sessions, 358 (85.2%) carried a PAS-2 vocal intensity rating and received a hybrid-target label for supervised learning. For 35 of these, RMSSD was not observed and the physiological component was evaluated on the imputed value (the training period median), so their labels are less certain than the 323 (76.9%) based on observed RMSSD. The 62 (14.8%) sessions without a PAS-2 rating received a missing target mask: they were excluded from loss computation during training but retained in the input sequences to preserve each resident’s temporal context.

As defined in the Introduction, the modeling objective is environment-level discrimination: for each resident, the model must use environment features alongside physiological and behavioral inputs to rank the 5 VR environments by predicted probability of meeting the hybrid calm target. A model that captures only person-level effects, stable between-resident differences in baseline physiological and behavioral state, would not achieve this objective, even if its aggregate predictive accuracy were high. The physiology-only ablation model specified later in this section, together with the environment-permutation test described in the Evaluation Strategy section, is designed specifically to disentangle these 2 sources of variance.

The hypersearch evaluated 2108 configurations across 7 prediction targets, 6 feature sets, and 5 architectural families (long short-term memory, gated recurrent unit, lightweight transformer, ridge regression, and random forest; S1-S5 in Multimedia Appendix 1), each evaluated by 3-fold expanding-window cross-validation within the training period (weeks 1‐7); mean cross-validation mean absolute error (MAE) across folds was the ranking criterion. Training was not seeded, so individual fits are not reproducible bit for bit. The monitored cross-validation MAE averages the masked error over all time steps of the validation sequences, including training-week history cells, and therefore understates validation-week error; it is reported here as the search’s ranking score rather than as an estimate of out-of-sample error. The full grid specification is in Multimedia Appendix 1.

Models were trained on sessions from weeks 1 to 7 (294 sessions) and evaluated on sessions from weeks 8 to 10 (126 sessions, 6 per participant); MAE-based metrics use the 108 test sessions with a defined label, whereas the ranking metrics use all 126. Each participant’s sessions formed one chronological sequence with full history retained; the prediction for a given session uses that session’s own physiological, behavioral, and environment features together with all preceding sessions as context (causal attention), so the model estimates calm for the current session rather than forecasting it from past sessions alone. Environment recommendation is obtained by rescoring a session under each of the 5 environment encodings (counterfactual prediction), as shown in Figures 1 and 2. Within the training period, configurations were ranked by MAE across temporal cross-validation folds. The test set was reserved for final evaluation and was not used in the hypersearch ranking; its role in the choice of the focal configuration among the retained candidates is described below.

From the hypersearch, 5 environment-aware configurations and the physiology-only ablation (6 in total) were retained for full evaluation. Among the environment-aware feature sets, the best configuration for each (target, feature set) pair was identified and the 5 best of these representatives by cross-validation MAE were retained (test MAE 0.247‐0.304); the best physiology-only configuration was added as a no-environment ablation baseline (98 features: physiological summary statistics—PPG HR, HRV time domain, RSA spectral, EDA, and beat counts—and 4 session-level recording metadata features, with no environment or behavioral features). The ablation configuration was also the best configuration of the entire search by cross-validation MAE. The selected focal model uses 23 input features—10 physiological summary features (HRV time- and frequency-domain, EDA), 3 behavioral observation features (PAS-2 vocal intensity, QUALID smile, and QUALID calm or vocal discomfort), and 10 semantic environment features—implemented in TensorFlow 2.20 and Keras 3.10 (Google) as a small transformer [47] (1 encoder layer, hidden dimension 64, dropout 0.3, and learning rate 0.001) trained with binary cross-entropy on the hybrid RMSSD target (T6; see Defining the Hybrid Calm Target section). Three considerations entered the choice of this focal configuration among the retained candidates, and it is important to distinguish them. The first 2 are independent of the test set: the compact model’s cross-validation MAE (0.114) was within 0.02 of the other retained configurations (344‐368 features, 0.095‐0.111), although it was the lowest-ranked of the 5 on that criterion, and it was the only retained configuration compact enough for the per-feature importance analysis required by RQ2 to be interpretable (23 vs 344‐368 features); parsimony alone therefore singles it out, whereas cross-validation MAE alone would not. The third consideration was not independent of the test set: the environment permutation test (described in the Evaluation Strategy section) had been computed on the test set, and the finding that the compact model—unlike the larger configurations—retained environment sensitivity was known when the focal choice was fixed. The test-set environment permutation result for the focal model should therefore be read as a confirmatory test performed under a configuration whose selection was partly informed by that same test set, not as a fully independent replication; the Limitations section returns to this point. Larger configurations were retained as ablation baselines.

