Effects of multi-sensory virtual reality training on gait adaptability and somatomotor network remodeling in patients with stroke: a randomized controlled trial Authors: Qiurong Xie, Xiao Xiong, Mingchao Liu, Yuxin Huang, Bo Sheng, Yanxin Zhang, Jia Huang, Qi Zhang Published: 2026 ABSTRACT ObjectiveTo investigate the effects of multisensory-integrated virtual reality (VR) training on gait adaptability and its regulatory mechanisms on the somatomotor network (SMN) in patients with stroke.MethodsIn this randomized controlled trial, 68 patients with stroke were allocated to a VR group (multisensory-integrated VR training) or a control group (conventional rehabilitation). Both groups received 30-minute sessions, 5 days/week for 4 weeks. The primary outcome was gait adaptability assessed by the Dynamic Gait Index (DGI). Secondary outcomes included the Timed Up and Go Test (TUGT), Berg Balance Scale (BBS), and Fugl-Meyer Assessment for Lower Extremity (FMA-LE). Functional near-infrared spectroscopy (fNIRS) measured resting-state functional connectivity within the SMN and task-evoked activation during stepping and obstacle crossing.ResultsThe VR group showed significantly greater improvements than the control group in DGI total score (P = 0.010), TUGT (P = 0.005), and BBS (P < 0.001 ). fNIRS analysis revealed that the VR group exhibited significantly greater increases in task-evoked activation in the right posterior parietal cortex (PPC) and supplementary motor area (SMA) during stepping (P = 0.029 and P = 0.032, respectively), and in the right SMA during unaffected-limb obstacle crossing (P = 0.048). Resting-state functional connectivity analysis showed significantly enhanced connections within the SMN, including left SMA-right PPC and right PPC-left dorsolateral prefrontal cortex (DLPFC) (both P < 0.05). Correlation analyses revealed that increased right SMA activation during obstacle crossing was positively correlated with TUGT improvement (r = 0.590, P = 0.001), while enhanced right PPC-left DLPFC connectivity was positively correlated with DGI improvement (r = 0.403, P = 0.041).ConclusionMultisensory-integrated VR training was associated with improvements in gait adaptability and balance in patients with stroke. The underlying mechanisms may involve enhanced activation in the SMA and PPC, along with changes in functional connectivity within the SMN and between the SMN and cognitive control networks. However, given that the significant improvements in DGI, TUGT, and BBS were not sustained under the most conservative assumptions about missing data, these findings should be considered preliminary and warrant confirmation in studies with lower attrition rates.Trial registration: Chinese Clinical Trial Registry, ChiCTR2500111919 (retrospectively registered). Registered 7 November 2025. Available from https//www.chictr.org.cn (registration number ChiCTR2500111919). Protocol The full trial protocol is available from the corresponding author upon reasonable request.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12984-026-02003-9. Introduction Gait adaptability impairment is a significant functional deficit in patients with stroke, specifically the ability to rapidly and appropriately adjust walking patterns in response to visual stimuli [1]. Studies indicate that approximately one-third of patients continue to face difficulties with community ambulation, struggling with terrain changes, obstacle avoidance, and dual-tasking, even after regaining walking ability [2]. This is primarily due to a significant decline in gait adaptability, which directly limits daily activity capacity and negatively impacts quality of life and social participation. Gait adaptability in real-world environments requires not only motor coordination but also cognitive processing, such as attention allocation, response inhibition, and visuospatial processing. This phenomenon, known as cognitive-motor interference, is particularly pronounced in patients with stroke due to impaired automaticity of gait [3]. Dual-task gait training, which combines motor tasks with concurrent cognitive tasks, has been shown to improve both gait performance and cognitive function in stroke populations [4, 5]. However, conventional dual-task training often lacks ecological validity and real-time feedback. Virtual reality(VR) offers a promising platform for integrating motor and cognitive demands in a controlled, engaging environment [6]. In recent years, VR-based gait rehabilitation training, which leverages its immersive, interactive, and multisensory-integrated characteristics (combining visual, auditory, tactile, and other stimuli), has provided patients with rich feedback and simulated training environments. It has been proven effective in enhancing patient engagement and walking capacity [7]. This training modality can safely simulate complex community scenarios (e.g., obstacles, varying terrain) for repeated adaptive practice. Evidence-based practice already recommends VR as an essential adjunct to conventional rehabilitation [8]. Research confirms that VR training is superior to standard overground training in improving walking, static balance, and gait adaptability [9]. Its mechanism of action may involve functional changes and the promotion of neuroplasticity within the central nervous system [10]. Furthermore, studies [11] have found that VR-based variable-speed walking training can specifically improve balance and gait parameters. The dynamic adjustment capacity of these gait parameters (e.g., walking speed, step length) in response to environmental changes is itself a key behavioral indicator of gait adaptability [12, 13]. The neural basis of gait adaptability is closely associated with the Somatomotor Network (SMN) [14]. The SMN encompasses key brain regions, including the supplementary motor area (SMA), posterior parietal cortex (PPC), primary motor