Study 1 was a single-arm pre-post study conducted from March to May 2024. Community-dwelling older adults aged 65 years or older were recruited through bulletin board postings at the community senior welfare center. Inclusion criteria were (1) the ability to communicate and follow instructions and (2) fear of falling. Exclusion criteria were (1) severe sensory impairments (hearing or vision), (2) cognitive or psychiatric impairments (eg, hallucinations), and (3) inability to ambulate independently.

Of 20 older adults screened, 18 met the eligibility criteria and agreed to participate. Three participants withdrew (2 before baseline because of sequelae of back surgery and 1 during the intervention because of scheduling conflicts), resulting in 15 participants (mean age 80.20 years, SD 6.88 years; 12/15, 80%, were women) who completed the study. The characteristics of the pilot study participants are reported in Table S1 in Multimedia Appendix 1.

To examine differential effects according to baseline cognitive function and sensorimotor risk levels, staff at the community senior welfare center assisted with recruitment to ensure balanced enrollment of participants across high- and low-risk categories.

Participants received the Brain Step program for 12 weeks, attending 50-minute sessions twice weekly. The program consisted of pressure sensor–based exercises, including center of gravity training, single- and double-foot stepping, and exergames (maze navigation, stepping stones, and mole catching). Sessions were conducted in small groups (≤8 participants per session because of space constraints at the Welfare Center) with trained facilitators, including 2 nonclinical welfare center staff members who served as session facilitators from this pilot phase onward, providing guidance and safety monitoring.

Primary outcomes included cognitive function (Korean Mini-Mental State Examination [K-MMSE]), physical function (Berg Balance Scale [BBS] and Timed Up and Go [TUG]), and sarcopenia-related measures (appendicular skeletal muscle mass [ASM], Appendicular Skeletal Muscle Mass Index [ASMI], and handgrip strength).

Secondary outcomes included physical fitness measures (30-second Sit-to-Stand, 2-Minute Step Test, and Figure-8 Walk) and device and program usability. Participants also completed qualitative interviews and satisfaction surveys. Pre-post changes were analyzed using generalized estimating equations (GEEs). Participant feedback was collected through focus group interviews with participants and senior welfare center staff.

Study 2 was a parallel-group RCT. Recruitment began in August 2024, the intervention was delivered from November 2024 to March 2025, and the final assessment was completed in April 2025. Participants were assessed at 3 time points: baseline (preintervention), 8 weeks (mid-intervention), and 16 weeks (postintervention).

Study 2 targeted low-income older adults living alone who were receiving the Customized Community Care Service described above. Participants were identified through the Community Care service database at the Welfare Center. Community Care providers facilitated recruitment by identifying eligible older adults from their caseloads, arranging initial contact with the research team, and coordinating transportation to the Welfare Center for intervention sessions. Through this involvement, Community Care providers also learned the intervention procedures and progressively assumed facilitator roles, further supporting the program’s sustainability within routine welfare center operations. The inclusion and exclusion criteria were identical to those of study 1.

Sample size was calculated using G*Power version 3.1.9.7. Based on a repeated-measures analysis of variance with a within-between interaction, a medium effect size (Cohen f=0.25) [27] was assumed. This assumption was based on meta-analytic evidence from exergaming interventions reporting pooled effect sizes of standardized mean difference of 0.35 for executive function [28] and standardized mean difference of 0.36-0.78 for balance outcomes in older adults [29], consistent with prior meta-analytic evidence [17]. Assuming an α level of .05, statistical power of 0.80, 2 groups, and 3 measurement time points, a minimum of 28 participants was required. Accounting for a 15% (5/33) dropout rate, the target sample size was 33 participants. To ensure balance between groups, 34 participants were planned for recruitment.

The random allocation sequence was generated using block randomization with a fixed block size of 4 by an independent researcher who was not involved in participant recruitment, intervention delivery, or outcome assessment. The allocation sequence was concealed from the other investigators until the time of assignment. Owing to the nature of the intervention, participants were not blinded and were aware of their group assignment, whereas outcome assessors were blinded to group allocation to minimize detection bias.

As shown in Figure 1, the Brain Step program consists of a large pressure sensor mat (1100 mm × 1100 mm) with 4 quadrants and 8 response zones, a display screen, and integrated safety bars.

