This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of Kwangwoon University, Republic of Korea, under IRB number 7001546-202300614-HR(SB)-005-01. Written informed consent was obtained from all participants before enrollment.

Participant confidentiality was strictly protected throughout the study. All EEG recordings and demographic information were anonymized and deidentified before analysis. No personally identifiable information was included in the study database, and no identifiable participant images are presented in this manuscript.

Participants were recruited on a voluntary basis and did not receive financial compensation for participation.

This was an exploratory, nonrandomized, single-session, 3-group repeated-measures EEG study. Participants with AD, MCI, and NC underwent repeated EEG measurements during pre-VR resting, immersive VR exposure, and post-VR resting states within 1 experimental session. Diagnostic group constituted the between-participant factor, and experimental state constituted the within-participant repeated-measures factor. All participants underwent the same VR procedure. No randomization, treatment allocation, separate non-VR comparison condition, or longitudinal follow-up was included. The study was designed to characterize short-term EEG functional network responses rather than to evaluate therapeutic efficacy.

Participants were recruited through advertisements at Chuncheon Sacred Heart Hospital within 1 month before the experimental sessions. Patients with AD and individuals with MCI were recruited from hospital patients whose diagnoses and clinical status were confirmed by experienced clinicians. Older adults with NC were recruited from the same hospital-based recruitment setting and were screened for the absence of neurological or psychiatric disorders.

Inclusion criteria were being aged 60 to 70 years, completion of at least secondary-school education, ability to complete EEG recording, and ability to tolerate VR stimulation. For the AD and MCI groups, participants were required to have a clinical diagnosis established by experienced clinicians. For the NC group, participants were required to have no history of neurological or psychiatric disorders.

Exclusion criteria included severe visual impairment preventing VR participation, inability to complete EEG recordings, severe psychiatric disorders, major neurological diseases other than AD or MCI, and intolerance to VR stimulation. Participants were instructed to avoid medications or substances that could acutely affect EEG activity for at least 24 hours before the experiment whenever clinically appropriate.

A total of 60 older adults participated in this study, including 20 patients diagnosed with AD, 20 individuals with MCI, and 20 participants with NC. All participants were aged 60 to 70 years and had completed at least a secondary-school level of education. Sex distribution was balanced across groups, with 10 male and 10 female participants in each group.

Diagnoses of AD and MCI were established according to standard clinical diagnostic criteria by experienced clinicians. Although standardized cognitive screening scores such as the Mini-Mental State Examination (MMSE) were not available for all participants, the clinical recruitment process ensured that participants within each diagnostic group had comparable clinical presentations based on hospital assessment.

Prior to the formal experiment, all participants completed a 1-minute VR familiarization session using VR content different from that used during the main experiment. This procedure was performed to evaluate tolerance to VR stimulation and minimize adaptation effects. No participant reported dizziness, cybersickness, or other adverse reactions during the familiarization procedure. Visual function was screened prior to enrollment. All participants were able to complete the VR procedure without corrective eyewear and reported no visual difficulties during VR stimulation.

A participant flow diagram summarizing recruitment, eligibility screening, group allocation, VR-EEG recording, and final analysis is provided in Multimedia Appendix 1. Available demographic and clinical characteristics are summarized in Table 1.

Because this was an exploratory single-session, 3-group repeated-measures EEG study, no formal a priori power calculation was performed. The sample size was determined based on participant availability during the recruitment period and the feasibility of simultaneous VR-EEG recordings in older adults with cognitive impairment. The final sample included 20 participants in each diagnostic group, allowing balanced group-level exploratory comparisons of EEG-derived network metrics.

EEG data were collected using a 3-phase experimental paradigm consisting of a pre-VR resting state, a VR stimulation state, and a post-VR resting state.

VR stimulation was delivered using a PICO 4 head-mounted display (PICO Technology). Participants viewed an immersive 3D roller-coaster environment from a first-person perspective. The VR environment consisted of a passive first-person roller-coaster scenario that provided dynamic visual motion cues within an immersive 3D space. The paradigm was designed to provide visual and vestibular-like motion stimulation through optic flow, but participants were not required to perform motor responses, make decisions, or complete memory tasks. The VR stimulation lasted approximately 5 minutes. Participants remained seated comfortably in an upright position throughout the experiment.