cortex (M1), and premotor cortex (PMC). Among these, the SMA is responsible for gait initiation and termination, and plays a central role within the network [15, 16]. The PPC specializes in visuomotor integration, updating motor plans online in response to environmental feedback [16]. The M1 serves as the final hub for executing precise motor commands [17]. During complex environmental walking, not only is intra-network coordination within the SMN enhanced (SMA planning, PPC integration, M1 execution), but activation in cognitive regions, such as the prefrontal cortex, also increases significantly to allocate more attentional resources [18]. Research indicates the PPC plays a crucial role in visually guided gait [19], and its functional integrity is essential for motor adaptability [20]. Stroke-induced damage to the SMN and related networks is the neurological root cause of impaired gait adaptability. Based on principles of central nervous system plasticity, task-oriented, repetitive training rich in multi-sensory stimulation is considered key to optimizing motor networks and promoting functional reorganization [10, 16]. Previous studies have established that such training can significantly activate regions including the PMC, SMN, and PPC in patients with stroke, with the degree of activation correlating with improvements in motor function [21, 22]. These findings provide an essential premise: multi-sensory stimulation engages frontoparietal regions involved in motor planning and visuomotor integration. However, whether multisensory-integrated VR training specifically modulates cortical activity within the SMN—particularly the SMA and PPC—in the context of gait adaptability, and whether such modulation correlates with behavioral improvements, remains unclear. Therefore, the present study aimed to test this hypothesis directly using fNIRS. Therefore, multisensory-integrated VR gait training could, in theory, facilitate the functional reorganization of the SMN by simulating complex tasks and providing rich feedback, thereby enhancing gait adaptability. However, while these previous studies demonstrate that VR-related activation occurs in regions such as the PMC and PPC, they do not directly test whether VR training specifically modulates cortical activity within the SMN—particularly in the SMA and PPC—in the context of gait adaptability. Moreover, the causal link between such cortical modulation and clinical functional improvements in gait adaptability remains to be established. In summary, multisensory-integrated VR gait training holds promise for improving gait adaptability in patients with stroke by potentially driving functional reorganization within SMN. Building on the established premise that multi-sensory stimulation engages frontoparietal regions [21, 22], Specifically, we hypothesize that: (1) compared to conventional rehabilitation, multisensory-integrated VR training will lead to greater improvements in clinical functional assessment outcomes reflecting gait adaptability; and (2) these clinical functional assessment outcomes improvements will be associated with enhanced activation in the SMA and PPC, as measured by functional near-infrared spectroscopy(fNIRS). In this framework, clinical functional improvements are considered primary outcomes expected to result from training-induced modulation of the SMA and PPC, while the neural changes serve to elucidate the underlying mechanisms of these behavioral effects. To test this hypothesis, this study will employ fNIRS to quantitatively analyze changes in SMN cortical activation before and after training, aiming to elucidate the underlying neural mechanisms of its therapeutic effect. Materials and methods Study design This study was a prospective, parallel-group, assessor-blinded, randomized controlled trial with a 1:1 allocation ratio. The Ethics Committee of the same institution approved the study protocol. Initial approval was obtained in August 2023 (Approval No. 2023KY-020-02) for investigating the behavioral effects of VR training. In May 2024, a protocol amendment was submitted to incorporate fNIRS measurements to explore neural mechanisms. This amendment was approved in June 2024 (Approval No. 2024YJS-007-02) before participant enrollment for the Dynamic Gait Index (DGI) and fNIRS data collection. No other changes were made to the study methods after trial commencement. All participants provided written informed consent after being fully informed about the study details before enrollment. The trial was prospectively registered with the Chinese Clinical Trial Registry. The full protocol was submitted by the end of 2023, and the registration (No. ChiCTR2500111919) was formally posted on the public platform in November 2025 upon completion of the standard administrative review process. Sample size The primary outcome measure for this study was DGI. Based on previous literature [23], the expected post-intervention DGI scores were 15.66 ± 2.69 for the control group and 19.2 ± 2.11 for the VR group. Using GPower software (version 3.1.9.2) with a two-tailed α = 0.05 and a power (1-β) of 0.85, a minimum of 27 subjects per group was calculated. Considering an estimated dropout rate of approximately 20%, the final total sample size was determined to be 68 subjects (34 per group). No interim analyses were planned or conducted, and there were no stopping guidelines. A post hoc power analysis indicated that the actual statistical power of this study was 0.81 (> 0.80), meeting the requirement. Participants The diagnosis of stroke was established according to the diagnostic criteria outlined in the “Key Diagnostic Points for Major Cerebrovascular Diseases in China (2019 Edition)”. The inclusion criteria were: (1) aged between 30 and 70 years; (2) within 2 weeks to 6 months post-stroke onset; (3) gait dysfunction secondary to stroke (Functional Ambulation Category ≥ 