Participants attended 50-minute sessions twice weekly for 16 weeks (32 sessions total). Sessions were conducted in small groups of up to 8 participants (limited by the available space and time at the Welfare Center), with morning and afternoon sessions available. Each session was led by 1 trained facilitator, with 1-2 research assistants providing individual guidance and safety monitoring. Instructional videos were provided to explain the program objectives and demonstrate each game task. Facilitators received training in device operation, safety monitoring, emergency procedures, and participant guidance before program initiation. To prepare for technology-assisted translation into routine community care practice, 2 welfare center staff members—who were not specialists in cognitive or physical interventions—were intentionally included as facilitator trainees, reflecting the typical workforce of community-based older adult care settings. Their operational needs were elicited through qualitative interviews and incorporated into the program design, including adjustable difficulty levels, automated game modules, and instructional videos demonstrating each task to support independent program delivery. Intervention fidelity was monitored through session checklists completed by research assistants and weekly supervision meetings. Adherence was tracked through individual session attendance records and game completion rates documented automatically by the Brain Step system.

As shown in Table 1, the main trial program was structured into 3 progressive phases. Phase 1 (weeks 1-4) focused on single-task training, including weight-shift exercises with real-time center of pressure feedback and side-reaching exercises (single- and double-foot stepping) to strengthen ankle mobility and lower limb muscle function. Phase 2 (weeks 5 and 6) introduced measurement system level 2, combining physical exercises with location memory tasks and learning system games to increase task complexity. Phase 3 (weeks 7-16) emphasized dual-task training through multidomain cognitive training targeting image association, arithmetic, and recognition. Participants selected answers by stepping on designated target zones among the 8 positions on the pressure sensor mat, thereby integrating cognitive decision-making with weight-shifting movements.

The software comprised 3 components (Figure 1): a balance assessment system (Figure 1B) providing real-time center-of-pressure visualization, a weight-shift training system (Figure 1C), and a cognitive-motor training system (Figure 1D) featuring 7 exergames inspired by classic arcade games from the 1980s and 1990s to enhance engagement through nostalgic familiarity. Detailed descriptions of each game are provided in the Methods S1 in Multimedia Appendix 1.

Participants in the control group received no intervention during the 16-week study period and attended only the 3 assessment sessions (baseline, 8 weeks, and 16 weeks). A passive (no-contact) control was selected at this feasibility stage to establish whether the intervention could be delivered and whether participants would engage with it in the real-world Public Community Care setting. As the intervention was delivered on-site within the existing infrastructure of the community welfare center, an active or attention-matched comparator was not logistically feasible at this stage. Limited space availability and a shortage of on-site staff precluded implementation of a parallel active-control program. Our priority was to evaluate the real-world applicability of the intervention.

Participants were stratified into high-risk (n=17) and low-risk (n=14) groups based on initial screening assessments. High-risk classification required the presence of 2 or more of the following: a history of recurrent falls, gait abnormalities, or cognitive test scores indicating mild cognitive impairment (K-MMSE ≤21 and Montreal Cognitive Assessment [MoCA] <23). This stratification was prespecified during the study design phase. Community stakeholders requested the inclusion of participants at the highest cognitive and sensorimotor risk, and the pilot study findings of differential response patterns informed the decision to classify participants by risk level using pretest screening (K-MMSE and fall history).

Comprehensive health outcomes were assessed to provide individualized feedback within the digital-human integrated personalized care framework. Outcomes were organized into 3 tiers: primary outcomes for which intervention effects were hypothesized, secondary outcomes to capture more detailed functional changes, and self-reported health outcomes routinely monitored in Community Care services. Primary outcomes included cognitive function (K-MMSE total and subscale scores), physical function (BBS and TUG), and sarcopenia indicators (ASM, ASMI, and grip strength). The K-MMSE [30] evaluates orientation, recall, attention/calculation, language, and visuospatial function (total score 0-30). Balance was measured using the Korean version of the BBS (14 items; total score 0-56) [31,32]. The TUG measured the time (in seconds) required to rise from a chair, walk 3 m, turn, return, and sit down [33]. Sarcopenia-related measures were assessed according to the Asian Working Group for Sarcopenia 2019 criteria [34]. ASM was measured using bioelectrical impedance analysis (InBody 370), and ASMI was calculated as ASM/height² (kg/m²). Grip strength was measured using a dynamometer (CAMRY EH101) [35,36], with the highest value from 3 trials of the dominant hand used for analysis.