Participants were monitored for discomfort, dizziness, and cybersickness during both the familiarization procedure and the main VR stimulation session. No participant discontinued the experiment because of VR-related discomfort. After the removal of the VR stimulus, a post-VR resting state was conducted to evaluate subsequent changes in brain activity. Each phase lasted approximately 5 to 6 minutes, and a 15-minute interval was provided between phases to minimize potential carryover effects.

The pre-VR, VR, and post-VR recordings were repeated within-session experimental states and did not represent longitudinal follow-up assessments.

The primary neurophysiological measures were EEG-derived graph-theoretical metrics, including GE, modularity, and nodal strength, calculated from PLV-based functional connectivity networks. Diagnostic group and experimental state were the main grouping variables. Standardized cognitive screening scores, such as MMSE, were not available for all participants and were therefore not included as covariates in the statistical analyses.

EEG signals were recorded using a HydroCel Geodesic Sensor Net (Electrical Geodesics) coupled with an EGI NetAmps 300 amplifier (Electrical Geodesics). A total of 64 electrodes were placed according to the International 10-20 system. Signals were sampled at 1000 Hz, and Cz was used as the reference electrode throughout acquisition.

Electrode impedance was maintained within clinical recording standards during data acquisition. Participants were instructed to remain seated, relaxed, and to minimize head, facial, and limb movements throughout recording. EEG data were collected during all 3 experimental states: pre-VR resting, VR stimulation, and post-VR resting. During the pre-VR and post-VR resting states, participants remained awake and visually fixated without active task performance. During VR stimulation, the headset was positioned carefully to avoid excessive pressure on the EEG electrodes and to minimize mechanical artifacts during simultaneous VR-EEG recording [31,32].

EEG preprocessing was performed using EEGLAB. Raw EEG data were imported from EGI MFF files, and electrode locations were assigned using a 68-electrode channel-location file. Before independent component analysis (ICA), the data were high-pass filtered at 1 Hz. The data were then rereferenced using T7 and T8 as reference channels; if the specified reference channels were unavailable or failed validation, average referencing was used instead.

ICA-based artifact correction was performed using the extended runica algorithm in EEGLAB. To account for rank reduction after rereferencing, ICA decomposition was performed with principal component analysis (PCA) dimensionality reduction, using up to 64 principal components. Independent components were classified using ICLabel. Components classified as muscle, eye, heart, line-noise, or channel-noise artifacts with a probability of 0.90 or higher were automatically flagged and removed, whereas components classified as brain or other were not automatically rejected. Across the analyzed ICA-cleaned EEG recordings, the number of removed independent components was 7.68 (SD 7.32) per recording, with a median of 6 and a range of 0 to 32 components. After ICA-based artifact correction, the data were low-pass filtered at 70 Hz and saved as ICA-cleaned EEG datasets.

Following artifact correction, EEG signals were decomposed into multiple frequency bands, including δ (1-4 Hz), θ (4-8 Hz), α (8-13 Hz), β (13-32 Hz), γ (32-70 Hz), and a full-spectrum band (1-70 Hz) [33]. After preprocessing, artifact-free 1-second epochs were extracted for further analysis. For each participant and experimental phase, 30 clean epochs were selected and used for subsequent functional connectivity analysis.

Functional connectivity was quantified using the PLV, which measures the consistency of phase differences between 2 signals over time. For channels i and j, PLV was computed as:

Functional brain networks were constructed based on PLV connectivity matrices derived from the preprocessed EEG signals. In the resulting network representation, each node corresponded to an EEG recording channel, and edges represented the PLV-based functional connectivity strength between pairs of channels.

For descriptive network visualization, a proportional thresholding approach was applied to the group-averaged PLV connectivity matrices. Specifically, only the top 5% of the strongest PLV connections were retained in the visualization graphs, while weaker connections were not displayed. This thresholding step was used to improve the interpretability of the network maps and to highlight the dominant spatial distribution of strong functional connections.

