Qualitative coding was conducted by the first author (XS) and one additional researcher (HQ). Both researchers worked from the same phase-based codebook and excerpt-level matrix, and the first author checked the Chinese-to-English translations used for reporting. No independent third adjudicator was used; disagreements were resolved through discussion and consensus. Reflexive management included coding memos, comparison of coding decisions, codebook revisions, and retention of original Chinese excerpts alongside translations and interpretations.

Some participants were previously known to members of the research team through community connections, whereas others had no prior relationship with the researchers. Prior familiarity was not used to determine eligibility, group allocation, or coding. Because familiarity could influence willingness to participate or what participants disclosed, the team used the same interview guide, study-code system, phase-based codebook, and consensus procedures across participants and did not treat familiarity as evidence for a preferred interpretation. Sample handling is summarized in Table 1.

The completed-protocol analysis used interview evidence, questionnaire responses, safety observations, and session-level performance exports from 34 community-dwelling older adults (VR, n=16; MR, n=18). One additional MR partial-exposure transcript was retained only as a diagnostic case. Participant characteristics relevant to interpretation, including age, sex, balance, exercise, and technology-use indicators, are reported in Table 2.

Participants were community-dwelling older adults recruited through local community centers and community-based outreach in Hunan, China. Eligibility criteria included age 60 years or older, ability to stand independently, ability to follow guided Tai Chi movements, and willingness to use a head-mounted display under supervision. Exclusion criteria included conditions that could increase risk or compromise safe participation, such as severe vertigo or motion-sickness susceptibility, history of seizure, severe uncorrected visual or auditory impairment, uncontrolled cardiovascular disease, severe musculoskeletal limitations that prevented safe standing movement practice, inability to provide informed consent, or other contraindications identified during screening.

The broader research program initially contacted 86 community-dwelling older adults through community-center promotion, on-site outreach, local newspaper advertising, and community contact networks in Hunan, China. After screening, 70 participants completed the initial single-session experiment; 40 of these participants voluntarily entered the repeated-exposure follow-up from which the current interview and session evidence was drawn. The archived follow-up materials contained 35 interview transcripts. For the current analysis, the completed-protocol quantitative and session sample and main qualitative coded sample each included 34 participants (VR, n=16; MR, n=18). Participant 37 was retained only as an MR partial-exposure diagnostic case and excluded from main theme counts. Participant 6 was a completed-protocol VR participant whose author-confirmed paraphrased coding contributed to participant-level theme counts but was not eligible for use as a verbatim quotation.

The sample size was determined by feasibility, participant availability, and the qualitative-dominant purpose of the study. The study was not powered to test VR and MR superiority or causal mechanisms. Quantitative estimates are presented with 95% CIs to describe precision and to support triangulation with qualitative findings. Small denominators, baseline imbalance, and retained allocation are treated as constraints on interpretation rather than as defects that can be overcome through statistical modeling.

Both conditions used the same Tai Chi exergame content delivered through a Meta Quest 3 head-mounted display. The VR condition presented a fully virtual environment that occluded the surrounding physical environment. The MR condition used full-color video passthrough with virtual overlays, allowing participants to retain visual access to the real environment. The same application build, task goals, and core settings were used across conditions; delivery mode was the intended difference.

The exergame implemented a simplified 16-form Tai Chi sequence guided by a virtual coach. For each movement, participants aligned the head-mounted display and the left and right controllers with 3 predefined target hotspots. Holding any hotspot within the success window for at least 3 seconds counted as one successful hotspot hold, and 48 successful holds represented one complete 16-movement loop. The sequence restarted when a loop was completed before the session ended. Participants received immediate visual guidance and end-of-session performance summaries. The repeated-exposure protocol comprised 6 sessions delivered across 6 consecutive days, with one session per day; days 1, 3, and 6 were the standardized performance checkpoints used in this analysis.

Masking was not applicable. Participants and research staff could see whether VR or MR was delivered. The follow-up retained the parent-study allocation, and all comparisons were descriptive.