Secondary outcomes were complementary, hypothesis-generating measures, including cognitive function (Korean Montreal Cognitive Assessment [MoCA-K] total and subscale scores), physical fitness (Figure-8 Walk Test, 30-second Sit-to-Stand, and 2-Minute Step Test [36,37]), and fall incidence. The MoCA-K [38,39] was included as a secondary cognitive measure because the K-MMSE has known ceiling effects in detecting subtle cognitive changes in individuals with mild cognitive impairment [40]. The MoCA-K provides greater sensitivity for mild cognitive impairment through more demanding assessments of executive function, visuospatial ability, and attention (total score 0-30) [39].

Fall incidence was assessed at 3 time points. At baseline, participants completed a screening question (Have you experienced a fall in the past year?). At the immediate postintervention assessment, they were asked, “Have you experienced a fall during the past 16 weeks?” At the 1-year postintervention follow-up, fall incidence was determined from community caregiver fall incident reports shared by the municipal welfare center as part of routine Public Community Care service monitoring. Fall incidence was included as a secondary outcome to complement the primary balance and mobility measures.

Other outcomes included sleep quality (Pittsburgh Sleep Quality Index [PSQI]; total score 0-21, with higher scores indicating poorer sleep quality) [41], depressive symptoms (Geriatric Depression Scale-Short Form [GDS-SF]) [42], and pain-related functional limitations (Western Ontario and McMaster Universities Osteoarthritis Index [WOMAC]) [43].

Game performance data were automatically recorded throughout the 16-week intervention for all participants in the intervention group (n=16). A total of 63 performance indicators were generated across 9 games, including play count, average score, play time, reaction time, and game-specific metrics.

Baseline characteristics were compared between groups to assess randomization balance using t tests (unpaired and 2-tailed) for continuous variables and chi-square or Fisher exact tests for categorical variables. The primary analysis examined intervention effects on primary, secondary, and other health outcomes using GEEs to account for the correlation among repeated measures within participants. GEE models specified an exchangeable working correlation structure with robust SEs. All randomized participants were analyzed according to their assigned groups (intention-to-treat). Missing observations were handled at the observation level using available-case analysis without imputation, with all 31 randomized participants retained in their originally assigned groups and each contributing all available time points to the estimation [44-46]. For grip strength, 1 participant could not be assessed because of hand surgery unrelated to the intervention and was therefore excluded from the grip strength model (n=30). All other outcomes included all 31 randomized participants (see the “Data Exclusion and Missing Data” section).

The full model included group (intervention vs control), time (baseline, 8 weeks, and 16 weeks), and the group × time interaction to examine whether intervention effects were significant. Risk-stratified analyses were also conducted, as prespecified during the study design phase. Years of education were included as a covariate in the models for cognitive function to account for potential confounding. Given the number of outcome variables tested, the Benjamini-Hochberg false discovery rate (FDR) correction was applied within the 6 primary outcomes (K-MMSE total, BBS, TUG, ASM, ASMI, and grip strength; 6 comparisons) at q=0.05. Secondary outcomes (MoCA-K total, Figure-8 Walk, 30-second Sit-to-Stand, 2-Minute Step Test, and fall incidence) and other outcomes (PSQI, GDS-SF, and WOMAC) were reported with unadjusted P values only. Subscale scores were reported descriptively without FDR correction. These analyses were distinguished by both outcome tier and inferential status. Confirmatory inference was restricted to the 6 prespecified primary outcomes with FDR correction. The secondary and other outcomes, the K-MMSE subscale analyses, and the prespecified risk-stratified subgroup analyses were exploratory and hypothesis-generating and were reported with unadjusted P values without confirmatory interpretation. The only post hoc, data-driven analyses were the exploratory correlations between game performance metrics and clinical outcomes (Figure S1 in Multimedia Appendix 1).