2); (4) stable vital signs; (5) adequate cognitive function (Mini-Mental State Examination score ≥ 24) to comprehend and follow instructions; (6) no unilateral spatial neglect or hemianopia; and (7) provision of voluntary written informed consent. Exclusion criteria were: (1) other neurological or musculoskeletal conditions that could significantly impair walking (e.g., Parkinson’s disease, severe arthritis); (2) severe cognitive impairment or depression hindering cooperation; (3) major systemic illnesses (e.g., cardiac, pulmonary, renal) or unstable cardiovascular conditions; (4) deep vein thrombosis in the lower limbs; (5) severe visual or auditory deficits that would interfere with training or assessment. Assessment of unilateral spatial neglect and hemianopia was based on medical record documentation or clinical examination by a neurologist before enrollment. Participants with a documented diagnosis of either condition were excluded. A total of 629 patients with stroke admitted to the inpatient and outpatient rehabilitation units of the Rehabilitation Hospital Affiliated to Fujian University of Traditional Chinese Medicine (Fuzhou, China) were initially screened for this study from June 2024 to January 2025. Following strict application of the inclusion and exclusion criteria, 583 individuals were excluded. Reasons for exclusion were: 505 did not meet the inclusion criteria, 22 met one or more exclusion criteria, and 56 declined to participate. Consequently, 68 eligible patients who provided voluntary written informed consent were enrolled. All enrolled participants were randomly assigned in a 1:1 ratio using a random number table to either the VR group or the control group, with 34 patients in each group. Randomization, allocation concealment, and blinding A computer-generated random number table (SPSS 26.0) was used to randomly allocate subjects to either the VR group or the control group in a 1:1 ratio, with 34 subjects per group. Simple randomization was used without any restrictions such as blocking or stratification. The random sequence was generated by an independent statistician not involved in recruitment, intervention, assessment, or data analysis, and was sealed in sequentially numbered, opaque envelopes to ensure allocation concealment. Eligible subjects who completed the baseline assessment were assigned by the study coordinator, who opened the corresponding envelope in the order of enrollment and informed the subject and their intervention therapist of the assigned group. This study was a randomized controlled trial with blinded assessors and statisticians. Due to the nature of the interventions (VR vs. conventional), blinding of subjects and intervention providers was not possible. However, the assessors who conducted all outcome measurements (functional assessments) and the researchers who performed the final statistical analysis remained blinded to group allocation throughout the study. An independent sequence keeper held group information (represented by codes A/B). Unblinding occurred only after the database was locked and the preliminary statistical analysis was completed. Intervention similarity Both groups received 60 min of daily rehabilitation, with the VR group receiving 30 min of conventional rehabilitation (identical to the control group) plus 30 min of VR training. The total dose and physical demands were matched between groups, while cognitive demands differed intentionally as the independent variable of interest. Further details on intervention equivalence are provided in the Interventions section. Intervention Control group The control group received conventional rehabilitation therapy supplemented with structured walking training. The total daily session lasted 60 min and was administered once daily, 5 days a week, over 4 weeks. The conventional rehabilitation component (30 min) was delivered by an experienced therapist based on individual functional assessments. Training, grounded in neurodevelopmental approaches (e.g., Bobath, Brunnstrom, Rood techniques), focused on weight shifting, trunk and pelvic control, sit-to-stand transfers, standing balance, stair negotiation, and basic walking training, aiming to inhibit abnormal movement patterns and facilitate isolated movements. The structured walking training (30 min) was designed to enhance lower limb function and gait adaptability, comprising the following progressive tasks: ① Unilateral weight-bearing training: Performed in front of parallel bars, involving alternating support on the affected and unaffected legs, progressing from partial to full weight-bearing, to maintain static balance for 1–2 min. ② Wall-based hip extension exercises: Standing with the back against a wall and heels approximately 20 cm away, participants performed hip thrusts to lift the shoulders and back off the wall, holding the isometric contraction for 10 s to strengthen hip and trunk extension control. ③ Step and obstacle training with affected limb support: Included step ascent and descent led by the affected leg, and obstacle crossing performed by the unaffected leg while maintaining a standing position with hip extension on the affected side. ④ Dynamic coordination training: Involved alternate stepping in place while maintaining the wall-based hip extension posture, and lateral/anteroposterior weight-shifting and stepping exercises within parallel bars. VR group The VR group received the identical conventional rehabilitation training as the control group (30 min/day). In addition, they underwent multisensory-integrated VR gait adaptability training (30 min/day). The total daily intervention duration, frequency, and 4-week period were thus the same for both groups. The multisensory-integrated VR training system was independently developed by our research team, based on principles of motor control and learning. The system used a Kinect Azure depth