Baseline variables included age, sex, education, daily smartphone use duration, Berg Balance Scale (BBS) [36], self-reported exercise intensity, self-reported health status, height, weight, and body mass index. Subjective experience was summarized using a Chinese-adapted 32-item version of the Game Experience Questionnaire (GEQ) core module [37]. Responses were coded on the 0‐4 analytic scale. The standard story-interest item was not administered; therefore, sensory and imaginative immersion was calculated from 5 items rather than the standard 6. The remaining analyzed dimensions were competence, flow, tension and annoyance, challenge, negative affect, and positive affect. Cybersickness was assessed with the 9-item Virtual Reality Sickness Questionnaire (VRSQ) [38]. The oculomotor score was calculated as the sum of items 1, 2, 4, and 5 divided by 12 and multiplied by 100; the disorientation score was calculated as the sum of items 3 and 6‐9 divided by 15 and multiplied by 100. The standardized VRSQ total was the mean of these two subscale scores. Automated game logs provided average accuracy rate and average time spent at standardized checkpoints on days 1, 3, and 6. Average accuracy rate was the proportion of logged frames in which the head and hand hotspots were within the game’s tolerance window. Average time spent was recorded in seconds, with lower values indicating faster task performance [39]. Trajectory was retained only as a diagnostic field because its direction and operational definition could not be verified and was therefore excluded from manuscript-ready analyses.

Measurement quality was evaluated using item-level completeness, observed score ranges, floor and ceiling concentrations, and consistency between questionnaire scoring rules and the available item data. Performance-field definitions and units were checked against the available game documentation and prior study records.

Within-sample Cronbach alpha was reported for analyzed GEQ dimensions and the VRSQ where appropriate. Test-retest reliability was not estimated because repeated administrations represented changing session experiences rather than a stability assessment.

Semistructured interviews were conducted by the research team across the repeated-exposure follow-up using selected postsession interview modules rather than a single endpoint interview. Interviews were audio recorded and generally lasted approximately 20 minutes. The interview guide prompted participants to describe initial impressions, unfamiliarity, bodily sensations, task difficulty, environmental experience, perceived safety, social reinforcement, perceived learning, trust, and willingness to continue use. Modules were aligned with early contact, adaptation after repeated practice, and later-session evaluation, so later prompts could ask participants to compare current experiences with earlier sessions. This design supported comparability across participants while allowing personally salient barriers and facilitators to be described [40].

Interview material was segmented into analyzable excerpts linked to participant study codes, group membership, DBF phase, theme label, and valence category. The excerpt-level approach was used because a single participant could describe several barriers and facilitators within the same interview. Segmenting these statements allowed the analysis to retain complexity while summarizing unique-participant theme endorsement at the phase level.

Initial Chinese-language transcripts were generated from the audio recordings using a Qwen-based automated transcription tool. Participant names and direct identifiers were replaced with study codes in the analytic matrix. The original Chinese transcripts were retained as the primary textual evidence, and Chinese-to-English translations used for analysis and reporting were checked by the first author. Quotations retained in the paper were checked against the corresponding Chinese transcript and shortened, where necessary, by removing repeated fillers or unclear fragments rather than silently reconstructing them. Author-confirmed summaries from participant 6 were used for coding and participant-level counts but not as direct quotations.

Qualitative analysis followed thematic analysis and framework-oriented procedures [41-43]. A DBF-informed codebook combined phase tags (entry, adaptation, and conditional adoption) with theme labels for barriers and facilitators. Initial codes were developed from the interview guide, DBF phases, and repeated reading of excerpts. Codes were then refined to distinguish conceptually adjacent themes such as Initial Uncertainty, Physical Discomfort, Skill Adaptation, Environmental Immersion, Trust Development, and Conditional Adoption.

To improve coding transparency, the analytic matrix retained the original excerpt, English translation, DBF phase, theme label, valence, and a short coding memo. For example, comments about headset heaviness, eye discomfort, or dizziness were coded as Physical Discomfort or Motion Discomfort, depending on the specific symptom. Comments about discussing strategies, comparing scores, or sharing ways to improve were coded as Social Reinforcement. Comments that expressed willingness only if equipment comfort, floor safety, content variety, or session duration improved were coded as Conditional Adoption or Adoption With Safety Measures. A codebook excerpt and representative coding examples are provided in Multimedia Appendix 1.