An exploratory analysis examined the associations between game performance metrics and significant changes in outcome measures in study 2, with P<.05 applied as the exploratory threshold for identifying significant outcome changes in the intervention group. Pearson correlation coefficients were calculated between game performance metrics and nominally significant changes in primary outcomes (pre-to-post and mid-to-post; for ASM, pre-to-post change, P=.04). For K-MMSE Recall (P=.01) and ASMI (P=.05), analyses were conducted in the high-risk subgroup (n=8) based on significant group × risk × time interactions. For ASM, analyses were conducted in the full intervention group (n=14).

Statistical significance was set at P<.05 (2-tailed). The FDR correction described above was applied to account for multiple outcome comparisons. The study was not powered to detect intervention effects for all outcomes or subgroup effects. All risk-stratified and correlation analyses were exploratory and hypothesis-generating. Effect sizes were calculated as Cohen d based on the between-group differences in pre-to-post change scores divided by the pooled SD of the change scores, with 95% CIs [27]. Analyses were performed using Python 3.12 (Python Foundation) with the statsmodels package for GEE models and SciPy for correlation analyses. Additional statistical details, including the analysis of fall incidence outcomes, are provided in Methods S1 in Multimedia Appendix 1.

Of the 31 randomized participants, all completed the baseline assessment except for 1 intervention participant who declined the grip strength measurement because of a history of bilateral hand surgery. This participant subsequently withdrew before the 8-week assessment. By the 8-week midpoint, 5 participants had withdrawn (2 intervention and 3 control), leaving 26 participants with data at that time point. By the 16-week postintervention assessment, 1 additional control participant had withdrawn because of scheduling conflicts, leaving 25 participants with data at that time point. The full longitudinal data set thus comprised 82 participant-by-time point observations across the 3 assessments.

Missing data resulted primarily from participant attrition rather than item-level nonresponse and were not imputed. All GEE analyses used available-case analysis at the observation level. Each participant-by-time point record was included when the outcome value was nonmissing, whereas records missing key covariates (group assignment or baseline risk stratum) were excluded. This approach leverages GEE’s ability to accommodate unbalanced repeated-measures data without listwise deletion at the participant level [45,46], such that all 31 randomized participants were retained in their originally assigned groups, with each contributing all available observations to the estimation.

Participant flow is presented in Figure 3. Of the 40 older adults screened, 31 were randomized to the intervention (n=16) or control (n=15) group. Six participants withdrew during the study. All 31 randomized participants were included in the intention-to-treat analysis (82 total observations across 3 time points). See Multimedia Appendix 2 for the study’s CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) checklist.

Baseline characteristics are presented in Table 3. Participants (mean age 81.5 years; 26/31, 84%, female) were low-income older adults living alone who were receiving Public Community Care services. The groups were comparable at baseline (see Table 3 for P values).

As shown in Table 4, there were no significant group × time interactions for cognitive function. The K-MMSE total score showed no significant interaction (P=.27, FDR-adjusted P=.38). No significant interactions were observed for the K-MMSE subscales (Orientation, P=.36; Recall, P=.09; Attention, P=.37; Language, P=.27; and Visuospatial, P=.12).

A nominally significant group × time interaction was observed for the sarcopenia-related index ASM (P=.04; Cohen d=0.39, 95% CI −0.41 to 1.18). The intervention group maintained ASM over the 16-week period (baseline: 15.35 kg; postintervention: 15.77 kg), whereas the control group showed a decline (baseline: 17.03 kg; postintervention: 15.48 kg; Table 4). The group × time interaction for ASMI did not reach statistical significance (P=.05; d=0.38, 95% CI −0.41 to 1.18). This pattern is consistent with possible preservation of muscle mass, although the finding did not survive FDR correction. The BBS showed a marginally significant group × time interaction (P=.05; d=−0.27), likely reflecting a ceiling effect given the high baseline scores in both groups (intervention: 52.88; control: 52.73 of a maximum score of 56). No significant group × time interactions were observed for the other sensorimotor frailty measures (Timed Up and Go, P=.65; for grip strength, P=.32).