The selected model and its ablation baseline were evaluated using 6 complementary approaches, each addressing a different aspect of model validity. Two sets of sensitivity analyses, added at revision in response to peer review and not prespecified, are described at the end of this subsection.

Of 420 scheduled VR sessions, 358 (85.2%) received postsession PAS-2 ratings from trained observers. Among rated sessions, 234 (65.4%) scored 0 on the vocal intensity scale (agitation absent), 116 (32.4%) scored 1 (low-volume, nondisruptive vocalization), and 8 (2.2%) scored 2 (moderately disruptive). No session reached the 2 highest agitation levels (scores 3 or 4). Physiological recordings were obtained for 374 sessions (89%); of these, 371 passed PPG quality control, and 218 produced usable EDA. These figures confirm that immersive VR comfort sessions were generally well tolerated in this cohort. They also illustrate the ceiling effect that motivated the hybrid target: with nearly two-thirds of sessions already scoring as behaviorally calm, the PAS-2 vocal item alone provides limited between-environment differentiation.

This section tests hypothesis 2 (hybrid target) via the model search results described in the Predictive Modeling section and feature-family permutation importance reported below, and tests hypothesis 3 (water presence) via the individual-feature permutation importance in Table 4.

Two sets of analyses, added at revision in response to peer review and not prespecified, examined how far the environment-level result depends on the definition of the target and on the way the focal configuration was selected (see Evaluation Strategy section).

The hybrid label was robust to the threshold. Alternative within-person cut points (40th percentile, 60th percentile, and upper tertile) changed 19 to 31 of the 358 labels (κ=0.79‐0.89 relative to the median split), and a 10% dead band around the median removed 98 near-median sessions without changing any retained label (Table 5). The environment permutation result was not robust to the same changes. Rescoring the trained model’s unchanged test predictions against each alternative labeling, shuffling the environment features still increased the error in every variant except the dead-band one (d=0.93‐1.37), but none reached significance (P=.10‐.20; dead band P=.57; d=−0.13). Retraining the focal configuration with a continuous target—within-person z-scored session RMSSD on PAS-2-calm sessions—gave the same picture across 3 seeds: environment shuffling increased the error (d=0.65‐1.60), with P=.06, .23, and .25 (Table 5). The direction of the effect was therefore consistent, but its statistical significance was specific to the median-split labeling of the single trained model reported under RQ1.

A selection criterion computed on training period data only did not select the compact model. When the 5 retained configurations were retrained on weeks 1 to 7 and the environment permutation test was computed on validation-week cells, no configuration showed significant environment sensitivity, and the compact model showed the least: shuffling its environment features slightly improved its validation error (P=.80‐.83; d=−1.02 and −0.80), whereas the 344-feature rank-1 configuration came closest to significance (P=.15‐.19; d=1.04 and 0.86; Table 3). By monitored cross-validation MAE alone, the criterion would have selected the physiology-only configuration and, among environment-aware configurations, the 344-feature rank-1 configuration. The validation-week-only MAE (0.25‐0.32 for every configuration) also confirms that the monitored cross-validation scores used to rank the search (0.09‐0.11) understate out-of-sample error, as noted in the Methods section.

Taken together, the environment-level signal reported under RQ1 should be read as exploratory. Its direction was consistent across target definitions and seeds, but its statistical significance was not reproducible under alternative labelings, and the property that motivated the focal choice—sensitivity to environment features—was not visible on training period data. Two limitations of these analyses are noted: the threshold analysis rescored a model trained on the median-split labels rather than retraining per threshold, which may penalize the alternative labelings; and the retraining analyses used 2 or 3 seeds of an unseeded pipeline, so their P values are indicative rather than precise.

This study set out to determine whether a data-driven system could identify which VR environment produces the calmest experience for a given resident with a major neurocognitive condition, using wearable physiology, behavioral observations, and semantic environment descriptors. Three findings ground the contribution.