Two researchers (XS and HQ) participated in coding, compared coding decisions, and resolved disagreements through discussion and consensus. Coding memos and codebook revisions were retained to support an audit trail. Thematic saturation was not treated as a statistical threshold; operationally, thematic sufficiency was judged to have been reached because later transcripts primarily elaborated existing phase-level categories, such as uncertainty, physical discomfort, immersion, social reinforcement, trust, and conditional adoption, rather than generating new dominant phase-level themes [44,45]. Theme counts represent unique participants endorsing a theme within the completed-protocol qualitative sample and are interpreted as descriptive evidence coverage rather than prevalence estimates.

Methodological integrity was supported through 4 linked procedures. Adequacy was addressed by using repeated-phase prompts across the completed-protocol qualitative sample and by retaining discrepant as well as convergent accounts. Researcher perspectives were managed through the shared interview guide, explicit phase and valence fields, coding memos, comparison of coding decisions, and consensus discussion. Groundedness was supported by retaining original Chinese excerpts, linking every coded statement to a participant and session phase, and presenting representative quotations. Meaningfulness and coherence were examined through integration with questionnaire, safety, and performance evidence without forcing convergence when data sources diverged. No participant member checking was conducted, and this absence is treated as a limitation rather than implied as a completed procedure.

Integration occurred at three levels. First, qualitative themes were quantified as unique-participant endorsements within each DBF phase. Second, theme endorsement was linked descriptively to GEQ, VRSQ, and performance summaries. Third, discrepant and illustrative cases were examined to clarify how adoption could proceed despite early uncertainty, mild symptoms, or conditional reservations. This integration strategy was intended to connect experience, performance, and interpretation without converting the study into a confirmatory quantitative comparison [46,47].

The integration priority was explanation rather than convergence alone. When interview themes, questionnaire profiles, and performance logs pointed in the same direction, they were used to strengthen interpretation. When they diverged, the qualitative account was used to clarify why a participant could improve technically while remaining cautious about comfort, safety, or future use.

Baseline characteristics were summarized using means and SDs for continuous variables and counts and percentages for categorical variables. Standardized mean differences and absolute proportion differences were used to describe baseline imbalance. Item-level completeness, Cronbach alpha, observed score ranges, and floor and ceiling concentrations were examined for each GEQ dimension and the VRSQ. Performance was summarized at days 1, 3, and 6 for average accuracy rate and average time spent (seconds), and day 1 to day 6 change was presented with 95% CIs. Trajectory was excluded from manuscript-ready performance results because its definition and direction could not be verified. GEQ and VRSQ profiles were interpreted descriptively. No P values or confirmatory between-group superiority tests were reported because the study was exploratory and quasi-experimental.

Data diagnostics included item completeness, observed ranges, floor and ceiling concentrations, session availability, and verification of performance-field definitions. No outlier exclusions, transformations, or imputations were used. Trajectory was excluded because its direction and operational definition could not be verified.

The completed-protocol quantitative dataset and main qualitative coded dataset each included 34 participants (VR, n=16; MR, n=18). The partial-exposure MR transcript was retained only for diagnostic reconciliation and excluded from main theme counts. One completed-protocol VR transcript was checked and added to the coding matrix using verified coding rows. Given the small sample size and the nonconfirmatory design, analyses used complete-case summaries by data source, and missingness was reported descriptively rather than modeled with imputation. A formal missing completely at random test and multiple imputation were not conducted because the study did not make confirmatory estimates.