The 5% proportional threshold was selected as a literature-informed and visualization-oriented threshold. Previous EEG graph-theoretical studies have displayed only the strongest subset of functional connections to improve the interpretability of network topology and connection changes. For example, Liang et al [35] visualized only the upper 30% of mutual information values in a 19-channel EEG network to better illustrate network connection changes. In the present study, the number of EEG nodes and possible pairwise connections was substantially larger. Therefore, retaining higher proportions of connections, such as 20% or 10%, produced dense and visually cluttered graphs, with many overlapping edges that obscured the dominant spatial connectivity patterns. The 5% threshold provided the clearest visualization of the strongest PLV connections and regional network differences. Therefore, the top 5% of strongest PLV connections were retained only for network visualization and descriptive connectivity mapping. GE and modularity were calculated from weighted PLV matrices, whereas the top 5% proportional threshold was applied only to descriptive network visualizations.

Brain network visualizations were generated to illustrate the spatial distribution of functional connectivity and to highlight long-range and interhemispheric connection patterns across experimental conditions. Connectivity graphs were projected onto scalp topography to provide an intuitive representation of network organization for each group and experimental state [35,39].

For statistical evaluation, graph-theoretical metrics were compared across diagnostic groups (AD, MCI, and NC) and experimental states (pre-VR resting, VR stimulation, and post-VR resting). Between-group differences within each experimental state were first assessed using 1-way ANOVA. When significant group effects were detected, post hoc pairwise comparisons using independent-sample 2-tailed t tests were performed to identify specific between-group differences. Within-group comparisons across experimental states were performed using paired-sample t tests. These statistical analyses were conducted to determine whether VR stimulation was associated with measurable modulation of brain network topology across the different cognitive states [40,41]. When reported, 95% CIs were calculated using the t distribution based on the corresponding group sample size.

Statistical significance was initially defined as P<.05. To control for multiple comparisons, P values were adjusted using the Benjamini-Hochberg false discovery rate (FDR) procedure [42]. FDR correction was applied separately within each graph-theoretical metric and comparison family. Specifically, between-group comparisons within the same experimental state and within-group comparisons across experimental states were treated as separate comparison families. Both raw P values and FDR-adjusted P values are reported where applicable.

No participant-level data were missing in the final dataset. During EEG preprocessing, artifact-contaminated segments were removed according to the predefined quality-control procedures. For each participant and experimental condition, 30 artifact-free epochs were retained for subsequent analyses, ensuring an equal number of observations across groups and experimental states. The overall VR-EEG experimental and functional network analysis workflow is shown in Figure 1.

Unless otherwise specified, inferential statistical conclusions were based on graph-theoretical metrics, including GE and modularity. PLV matrices, nodal strength distributions, and node connection graphs were used primarily as descriptive visualizations to illustrate spatial connectivity patterns and to support interpretation of the quantitative network results.

The PLV connectivity matrices were used as descriptive visualizations to illustrate broadband functional connectivity patterns across diagnostic groups and experimental states. In the pre-VR resting state, the NC group visually showed stronger broadband coupling, whereas the AD group showed weaker and less distributed connectivity patterns, with the MCI group showing intermediate features.

During VR stimulation, the PLV matrices suggested a change in broadband connectivity organization, particularly in the NC and AD groups. In the NC group, broadband PLV patterns appeared less globally dense during VR stimulation, whereas the AD group showed relatively limited changes in overall coupling strength. In the post-VR resting state, visual differences among the 3 groups appeared less pronounced than those observed at baseline.

Because these PLV matrices were intended primarily for visualization, formal statistical interpretation was based on subsequent graph-theoretical metrics, including GE and modularity. These descriptive PLV patterns provide spatial context for the quantitative network changes reported below. The broadband PLV connectivity matrices across diagnostic groups and experimental states are shown in Figure 2.

Figure 3 illustrates GE values across frequency bands and experimental states in the NC, MCI, and AD groups. Descriptively, GE values tended to be higher in the 13 to 32 Hz range than in lower-frequency and higher-frequency bands across several groups and experimental states. This pattern suggests that beta-band activity may contribute to large-scale network integration in the present EEG dataset. However, because frequency-band comparisons were primarily descriptive, this observation should be interpreted cautiously and not as confirmatory evidence of a frequency-specific network mechanism.