After FDR correction, none of the group × time interactions remained statistically significant (all FDR-adjusted P>.16; see Table 4). However, effect sizes are presented to provide context for the observed differences. Before correction, the nominally significant findings showed small-to-medium effect sizes (ASM d=0.39; ASMI d=0.38), and the risk-stratified analyses revealed medium-to-large effect sizes among high-risk participants for K-MMSE Recall (d=0.73) and ASMI (d=0.90), warranting confirmation in larger trials with risk-enriched samples.

As a secondary outcome, self-reported fall incidence was examined at 3 time points (Table S5 in Multimedia Appendix 1). The group × time interaction for pre-post fall incidence was not statistically significant (Wald χ²₁=0.66, P=.42; odds ratio [OR] 0.47, 95% CI 0.07-2.94). At the postintervention assessment, fall incidence was numerically lower in the intervention group than in the control group (5/14, 35.7% vs 7/11, 63.6%; OR 0.32, 95% CI 0.06-1.64). Although this between-group difference did not reach statistical significance (Fisher exact P=.24), the direction and magnitude of the effect favored the intervention group. As a secondary outcome, this finding should be interpreted as hypothesis-generating only.

A 1-year postintervention follow-up assessed fall incidence through community caregiver reports collected as part of routine Public Community Care service monitoring. No fall events were reported for any participant in either the intervention or control group during the 12-month follow-up period (intervention, 0/14; control, 0/14). A separate GEE model comparing baseline with the 12-month follow-up also showed no significant group × time interaction (Wald χ²₁=0.13, P=.71; OR 1.39, 95% CI 0.24-8.05), although this model should be interpreted with caution because of quasi-complete separation resulting from the absence of events at follow-up. Although the absence of falls during follow-up is notable for this population with elevated fall risk, the floor effect precluded between-group comparisons, and the absence of a nonparticipating comparison group limited causal interpretation.

No significant group × time interactions were observed for the secondary outcomes (MoCA-K, P=.12; Figure-8 Walk, P=.65; 30-second Sit-to-Stand, P=.72; and 2-Minute Step Test, P=.62; Tables S3 and S4 in Multimedia Appendix 1) or the self-reported health outcomes (PSQI, P=.19; GDS-SF, P=.55; and WOMAC, P=.68).

Figure 4 presents outcomes with risk-level interactions. These analyses are exploratory and hypothesis-generating because of the small subgroup cell sizes (n=5-8 per cell). High risk was defined as meeting the criteria for cognitive impairment (K-MMSE ≤21 or Montreal Cognitive Assessment [MoCA-K] <23) combined with a history of recurrent falls.

A significant uncorrected 3-way interaction (group × risk × time) was observed for K-MMSE Recall (Wald χ²₂=6.16, P=.01; Figure 4A). Among the high-risk participants, the intervention group showed an improvement in Recall (4.70-5.00; d=0.73), whereas the control group declined (5.14-4.40). The low-risk participants showed the opposite pattern (d=−0.99), suggesting that cognitive effects may be concentrated among individuals with greater baseline impairment.

The 3-way interaction for ASMI reached borderline significance (Wald χ²₂=6.00, P=.05; Figure 4B). Among the high-risk participants, the between-group effect size was large (d=0.90): the control group showed a substantial decline (8.0-6.3 kg/m²), whereas the intervention group maintained ASMI levels (6.2-6.5 kg/m²). No other outcome variables showed significant uncorrected 3-way interactions (group × risk × time), including the MoCA-K total (P=.71) and subscale scores (Memory, P=.83; Attention, P=.71; Language, P=.71; Visuospatial, P=.75; Orientation, P=.89), K-MMSE total (P=.71) and other subscale scores (Recall, P=.37; Orientation, P=.85; Attention, P=.75; Language, P=.71; Visuospatial, P=.75), and sensorimotor function measures (BBS, P=.82; TUG, P=.83; Figure-8 Walk, P=.75; grip strength, P=.81; 30-second Sit-to-Stand, P=.70; 2-Minute Step Test, P=.70; and ASM, P=.70).

Post hoc exploratory correlations between game performance metrics and clinical outcomes (Figure S1 in Multimedia Appendix 1) suggested potential associations between cognitive challenge during gameplay and cognitive outcomes, although the small subgroup sizes require cautious interpretation.