All sessions were supervised by the research team. Participants received orientation before headset use, assistance with equipment fitting, and reminders that they could pause, rest, or stop at any point. A researcher remained nearby to monitor movement safety, dizziness, fatigue, discomfort, headset pressure, anxiety, and potential fall risk. Predefined stopping criteria included participant request to stop, visible instability, worsening dizziness or nausea, marked fatigue, anxiety, headset discomfort that did not resolve after adjustment, or any researcher concern that continued movement could increase fall risk. Monitoring procedures included presession orientation, in-session observation, symptom checks during pauses or after reported discomfort, immediate rest or headset removal when needed, and postsession confirmation that symptoms had resolved or remained mild. The training area was arranged to reduce obstacles and allow researcher assistance if needed. No falls, serious adverse events, or session terminations were recorded.

Entry-stage responses were characterized by unfamiliarity, uncertainty, and bodily adjustment. Initial Uncertainty was endorsed by 7 of 16 participants (43.8%; 95% CI 23.1%‐66.8%) in VR and 6 of 18 participants (33.3%; 95% CI 16.3%‐56.3%) in MR. Physical Discomfort was endorsed by 4 of 16 participants (25.0%; 95% CI 10.2%‐49.5%) in VR and 3 of 18 participants (16.7%; 95% CI 5.8%‐39.2%) in MR. Initial Adaptation was endorsed by 1 of 16 participants (6.2%; 95% CI 1.1%‐28.3%) in VR and 5 of 18 participants (27.8%; 95% CI 12.5%‐50.9%) in MR. These findings indicate that early resistance was not reducible to physiological discomfort alone. Participants also described unfamiliar equipment, uncertainty about whether they could operate the system, and concern about performing movements correctly.

A VR participant stated, “I was a little worried.” An MR participant described the first five minutes as “It felt very difficult.” These transcript-grounded excerpts illustrate competence and task uncertainty rather than physical intolerance alone. Other participants described novelty and curiosity, suggesting that unfamiliarity could operate as both a barrier and a facilitator depending on whether orientation and reassurance were sufficient.

This entry pattern has practical importance because it suggests that support should begin before performance feedback. Older adults who worried about operating the system or performing incorrectly needed clear orientation, slow-paced familiarization, and visible researcher support. Early adoption was also a response to whether participants felt that errors, pauses, and bodily sensations could be managed safely.

During adaptation, participants increasingly described the training in relation to perceived improvement, environmental experience, and encouragement from others. Environmental Immersion was endorsed by 9 of 16 participants (56.2%; 95% CI 33.2%‐76.9%) in VR and 8 of 18 participants (44.4%; 95% CI 24.6%‐66.3%) in MR. Social Reinforcement was endorsed by 7 of 16 participants (43.8%; 95% CI 23.1%‐66.8%) in VR and 9 of 18 participants (50.0%; 95% CI 29.0%‐71.0%) in MR. Skill Adaptation was endorsed by 6 of 16 participants (37.5%; 95% CI 18.5%‐61.4%) in VR and 9 of 18 participants (50.0%; 95% CI 29.0%‐71.0%) in MR. Emotional Engagement was endorsed by 1 of 16 participants (6.2%; 95% CI 1.1%‐28.3%) in VR and 6 of 18 participants (33.3%; 95% CI 16.3%‐56.3%) in MR.

VR accounts often connected engagement to scenery, music, and emotional comfort. One VR participant explained, “I preferred the openness of nature, the boundless mountains...it felt as though I was immersed in nature.” MR accounts more often connected adaptation to remaining oriented to the real environment while learning the movement sequence; one participant stated, “I think I could get better the more I played.” Social reinforcement appeared in both groups. A VR participant described strategy sharing directly: “I wanted to share my method with them so that they could also achieve higher scores.” These accounts show that peer comparison and shared problem solving operated alongside immersive scenery and system feedback.

GEQ profiles supported this interpretation descriptively: VR had higher sensory and imaginative immersion (3.69; SD 0.47) and flow (3.83; SD 0.41) than MR (2.64; SD 0.58 and 2.57; SD 0.70), respectively. These questionnaire summaries are used only to contextualize qualitative interpretation, not to test modality superiority.