camera to capture real-time lower-limb movements, which were simultaneously projected onto a 43-inch Liquid Crystal Display (LCD) screen, providing immersive visual feedback. The training integrated visual and auditory cues, simulating community environments (e.g., forest paths, streets), and consisted of four core task modules: ① Obstacle crossing: Participants were required to lift their legs to cross virtual obstacles in response to on-screen visual and auditory prompts. This task primarily targets motor planning and coordination, while also requiring visuospatial processing to judge obstacle height and timing. The system recorded the number of successful crossings and task completion time. ② Target kicking/avoidance tasks (Soccer/Kicking bombs): Different colored soccer balls or bomb icons appeared on the screen. Participants needed to quickly lift the corresponding leg (indicated by color bands for left/right) to “kick” the ball or “avoid” the bomb. Successful hits and avoids were recorded. This task incorporates Go/No-Go inhibition (responding to balls while inhibiting response to bombs) and selective attention, representing a cognitive-motor dual-task paradigm. ③ Directional reactive stepping: Viewing the screen from a top-down perspective, participants were required to step quickly in different directions (forward, backward, left, right) to catch approaching arrows. Correct direction and timing were judged as successful, triggering real-time visual (an arrow turning green) and auditory feedback, with success rates recorded. This task demands rapid visuospatial processing, attention switching, and motor planning, engaging both cognitive and motor systems simultaneously. ④ Squat-stand task: Participants needed to promptly squat down, mimic a holding action, and return to standing in response to a virtual wooden barrel rolling towards them on the screen. The number of successful completions was recorded. This task involves timing prediction and reactive postural control. Details are shown in Fig. 1. Screen-capture videos demonstrating each of the four game modules are available as Supplementary Videos 1–4. Training implementation and progression All training was conducted in a standing position on a stable floor surface (stepping in place) under the supervision of a therapist, with a safety harness used as necessary. Licensed physical therapists delivered all interventions with at least 3 years of experience in stroke rehabilitation. Before the study, all therapists underwent a 2-day standardized training session on the VR system and the structured walking protocol to ensure consistency across participants. Task difficulty (e.g., obstacle height, presentation frequency) was dynamically adjusted based on individual performance to ensure the training remained appropriately challenging. The 30-minute VR training session comprised actual active training time, excluding setup and rest periods. The session was structured as follows: participants completed the four task modules in sequence, with each module lasting approximately 7–8 min. Brief rest intervals (1–2 min) were provided between modules to prevent fatigue, as continuous standing and stepping tasks could be physically demanding for patients with stroke. During these rest intervals, participants remained standing or sat down as needed, and water intake was permitted. The therapist monitored participants’ vital signs and perceived exertion throughout the session, and additional rest was allowed if necessary. Thus, the pure active exercise time was 30 min, while the total session duration (including rest breaks) was approximately 35–40 min. VR training was conducted in a quiet, dimly lit room with a minimum space of 3 m × 3 m to allow safe stepping movements. The therapist remained within arm’s reach throughout the session. Fig. 1The four core modules of the VR training system. A Obstacle crossing: Participants lift their legs to cross virtual obstacles in response to visual and auditory prompts. B Target kicking/avoidance (Soccer/Kicking bombs): Participants kick soccer balls (Go) and avoid bombs (No-Go) based on color cues. C Directional reactive stepping: Participants step quickly in different directions to catch approaching arrows. D Squat-stand task: Participants squat down and return to standing in response to a rolling virtual barrel. Videos of each task are available as Supplementary Videos 1–4 The four core modules of the VR training system. A Obstacle crossing: Participants lift their legs to cross virtual obstacles in response to visual and auditory prompts. B Target kicking/avoidance (Soccer/Kicking bombs): Participants kick soccer balls (Go) and avoid bombs (No-Go) based on color cues. C Directional reactive stepping: Participants step quickly in different directions to catch approaching arrows. D Squat-stand task: Participants squat down and return to standing in response to a rolling virtual barrel. Videos of each task are available as Supplementary Videos 1–4 Equivalence of interventions To ensure that the observed effects could be attributed specifically to the multi-sensory VR component, we carefully considered the equivalence between the control and VR interventions across several dimensions: Training dose and intensity: Both groups received a total of 60 min of daily intervention. The VR group received 30 min of conventional rehabilitation (identical to the control group’s conventional component) plus 30 min of multisensory-integrated VR training. The control group received 60 min of conventional rehabilitation, which included 30 min of conventional neurodevelopmental therapy and 30 min of structured walking training. Thus, the additional 30 min of VR training in the VR group represents the specific component being tested, while the conventional rehabilitation dose (30 min) and total intervention duration (60 min) were matched between