In the broadband frequency range (1-70 Hz), between-group differences in GE were most evident during the pre-VR resting state. In the pre-VR condition, patients with AD exhibited higher GE than both NC participants (AD: 0.425, SD 0.036, 95% CI 0.408-0.442; NC: 0.337, SD 0.081, 95% CI 0.299-0.375; raw P=.009; FDR-adjusted P=.04) and patients with MCI (MCI: 0.343, SD 0.064, 95% CI 0.313-0.373; raw P=.003; FDR-adjusted P=.03). During VR stimulation, the difference in GE between the AD and NC groups was no longer significant (AD: 0.411, SD 0.033, 95% CI 0.396-0.426; NC: 0.412, SD 0.065, 95% CI 0.382-0.442). GE in the AD group was higher than that in the MCI group during VR stimulation in the uncorrected analysis (MCI: 0.355, SD 0.071, 95% CI 0.322-0.388; raw P=.04); however, this comparison did not remain significant after FDR correction (FDR-adjusted P=.12) and was therefore interpreted as an exploratory trend. In the post-VR state, GE values in the AD and NC groups converged, and no significant difference was observed between the 2 groups.

A comparable descriptive trend was observed in the high-frequency band (32-70 Hz); however, no between-group or within-group state comparison in this frequency band remained significant after FDR correction. Therefore, high-frequency GE findings were interpreted as exploratory patterns rather than confirmatory statistical effects. Across all experimental states, the MCI group generally exhibited intermediate GE values between the AD and NC groups, accompanied by moderate state-dependent variations. These results suggest that VR stimulation was associated with dynamic changes in network integration, with reduced AD-NC group differences in GE during and after VR stimulation. Each panel shows the distribution of GE values for the 3 diagnostic groups, illustrating group-level differences in network integration under different VR conditions in Figure 4.

These findings suggest distinct patterns of GE modulation across the clinical groups. At baseline, patients with AD exhibited elevated GE, suggesting altered or dysregulated network integration. During VR stimulation, GE in the AD group showed a decreasing trend; however, within-group GE changes in the AD group did not remain significant after FDR correction.

In contrast, participants in the NC group demonstrated a significant increase in GE during VR stimulation (pre-VR: 0.337, SD 0.081, 95% CI 0.299-0.375; VR: 0.412, SD 0.065, 95% CI 0.382-0.442; raw P=.007; FDR-adjusted P=.04). This increase suggests state-dependent upregulation of large-scale functional integration in response to sensory stimulation. The MCI group did not show an immediate increase in GE during the VR phase; instead, GE increased from pre-VR resting to the post-VR state after FDR correction (raw P=.009; FDR-adjusted P=.04), suggesting a delayed network response following stimulation.

This temporal pattern suggests that network responsiveness in MCI may be partially preserved compared with AD, but that the dynamic regulation of global integration may occur more slowly. Taken together, these findings suggest a possible continuum of network dysfunction from NC to MCI to AD, characterized by rapid GE modulation in NC, delayed GE changes in MCI, and limited within-group GE modulation in AD. Each panel illustrates the distribution of GE values across experimental states, highlighting state-dependent modulation of network integration within the AD, MCI, and NC groups in Figure 5.

Modularity analysis revealed a robust baseline group difference and descriptive state-dependent patterns across clinical groups, partially paralleling the patterns observed for GE [43]. In the pre-VR resting state, patients with AD exhibited higher modularity than NC participants, and this difference remained significant after FDR correction (AD: 0.446, SD 0.036; NC: 0.273, SD 0.120; raw P=.001; FDR-adjusted P=.01), while the MCI group showed intermediate values (0.363, SD 0.146). During VR stimulation, modularity values converged across the 3 groups, and no significant between-group differences were observed after FDR correction. In the post-VR state, modularity values in the AD and MCI groups were comparable, and both were numerically higher than those of the NC group. The MCI-NC difference was significant in the uncorrected analysis but did not remain significant after FDR correction (raw P=.02; FDR-adjusted P=.07); therefore, this result was interpreted as an exploratory trend, as shown in Figure 6.

Within-group comparisons across experimental states showed descriptive temporal modulation patterns, but no within-group state comparison remained significant after FDR correction. In the AD group, modularity numerically decreased during VR stimulation, suggesting a possible reduction of the highly segregated network organization observed at baseline. The MCI group showed minimal numerical change during VR but exhibited higher modularity in the post-VR phase, suggesting a possible delayed network reconfiguration pattern. In contrast, participants in the NC group showed a numerical increase in modularity during VR stimulation followed by a reduction in the post-VR state. These within-group modularity patterns should therefore be interpreted as descriptive trends rather than confirmatory statistical effects.