First, an environment-level signal is statistically detectable from the available data. The compact 23-feature model showed a significant drop in prediction accuracy when environment features were permuted (P=.045; Cohen d=1.75), while the 4 larger environment-aware configurations (344 to 368 features) ignored environment entirely (all P>.79). The signal is small—person-level variance dominated at partial η²=0.42—and borderline in statistical terms, and it was recovered only when the feature set was constrained to keep environment information from being buried by physiology and missingness patterns. As the focal configuration was selected in part with knowledge of this test-set result and because sensitivity analyses showed the significance to be specific to the median-split labeling while the direction of the effect was consistent, the finding is exploratory and requires independent replication.

Second, the hybrid behavioral-physiological target, as operationalized here, was both necessary and sufficient for that signal to emerge. No purely behavioral target was among the 5 configurations retained from more than 2100 evaluated: with 65% of sessions already rated as behaviorally calm, PAS-2 vocal intensity alone has too little variance to differentiate environments. Among the environment dimensions, water presence carried the most weight, ranking first among the 10 environment features and fifth overall.

Third, automated plain language reporting from the model’s outputs is technically feasible, with a documented audience-specific trade-off: staff-targeted reports achieved higher fidelity ratings (4.70/5) but drifted toward clinical register (compliance 2.83/5), while family-targeted reports stayed firmly nonmedical (4.37/5) but sacrificed precision (3.77/5). All 60 generated reports passed binary safety checks.

Two distinct claims warrant separation. The first—that the model uses environment information at inference time, in a person-specific way, and that its outputs can be surfaced in plain-language reports—is what the primary analyses suggest, with the qualifications documented in the sensitivity analyses. The second—that the rankings the model produces are clinically optimal, in the sense that the environment placed first for a given resident is in fact the one most likely to benefit them in routine care—is not and would require prospective validation against in-session distress outcomes in a held-out cohort. We frame the present results as motivation and effect-size estimate for such a validation study, not as evidence that the system is ready to inform care decisions.

The recommendation pipeline reported here occupies an operational niche left open by 4 adjacent literatures. Within VR personalization specifically, the systematic review by Li et al [13] of 132 studies cataloged real-time adaptation in cognitively intact users, with most systems operating during a session via EEG, GSR or EDA, or HR signals. Between-session environment selection from a curated library—what the present pipeline does—was not an explicit category in their taxonomy, and residential long-term care for residents with major neurocognitive conditions was not a represented application domain. The contribution here is therefore at a complementary timescale and population to the dominant real-time adaptation paradigm.

Within wearable-physiology work in dementia care, Iaboni et al [7] established that person-specific multimodal models can classify behavioral symptoms substantially better than population-level models. The present pipeline extends that finding in 2 directions: it adds semantic environment features to the same physiological-modeling pattern, and it shifts the prediction target from agitation detection (“Is this resident agitated now?”) to environment selection (“Which environment is most likely to keep them calm next session?”). The dominance of person-level variance reported here (partial η²=0.42) is consistent with the person-specific result of Iaboni et al. What is added is preliminary evidence that, layered on top of person-specific physiology, a smaller environment-level signal may be recoverable when feature parsimony preserves it—a methodological point with practical consequences: in applied dementia care machine learning, exhaustive feature extraction can bury the signal of interest, and the choice of feature-set size is itself an RQ-aligned design decision rather than a regularization detail.

Within the tailoring tradition in dementia care, Cohen-Mansfield et al [11] and the Tailored Activity Program of Gitlin et al [12] established that individualized nonpharmacological interventions, matched to each resident’s unmet needs or preferences, improve agitation and functional outcomes. Both relied on human-administered tailoring at the authoring or delivery stage. The present pipeline contributes a data-driven environment recommendation mechanism in that lineage—supporting between-session decisions by care staff with a resident’s own session-to-session response data, rather than executing within-session content adaptation. The hybrid behavioral-physiological target used here extends naturally to other structured activities in dementia care where observer ratings are available alongside physiological recording—music therapy, multisensory stimulation, and outdoor walks—in which neither channel alone is sufficient to define comfort [7,18].