groups. Physical demands: Both interventions required participants to maintain a standing position and to perform lower-limb movements throughout the training session. In the control group, the structured walking training included unilateral weight-bearing, wall-based hip extension exercises, step and obstacle training with affected limb support, and dynamic coordination tasks (e.g., alternate stepping, weight-shifting). In the VR group, the physical demands involved similar weight-bearing and stepping activities within the virtual environment, including obstacle crossing, directional stepping, and squat-stand tasks. Although the specific movement patterns differed, both interventions were designed to be physically engaging and functionally relevant to gait rehabilitation, with comparable emphasis on lower limb loading, postural control, and dynamic balance. Notably, the total duration of physical activity (60 min) was identical between groups. Cognitive and attentional demands: This dimension differed intentionally between groups, as the cognitive engagement provided by VR is hypothesized to be a key mechanism of its therapeutic effect. The control group received conventional rehabilitation, which primarily focuses on motor execution with minimal explicit cognitive challenge. In contrast, the VR group performed tasks that explicitly required response inhibition (Go/No-Go), visuospatial processing, attention switching, and dual-task coordination (as detailed in the VR training description). This difference in cognitive demand is not a confounder but rather the independent variable of interest—the “active ingredient” hypothesized to drive additional improvements in gait adaptability. Therefore, rather than matching cognitive demands, we intentionally varied them to test whether multi-sensory VR training confers benefits beyond conventional motor rehabilitation. In summary, the two interventions were matched in total training dose and comparable in physical demands, while differing in cognitive and attentional engagement—a difference that constitutes the core experimental manipulation of this study. Outcome measures General information, including demographic data (age, gender, years of education, body mass index) and stroke-related clinical characteristics (type, lesion side, lesion location, duration, medical history, etc.), was collected at baseline by blinded assessors. All functional outcome measures were assessed by blinded assessors at baseline (before the intervention) and immediately after the 4-week intervention period. The primary outcome was gait adaptability, estimated using DGI. This scale consists of 8 tasks simulating daily environmental challenges (e.g., walking at changing speeds, turning the head, stepping over obstacles, pivoting), each scored from 0 to 3, yielding a total score of 24. A higher score indicates better walking and balance capacity in complex environments [24]. Secondary outcomes included: ①Functional mobility, assessed using the Timed Up and Go Test (TUGT), which records the time (in seconds) required to stand up from a chair, walk 3 m, turn around, walk back, and sit down. A shorter time indicates higher functional mobility efficiency and lower fall risk [25]. ② Static and dynamic balance function, assessed using the Berg Balance Scale (BBS). This 14-item scale, with a total score of 56, is widely used to evaluate balance in patients with stroke; a score below 40 indicates increased fall risk [26]. ③Lower limb motor function, assessed using the Fugl-Meyer Assessment for Lower Extremity(FMA-LE). This scale evaluates reflexes, synergies, and isolated movements, among other aspects, with a total score of 34. A higher score indicates better basic motor function [27]. Throughout the study period, all adverse events were closely monitored and recorded. Any unexpected injuries, newly emerged illnesses, or discomfort were documented in detail in the “Adverse Event Record Form.” This documentation included the time of occurrence, symptoms, duration, management measures taken, and an assessment of the causality between the event and the study intervention. Adverse events were categorized and recorded by severity. Protocol amendments regarding outcome measures The initial study protocol (approved August 2023, Approval No. 2023KY-020-02) included the TUGT, BBS, and FMA-LE as the primary clinical functional outcomes. In May 2024, before participant enrollment for fNIRS data collection, the DGI was added as the primary outcome to better capture gait adaptability, as it is specifically designed to assess walking performance in complex environments and has been widely used for this purpose in stroke populations [24]. This amendment was approved by the ethics committee in June 2024 (Approval No. 2024YJS-007-02). No other changes were made to the outcome measures after trial commencement. fNIRS data acquisition and analysis fNIRS data were acquired using a 40-channel LABNIRS system. The system employed integrated L-shaped fiber-optic probe holders with alternating arrangements of light sources and detectors, with a source-detector separation of 3 cm and a sampling frequency of 10.101 Hz. This setup enabled synchronous recording of relative concentration changes in oxygenated hemoglobin (HbO₂), deoxygenated hemoglobin (HbR), and total hemoglobin. Data collection was conducted in a sound-attenuated and dimly lit laboratory. To optimize signal quality, a black cap was used to secure the probe holder and minimize interference from ambient light. Hair was parted using a comb before probe placement to ensure optimal optode-scalp contact. Before acquisition commenced, participants received detailed instructions on the procedure and were asked to minimize head movements during task execution. The total acquisition protocol lasted approximately 25 min and proceeded in the following order: First, a 5-minute