Late-stage adoption was conditional rather than unconditional. Conditional willingness with improvements was endorsed by 6 of 16 participants (37.5%; 95% CI 18.5%‐61.4%) in VR and 1 of 18 participants (5.6%; 95% CI 1.0%‐25.8%) in MR. Conditional willingness was endorsed by 2 of 16 participants (12.5%; 95% CI 3.5%‐36.0%) in VR and 4 of 18 participants (22.2%; 95% CI 9.0%‐45.2%) in MR. Trust Development was endorsed by 0 in VR and 5 of 18 participants (27.8%; 95% CI 12.5%‐50.9%) in MR, while Trust and conditional willingness was endorsed by 5 of 16 participants (31.2%; 95% CI 14.2%‐55.6%) in VR and 3 of 18 participants (16.7%; 95% CI 5.8%‐39.2%) in MR. These patterns should be interpreted cautiously because baseline and follow-up conditions differed descriptively, but they suggest different experiential emphases.

Some participants expressed willingness to continue immersive Tai Chi if safety, headset comfort, session duration, content variety, and the physical setting were improved. An MR participant answered “I trusted it” and later described the equipment as “it felt reliable.” A VR participant described a possible place in the daily routine: “I could dance for exercise in the morning and play this in the afternoon; that would be quite good.” Another linked willingness to a practical environmental safeguard: “Just put down a thin carpet.then I would not be worried.” These accounts support conditional adoption rather than evidence of long-term unsupervised adherence.

Conditional adoption was therefore not a weak form of acceptance; it was the dominant form of late-stage evaluation in this supervised setting. Participants did not simply accept or reject the system. They specified conditions under which continued use would be reasonable, such as shorter sessions, more comfortable headset fitting, safer space arrangement, clearer guidance, or richer content (Table 5).

Triangulation across interviews, questionnaires, and performance logs supported a phase-sensitive interpretation. Accuracy improved in both groups, but participants interpreted improvement through different experiential frames. VR accounts often connected engagement to environmental immersion and affective value, whereas MR accounts more often emphasized reality grounding, perceived control, and trust. The same performance pattern could be interpreted differently depending on comfort, flow, sickness symptoms, and perceived control. Figure 2 synthesizes how entry uncertainty and physical discomfort gave way to practice-supported adaptation and then to trust or conditional willingness; the arrows organize the interpretation and do not represent tested causal effects. The map is a qualitative-dominant visual synthesis.

Safety-related interview content included dizziness, fatigue, eye discomfort, headset pressure, and concern about falling. Strict safety and discomfort themes were present in 7/16 (43.8%) of VR participants and 11/18 (61.1%) of MR participants. No falls, serious adverse events, or session terminations were recorded. These mild symptoms nevertheless shaped conditional adoption because participants evaluated continued use in relation to whether the system felt safe, manageable, and worthwhile (Table 6).

The VR and MR comparison should be interpreted cautiously. Recent reviews support the potential of immersive training to improve balance and physical activity outcomes in older adults [9,10], and a preliminary randomized trial reported promising findings for MR-based exercise in older women with sarcopenia [29]. Those studies establish clinical and implementation plausibility, but they do not determine whether VR or MR is superior for the present Tai Chi application. Our qualitative evidence instead identifies different experiential emphases: VR accounts foregrounded scenery, absorption, and affective value, whereas MR accounts more often foregrounded visual grounding, control, and trust. These are design hypotheses rather than causal modality effects. Alternative explanations remain plausible because baseline imbalance, prior technology experience, novelty, supervised delivery, community familiarity, and individual preference can shape immersive-system responses [11-16,50]. Preserving access to the physical room may support orientation for some users but reduce absorption, whereas full visual immersion may enhance engagement while increasing uncertainty for others [26-28,50]. Future comparisons should therefore manipulate occlusion, feedback, and supervision directly rather than treating the present group pattern as a stable property of either modality.

Social reinforcement deserves separate emphasis. Participants described discussing strategies, comparing scores, sharing methods, and wanting multiplayer or encouragement features. This is consistent with recent group and remote VR studies in which trusted facilitators, peer connection, and group delivery shaped acceptance and engagement [25,30], and with qualitative evidence that older users wanted stronger social engagement in VR exercise games [31]. For community programs, social support is therefore not an accessory to the headset. It is part of the intervention mechanism and may remain motivational after technological novelty declines.