This study investigated alterations in EEG-based nodal connection strength during the pre-VR and VR conditions, illustrating VR-associated changes in functional brain network patterns in patients with AD. As shown in Figure 7, during the pre-VR resting state, the AD group exhibited pronounced right-lateralized connectivity, with higher nodal strength in the right hemisphere than in the left, whereas the NC group showed relatively balanced interhemispheric connectivity patterns.

During VR stimulation, the AD group demonstrated a bidirectional modulation pattern: most nodes in the left hemisphere showed increased connection strength, while several nodes in the right hemisphere displayed decreased strength. This “left-up and right-down” pattern suggested a shift in interhemispheric connectivity distribution during VR stimulation, with the AD connectivity pattern appearing more balanced than in the pre-VR resting state.

Analysis of key midline hub nodes (Pz, POz, Fz, and Oz) further revealed that the NC group exhibited variable changes with generally higher nodal strength, whereas the AD group showed consistent increases in nodal strength across these central regions during VR stimulation. As illustrated in Figure 7, these changes appeared to reduce visual group differences in these core integration nodes.

Taken together, these descriptive nodal strength patterns suggest that VR stimulation was associated with a shift in nodal connectivity distribution in AD, characterized by increased left-hemisphere and midline involvement and reduced apparent right-lateralized dominance. Because nodal strength maps were used primarily for descriptive visualization, these findings should be interpreted as spatial patterns of regional connectivity rather than evidence of therapeutic efficacy or functional improvement. These descriptive findings provide spatial context for the quantitative GE and modularity results, which indicated state-dependent changes in large-scale network organization.

Node connection graphs were used as descriptive visualizations of the top 5% strongest PLV connections and were not treated as independent inferential statistical tests or threshold-sensitivity analyses. Using preprocessed EEG data in the .set format and retaining the top 5% of strongest connections, the NC group showed stronger and more spatially distributed retained PLV connections than the MCI and AD groups, with relatively balanced interhemispheric organization. During VR stimulation, the retained connections in the NC group appeared more selective and spatially distributed across the network, suggesting a descriptive shift in the organization of strong PLV connections under sensory stimulation [44].

In contrast, the AD group exhibited marked hemispheric asymmetry at rest, with connectivity predominantly concentrated in the right hemisphere, particularly in central and centroparietal regions. This central and centroparietal concentration may reflect compensatory reliance on sensorimotor integration hubs and midline relay regions when distributed long-range coordination is reduced, consistent with more locally constrained coupling in disease. During VR stimulation, hemispheric asymmetry decreased, and the network adopted a more symmetric configuration, suggesting a state-associated shift toward a more bilateral distribution of retained connections and redistributed connectivity away from an overly lateralized resting-state pattern. The MCI group showed intermediate features between NC and AD but lacked a stable configuration across states, consistent with a transitional and heterogeneous network phenotype.

Overall, VR stimulation was associated with a more selective and spatially balanced distribution of retained connections in NC networks. In AD, VR stimulation was associated with reduced apparent resting-state lateralization and a more bilateral distribution of retained connections. These connectivity reorganization patterns across groups and states are illustrated in Figure 8.

To further evaluate VR-related network changes in AD, differential connectivity maps were generated by subtracting the pre-VR network from the VR-state network within the AD group across frequency bands. In the broadband range (1-70 Hz) and in the mid-frequency band (8-13 Hz), VR-related connectivity increases were mainly localized to the left hemisphere, with prominent involvement of central and centroparietal regions. This pattern contrasts with the right-hemisphere dominance observed at rest and suggests a state-associated shift toward more bilateral hemispheric distribution.

In the high-frequency band (32-70 Hz), descriptive maps showed a different spatial distribution of retained long-range connections during VR stimulation. Because high-frequency GE findings did not remain significant after FDR correction, this pattern should be interpreted as visualization-based spatial context rather than confirmatory evidence of enhanced high-frequency communication.

Taken together, the node connection graphs descriptively suggested that VR stimulation was associated with a shift in AD connectivity patterns from a more lateralized resting-state configuration toward a more bilateral distribution. These visualization-based observations should be interpreted together with the quantitative GE and modularity findings rather than as independent statistical evidence. Overall, the connectivity maps provide spatial support for the interpretation that VR stimulation was associated with short-term changes in large-scale functional network organization.