Finally, the prominence of water presence among environment features replicates, at the feature importance level, the result of Tanja-Dijkstra et al [40], who found that VR beach walks produced greater anxiety reductions than VR city walks in a randomized trial. The 3 environments with the highest water-presence ratings (forest and stream 1.0, beach 0.67, and Île d’Orléans 0.33, tied with active forest) showed the most differentiated calm responses across residents here. The convergence of these 2 independently obtained results—a randomized between-environment effect in cognitively intact users, and a feature-importance ranking in residents with major neurocognitive conditions—is consistent with the possibility that the blue-space exposure literature [38,39] transfers to immersive VR in dementia care, although the present feature-importance result is preliminary. Social presence and interaction demand also contributed positively, suggesting that the social and interactive character of a virtual scene carries person-specific information; for content developers, a tentative implication is that water elements, social density, and interactivity level deserve attention as design variables when curating VR libraries for this population. The audience-specific trade-off in automated reporting—fidelity preserved for staff at the cost of clinical drift, nonmedical register preserved for families at the cost of precision—is, to our knowledge, the first quantified demonstration of this tension in dementia care LLM applications and is consistent with broader cautions about LLMs in health care communication [32,33].

Several limitations should be considered when interpreting these results.

Generalizability of the findings is constrained along 3 specific dimensions. First, the VR environment library comprises 5 scenes drawn from a specific production context: these environments sample a narrow slice of the space of possible dementia care VR content, and results should not be extrapolated to commercial or substantially larger libraries without separate evaluation. Second, the 10-dimension semantic encoding used to represent environments was derived inductively from this specific library; while the dimensions are plausible on a priori grounds, their mapping onto different content—therapeutic reminiscence scenes, interactive games, music-driven experiences—has not been validated, and the transferability of the encoding protocol itself remains an open question. Third, the study cohort comprises 21 residents in a single long-term care facility in Québec, predominantly with Alzheimer disease, at ISO-SMAF levels 7 to 10 (marked to severe functional disability); findings cannot be assumed to generalize to earlier disease stages, different cognitive etiologies, or different cultural and care system contexts. Replication across multiple sites, broader cohorts, and independent content libraries is a necessary prerequisite to any clinical recommendation.

Only 4 of 5 environments appeared in the test period (weeks 8‐10); the beach environment was not scheduled during those weeks. All ranking evaluations are therefore 4-way comparisons, and the model’s ability to recommend the missing environment remains untested.

The environment permutation test yielded a P value of .045 (P=.04 with 2000 permutations)—statistically significant but borderline, obtained from a single model configuration, and accompanied by a top-1 hit rate CI (19.8%‐35.7%) that includes the chance baseline. In addition, as stated in the Methods section, the choice of the focal configuration was partly informed by the test-set environment permutation result, so the test set is not fully independent for this inference. Shuffling feature columns after training cannot retroactively change what the model learned, so the test remains informative about whether the trained model uses environment information; what it cannot provide is an unbiased estimate of how often such a signal would be detected in a configuration chosen without reference to the test set. The remedy—a selection criterion defined on training period data only—was applied at revision: it does not select the compact model, and no retained configuration shows significant environment sensitivity on validation folds (Table 3). The test set result therefore stands as an exploratory finding, and the property that motivated the focal choice was not reproducible from training period data. We interpret the present results as preliminary evidence that an environment-level signal may exist—its direction was consistent across labelings and seeds—not that it is large, robust, or clinically useful.

The hybrid target is a constructed criterion rather than a validated well-being measure. Its within-person median split labels approximately half of each participant’s sessions as physiologically “better” by construction, irrespective of the magnitude of the underlying differences in RMSSD, and labels for sessions close to a participant’s median are sensitive to measurement noise. The sensitivity analyses (Table 5) show that the label itself is robust to the threshold (κ=0.79‐0.89) but that the environment permutation significance is not: alternative thresholds and a continuous target preserved the direction of the effect without reaching significance. Two further caveats apply: the threshold analysis rescored a model trained on the median-split labels rather than retraining per threshold, which may penalize the alternative labelings; and the within-person median was computed over all sessions, including the test period, so training labels depend slightly on test period data. “Calm” should therefore be read strictly as the operational label defined in the Methods section, and not as a direct measure of the resident’s subjective state.

Sense of presence was not measured. Between-environment differences in physiological response could therefore reflect differences in presence or immersion—driven, for example, by camera motion, visual complexity, or the interactive element of active forest—rather than differences in comfort per se. The semantic encoding captures some of these properties, but the present design cannot separate presence from calm. Questionnaire-based presence measures are impractical in this population, and neurophysiological presence indices [25,26] have so far been validated only in cognitively intact adults; adapting them to residents with major neurocognitive conditions would be a prerequisite for disentangling the 2 in future work.