Safety concerns were not peripheral; they were part of the adoption process. Average VRSQ scores were low, and no falls, serious adverse events, or session terminations were recorded, but interviews still documented dizziness, fatigue, eye discomfort, headset pressure, and concern about falling. This combination is consistent with reviews describing generally feasible immersive exercise alongside persistent usability and tolerability concerns [9,10,50]. Symptom averages can therefore understate the importance of episodic discomfort for an individual’s decision to continue. Short initial sessions, stable layouts, headset-fit checks, visible stopping procedures, symptom monitoring, and instructor support serve both risk control and trust formation [9,10,50]. The absence of serious adverse events in supervised sessions is not evidence of safety for unsupervised community or home use because supervision, equipment setup, and immediate assistance were part of the conditions under which participants evaluated the system. Recent adherence research likewise notes that XR exercise evidence is dominated by short, supported programs and rarely includes long-term community follow-up [32]. Any transfer to home use should therefore be preceded by explicit risk assessment, an orientation protocol, clear stopping rules, and a plan for technical or caregiver support [32,51-53].

For community and home implementation, immersive Tai Chi should be delivered as a phase-sensitive service rather than a stand-alone product. Entry support should reduce procedural, performance, bodily, and environmental uncertainty through orientation, simple controls, fit checks, visible safety precautions, and permission to pause. Adaptation support should combine understandable performance feedback, progressive challenge, peer comparison, and encouragement. Conditional adoption requires lighter and easier-to-fit equipment, meaningful content variety, transparent safety procedures, and realistic access to instructor, caregiver, or technical support. These priorities align with calls for age-appropriate challenge, social engagement, and service continuity in XR exercise [31,32] and with broader evidence that older adults’ health-technology engagement depends on usability, perceived value, support, and fit with daily life [51-53]. For VR delivery, the priority is to preserve the motivational value of immersive scenery while reducing disorientation, headset burden, and uncertainty about the surrounding space [9,10,50]. For MR delivery, environmental visibility should be paired with clear virtual guidance, trustworthy feedback, and sufficient content richness; recent MR exercise evidence also supports supervised, adaptive task progression rather than unassisted exposure [29]. For both modalities, community programs should plan staff training, safe layouts, symptom checks, and gradual progression before considering independent home use [32,51-53].

Several limitations should be considered. First, the follow-up retained the parent study allocation and was not rerandomized; the VR and MR comparison is exploratory. Second, the small community sample from Hunan, China limits transferability. Third, one MR partial-exposure case was excluded from main theme counts, and participant 6 contributed author-confirmed paraphrased coding rather than direct quotations. Fourth, baseline differences in sex distribution, smartphone use, exercise intensity, and BBS may have influenced engagement and trust. Fifth, some participants were previously known through community connections, which may have affected participation or disclosure despite standardized procedures. Sixth, automated transcription may contain recognition errors; quotations were therefore shortened to transcript-grounded excerpts and should remain subject to final source-audio checking before submission. Seventh, six supervised sessions do not demonstrate long-term adherence or unsupervised safety. Eighth, the system-defined average accuracy rate and average time spent quantify game-task performance rather than validated biomechanical technique; the backend’s numeric spatial-tolerance threshold was not preserved in the available export, and trajectory direction and definition were unverifiable. Performance findings were therefore restricted to directly reproducible descriptive fields, and trajectory was excluded. Finally, theme counts indicate evidence coverage within this qualitative sample, not population prevalence.

Future research should prospectively register larger and more balanced trials, predefine adverse-event and stopping procedures, and preserve backend metric specifications, including spatial-tolerance thresholds, before data collection. Repeated-measures designs should distinguish short-term task familiarization from retained motor learning and should test whether phase-tailored orientation, feedback, social support, and symptom management improve adherence. VR and MR comparisons should hold exercise content constant while manipulating environmental visibility and feedback, and they should include longer community or home follow-up. This agenda is consistent with current evidence calling for multidimensional adherence measurement, long-term tracking, and direct evaluation of age-sensitive XR service design [9,29,32].