Sensory status was characterized only by recorded use of corrective lenses and hearing aids. Hearing was not formally assessed, and because only one participant used a hearing aid in a population in which hearing loss is common, the audio dimensions of the environment encoding (nature sounds and music) may have been perceived unevenly across residents, which could contribute to their weak or negative permutation importance.

LOPO cross-validation showed high variability (MAE range 0.00‐1.00), indicating that some individuals’ profiles are not well predicted from the patterns of the remaining 20 residents. The test period (weeks 8‐10) may also be subject to temporal confounding from habituation effects, seasonal changes, or progressive cognitive decline.

Finally, all report evaluation was automated (LLM-as-judge). While this approach enables systematic evaluation at scale, it has not been validated against the judgments of actual care staff or family members. Human evaluation is planned as future work.

A residual ethical question, distinct from the procedural safeguards described in the Ethical Considerations section, concerns what consent can meaningfully mean for continuous physiological monitoring in a population that cannot fully articulate its preferences. Institutional approval and per-session willingness verification address the procedural dimension of this question but do not resolve the deeper concern that decisions about participation are necessarily made on behalf of residents rather than by them. This caveat extends to any future deployment of the recommendation pipeline beyond a research context.

Two deployment-level risks warrant attention before any clinical translation of this pipeline. The first is care-staff overreliance on algorithmic recommendations. A system that outputs a ranked list of “calmest environments” for each resident has the potential to displace rather than support staff judgment, particularly in time-pressured care settings; any prospective deployment will require safeguards that preserve staff discretion, including explicit communication of model uncertainty and clear override protocols. The second concerns the report-generation component. The fidelity-compliance trade-off documented in the RQ3 results—where reports passed all binary safety checks while drifting toward clinical register for staff audiences and sacrificing numerical precision for family audiences—illustrates that LLM-generated text can satisfy structured safety filters yet still misrepresent a resident’s experience in ways that are not immediately obvious to a nontechnical reader. Routine human review of generated reports, at least during initial deployment periods, should be a precondition for any operational use.

The economic case for this approach turns on a simple substitution: if matched environment selection reduces the frequency of distress episodes that currently trigger pharmacological response or one-on-one care, the wearable infrastructure and pipeline costs could be offset by reductions in those downstream interventions. This study does not test that substitution. A formal cost-effectiveness analysis alongside any prospective validation trial is identified as necessary future work.

This feasibility and proof-of-concept study shows that a complete decision-support pipeline—from wearable physiological sensor to hybrid calm target to data-driven environment ranking to plain-language report—is technically implementable for VR comfort sessions in residential dementia care. An environment-level signal was detectable (permutation P=.045; P=.04 with 2000 permutations), person-specific, and interpretable, with water presence emerging as the most influential environment characteristic. However, the signal was small, borderline in statistical terms, and obtained from a single configuration whose selection was partly informed by the same test set; its significance did not survive alternative target definitions in sensitivity analyses, although its direction did. It should be regarded as preliminary evidence requiring independent replication, not as evidence of validated clinical usefulness. A compact model using 23 features outperformed larger models not in prediction accuracy but in its ability to use environment information—a finding that underscores the value of feature parsimony when the RQ concerns a small, specific signal.

The system is not intended to replace clinical judgment. Its purpose is to surface candidate patterns that are difficult to see unaided—for example, that one resident may relax more in environments with water while another may prefer social density—and to communicate them transparently so that staff can weigh them against their own observations.

Prospective validation is the necessary next step. A suitable study would enroll approximately 80 to 120 residents across 3 to 5 long-term care facilities to provide adequate power for multisite inference, with a stepped-wedge or cluster-randomized design (avoiding the practical complications of per-resident randomization in long-term care routines). The primary outcome should be an objective index of in-session distress—for example, PAS-2 vocal-agitation episode frequency per session, or a prespecified hybrid behavioral-physiological endpoint—with the comparator being usual care environment rotation or staff-selected scheduling. Ethical oversight should follow the substitute decision-maker framework applied here (see Limitations section), extended to a multisite consent protocol. Until such a study is completed, the present pipeline should be regarded as a demonstration of technical feasibility, not as a clinical tool ready for deployment.

Getting the data-to-report pipeline to work end-to-end, across 21 residents and 420 sessions, in a real long-term care setting, was the central contribution of this work; obtaining preliminary evidence that an environment signal may be recoverable from such data is a modest step toward dementia care systems that listen to the resident’s body, not just the observer’s notes—a step whose value depends on the independent replication we have identified as the necessary next test.