resting-state block was recorded, during which participants sat quietly and fixated on a central crosshair on a screen. This was followed by three task-state blocks, separated by 5-minute breaks. Task instructions were presented using E-Prime software in a block design alternating between “task” and “rest” periods. The stepping-in-place task required participants to march in place for 20 s upon cue, followed by a 15-second rest period, repeated for 5 cycles. The obstacle-crossing task required participants to step over a 15-cm-high physical obstacle with either their affected or unaffected foot, according to on-screen prompts, for 20 s, followed by a 15-second rest, alternating sides for a total of 10 cycles (5 cycles per side). The sit-to-stand task required participants to repeatedly rise from and sit down on a standard-height chair for 20 s, followed by a 15-second rest, repeated for 5 cycles. Details are shown in Fig. 2. Data processing and analysis were performed in MATLAB, primarily using the NIRS_KIT and NIRS_SPM toolboxes. Raw light-intensity signals were first converted to optical density and then to a time series of hemoglobin concentration changes using the modified Beer-Lambert law, with a differential pathlength factor set to 6.0. Two independent analysts, blinded to group allocation, performed quality control on the raw data to identify and mark channels and time segments containing significant noise or motion artifacts. The preprocessing pipeline included: temporal alignment of data segments based on task markers; detrending using a polynomial regression model coupled with motion artifact correction via the temporal derivative distribution repair method; bandpass filtering (0.01–0.08 Hz) to suppress low-frequency drift and high-frequency noise; and finally, quantitative signal quality index-based assessment of each channel, leading to the exclusion of channels that remained substandard (e.g., SQI < 2) after correction. Analytical metrics were extracted from the preprocessed data. For task-state data, cortical activation was estimated using a general linear model-based statistical parametric mapping approach. The HbO₂ time series from each channel was modeled against the task regressor convolved with a hemodynamic response function, while controlling for covariates such as age and sex. The resulting beta coefficient (β) for each channel served as an indicator of its task-related activation strength. The concentration of HbO₂ was used as the primary hemodynamic metric for inference in this study. For resting-state data, two analytical approaches were employed. First, functional connectivity was calculated as the Pearson correlation coefficient between the HbO₂ time series from different region-of-interest (ROI) pairs, which is the standard metric for resting-state fNIRS analysis. Second, as an exploratory analysis, we also examined the mean HbO₂ concentration during the stable period for each channel. After discarding the initial 30 s, the mean HbO₂ concentration was calculated for each channel. We acknowledge that this metric is susceptible to systemic physiological noise; therefore, these exploratory findings should be interpreted with caution. Fig. 2Schematic diagrams of the fNIRS setup. A Channel layout schematic 1 B Channel layout schematic 2 C fNIRS data acquisition protocol Schematic diagrams of the fNIRS setup. A Channel layout schematic 1 B Channel layout schematic 2 C fNIRS data acquisition protocol Statistical analysis Statistical analyses were performed using SPSS (version 26.0) with α = 0.05 (two-tailed). Baseline demographic and clinical characteristics were compared between groups using independent samples t-tests, Mann-Whitney U tests, or chi-square tests, as appropriate based on data distribution, to assess the balance achieved by randomization. Analyses of primary and secondary clinical outcomes were based on the per-protocol set. Between-group comparisons after the intervention were conducted using analysis of covariance (ANCOVA), with baseline scores as covariates. If the data violated the assumptions for parametric tests, nonparametric tests would be used to compare within-subject change scores (post-intervention minus baseline) between groups. For fNIRS data, analyses were conducted on available data after quality control. Participants with insufficient data quality (e.g., channels with signal quality index < 2 after preprocessing, excessive motion artifacts, or incomplete task recordings) were excluded from the corresponding analyses on a case-by-case basis. The exact sample sizes for each fNIRS analysis (e.g., stepping task, obstacle-crossing tasks, resting-state) are reported in the respective tables and figure legends. No imputation was applied to missing fNIRS data. Differences in task-evoked brain activation strength were assessed using independent samples t-tests to compare the extracted β-values from predefined regions of interest under specific task conditions between the VR and control groups. For resting-state data, two sets of analyses were conducted. First, functional connectivity values (Fisher-transformed Pearson correlation coefficients) were compared between groups using independent samples t-tests. Second, as an exploratory analysis, the mean HbO₂ concentration within predefined brain regions during the resting state was also compared between groups using independent samples t-tests. Given that the mean HbO₂ concentration is not a standard resting-state metric and may be influenced by systemic physiological factors, these results are considered exploratory. No correction for multiple comparisons was applied to the analyses of fNIRS outcomes, and these findings should be regarded as preliminary. To examine the relationship between neural changes and clinical functional improvements, partial correlation analyses were performed between changes in brain activation (β-values) or functional connectivity (Fisher-transformed correlation coefficients) and changes in clinical outcomes (DGI, TUGT, BBS) that showed significant between-group differences. These analyses were controlled for potential confounders, including age, gender, years of education, BMI, time since stroke onset, and hemiplegic side. No subgroup analyses were conducted. To assess the potential impact of dropout, we first compared baseline characteristics between completers (n = 56) and dropouts (n = 12) using independent t-tests, Mann-Whitney U tests, or chi-square tests, as appropriate. To evaluate the robustness of our findings, we performed two sensitivity analyses using extreme-case scenarios. In the worst-case scenario, participants who dropped out of the VR group were assigned the poorest observed endpoint value among completers, while those in the control group were assigned the best observed endpoint value. In the best-case scenario, the opposite assignment was applied. The analyses were then repeated, including all 68 randomized participants. Results Participant flow A total of 629 patients with stroke were screened for eligibility from June 2024 to January 2025. Based on the inclusion and exclusion criteria, 561 individuals were excluded: 505 did not meet the inclusion criteria, 22 met one or more exclusion criteria, and 34 declined to participate. Consequently, 68 eligible patients were enrolled and randomly assigned to the VR group (n = 34) or the control group (n = 34). After the 4-week intervention, 6 participants in the control group dropped out (3 transferred to local hospitals for further treatment and were lost to follow-up, 2 transferred due to family reasons and were lost to follow-up, and 1 due to worsening medical condition). Ultimately, 28 participants in the control group completed post-intervention clinical functional assessments, and 26 completed fNIRS assessments (2 lost to follow-up due to hospital transfer). In the VR group, 6 participants dropped out (4 transferred to local hospitals for further treatment and were lost to follow-up, 2 transferred due to family reasons and were lost to follow-up). Ultimately, 28 participants in the VR group completed post-intervention clinical functional assessments, and 28 completed fNIRS assessments. The trial ended after the target sample size was achieved. Detailed participant flow, including specific reasons for ineligibility, is shown in Fig. 3. Fig. 3Study CONSORT flow diagram Study CONSORT flow diagram Baseline characteristics Before the intervention, general demographic data, clinical characteristics, and gait/balance function scores were compared between the two groups. The results showed no statistically significant differences between the groups in general factors, including age, gender, body mass index, years of education, time since stroke onset, stroke type, or hemiplegic side (all P > 0.05). There were also no statistically significant differences in pre-intervention scores for DGI, TUGT, BBS, or FMA-LE(all P > 0.05). Specific data are detailed in Table 1. These results indicate that the baseline characteristics of the two groups were balanced and comparable. Attrition analysis Comparison of baseline characteristics between completers (n = 56) and dropouts (n = 12) revealed no significant differences in demographic and clinical variables, including age, gender, BMI, education, time since stroke, stroke type, hemiplegic side, and FAC grade (all P > 0.05). However, dropouts had significantly lower baseline DGI scores compared to completers (P < 0.001), indicating that participants with poorer baseline gait adaptability were more likely to drop out. No significant differences were found in baseline TUGT, BBS, or FMA-LE scores (all P > 0.05). Detailed data are presented in Supplementary Table 1. Comparison of clinical outcomes Before the intervention, no statistically significant differences were observed between the two groups in any of the outcome measures, confirming baseline comparability. Specifically, scores for DGI, TUGT, BBS, and FMA-LE were comparable between groups (all P > 0.05). Detailed data are presented in Table 2. Following the 4-week intervention, between-group comparisons of the change scores (post-intervention minus baseline) revealed significant differences favoring the VR group in several domains. For gait adaptability (DGI), the VR group demonstrated significantly greater improvement in the total score (P = 0.010) and in four specific sub-items: walking with vertical head turns (P = 0.022), walking while performing a cognitive task (P = 0.030), gait initiation (P = 0.021), and stepping over obstacles. Regarding functional mobility (TUGT), the reduction in completion time was significantly greater in the VR group compared to the control group (P = 0.005), indicating a more pronounced improvement in safe walking speed. For balance function (BBS), the VR group showed a significantly larger increase in the total score (P < 0.001) and in nine sub-items: sitting without back support (P = 0.045), transfers (P = 0.039), standing with eyes closed (P = 0.046), standing with feet together (P = 0.001), retrieving an object from the floor (P = 0.006), turning to look behind (P = 0.001), turning 360 degrees (P = 0.001), alternating stepping on a stool (P = 0.014), and standing on one leg (P = 0.011). In contrast, for lower limb motor function (FMA-LE), the between-group difference in change scores did not reach statistical significance (P > 0.05), although the VR group showed numerically greater mean improvement (Table 2; Fig. 4). These findings should be interpreted with caution due to the absence of correction for multiple comparisons. Fig. 4Clinical functional outcome measures. (A) Between-group comparison of DGI change scores (post-intervention minus pre-intervention). (B) Between-group comparison of TUGT change scores (post-intervention minus pre-intervention). (C) Between-group comparison of BBS change scores (post-intervention minus pre-intervention). (D) Between-group comparison of FMA-LE change scores (post-intervention minus pre-intervention). Asterisks indicate significant between-group differences (*P < 0.05). DGI Dynamic Gait Index, TUGT Timed Up and Go Test, BBS Berg Balance Scale, FMA-LE Fugl-Meyer Assessment for Lower Extremity Clinical functional outcome measures. (A) Between-group comparison of DGI change scores (post-intervention minus pre-intervention). (B) Between-group comparison of TUGT change scores (post-intervention minus pre-intervention). (C) Between-group comparison of BBS change scores (post-intervention minus pre-intervention). (D) Between-group comparison of FMA-LE change scores (post-intervention minus pre-intervention). Asterisks indicate significant between-group differences (*P < 0.05). DGI Dynamic Gait Index, TUGT Timed Up and Go Test, BBS Berg Balance Scale, FMA-LE Fugl-Meyer Assessment for Lower Extremity Sensitivity analysis To assess the robustness of our primary findings under the most conservative assumption regarding missing data, we performed a worst-case scenario analysis for all four primary outcomes using change scores (post-intervention minus baseline). Detailed results are presented in Supplementary Table 2. For DGI, TUGT, and BBS, the significant improvements observed in the VR group in the PP analysis (all P < 0.05) were no longer significant in the worst-case scenario (all P > 0.05). For FMA-LE, the non-significant difference in the PP analysis (P = 0.221) remained non-significant in the worst-case scenario (P = 0.193). These results suggest that the significant effects on DGI, TUGT, and BBS were not sustained under the most conservative assumptions about missing data and should be interpreted with caution. Comparison of fNIRS outcomes Brain activation changes (β-values) during task Stepping-in-place task Following the intervention, the between-group comparison of changes in activation (β-values) during the stepping-in-place task showed that the VR group exhibited a significantly greater increase in activation compared to the control group in the right PPC (P = 0.029) and the right SMA (P = 0.032). No significant between-group differences were found in the remaining brain regions (Table 3; Fig. 5). Table 1Comparison of general demographic data between the two groups (x̄̄ ± S / (X̄ ± S/M(P25,P75)/n༉ItemsControl Group (n = 28)VR Group (n = 28)t/Z/χ² P Age (years)55.93 ± 9.4555.82 ± 11.550.0380.970Gender (Male/Female)*23 / 523 / 5< 0.001> 0.999BMI (kg/m²) #23.23 (21.02, 25.36)21.86 (20.63, 23.22)1.1800.238Years of Education #9 (6, 15)12 (6.75, 14.25)0.1680.867Time since stroke onset (days)87.46 ± 59.4465.36 ± 52.741.4720.147Stroke Type (Hemorrhage/Infarction)*11 / 1713 / 150.2920.589Hemiplegic Side (Left/Right)*11 / 1710 / 180.2710.787Functional Ambulation Category (FAC) Grade*3 (3, 3)3 (3, 3)0.1620.872Data not conforming to a normal distribution are presented as Median (P25, P75) and analyzed using the Mann-Whitney U test*Categorical data are presented as counts (n) and analyzed using the Chi-square test Table 1Comparison of general demographic data between the two groups (x̄̄ ± S / (X̄ ± S/M(P25,P75)/n༉ItemsControl Group (n = 28)VR Group (n = 28)t/Z/χ² P Age (years)55.93 ± 9.4555.82 ± 11.550.0380.970Gender (Male/Female)*23 / 523 / 5< 0.001> 0.999BMI (kg/m²) #23.23 (21.02, 25.36)21.86 (20.63, 23.22)1.1800.238Years of Education #9 (6, 15)12 (6.75, 14.25)0.1680.867Time since stroke onset (days)87.46 ± 59.4465.36 ± 52.741.4720.147Stroke Type (Hemorrhage/Infarction)*11 / 1713 / 150.2920.589Hemiplegic Side (Left/Right)*11 / 1710 / 180.2710.787Functional Ambulation Category (FAC) Grade*3 (3, 3)3 (3, 3)0.1620.872Data not conforming to a normal distribution are presented as Median (P25, P75) and analyzed using the Mann-Whitney U test*Categorical data are presented as counts (n) and analyzed using the Chi-square test Comparison of general demographic data between the two groups (x̄̄ ± S / (X̄ ± S/M(P25,P75)/n༉ Data not conforming to a normal distribution are presented as Median (P25, P75) and analyzed using the Mann-Whitney U test *Categorical data are presented as counts (n) and analyzed using the Chi-square test Table 2Comparison of DGI scores between the two groups before and after intervention (X̄ ± S / M (P₂₅, P₇₅))Outcome measuresGroupBaseline4 Weeks post-interventionBaseline between-group P-valueBetween-group P-value (Change) DGI DGI-1 Level walking #VR Group (n = 28)2.00 (2.00, 3.00)3 (3, 3)0.5080.243Control Group (n = 28)2.00 (1.50, 3.00)3 (2, 3)DGI-2 Walking with horizontal head turns #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 3)0.4900.901Control Group (n = 28)2.00 (1.00, 2.00)2 (2, 3)DGI-3 Walking with vertical head turns #VR Group (n = 28)1.00 (1.00, 2.00)2 (2, 2.75)0.523 0.022 Control Group (n = 28)1.00 (1.00, 2.00)2 (1, 2)DGI-4 Walking while performing a cognitive task #VR Group (n = 28)1.50 (1.00, 2.00)2 (2, 2)0.971 0.030 Control Group (n = 28)1.00 (1.00, 2.00)2 (1, 2)DGI-5 Walking and stopping suddenly #VR Group (n = 28)1.00 (1.00, 2.00)2 (2, 2)0.2450.065Control Group (n = 28)1.00 (1.00, 1.00)1 (1, 2)DGI-6 Walking and starting suddenly #VR Group (n = 28)1.00 (1.00, 2.00)2 (1, 2)0.191 0.021 Control Group (n = 28)1.00 (1.00, 1.50)1 (1, 2)DGI-7 Walking and stepping over an obstacle #VR Group (n = 28)1.00 (1.50, 2.00)2 (2, 2)0.823 0.022 Control Group (n = 28)1.00 (1.00, 2.00)1.5 (1, 2)DGI-8 Walking and changing direction #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 3)0.9420.119Control Group (n = 28)2.00 (1.00, 2.00)2 (1.25, 2.75)DGI Total ScoreVR Group (n = 28)12.46 ± 4.0017.6429 ± 2.959160.643 0.010 Control Group (n = 28)11.96 ± 4.0314.9286 ± 4.01782 TUGT TUGT Total Score #VR Group (n = 28)21.77 (16.57, 30.31)17.6429 ± 2.959160.876 0.015 Control Group (n = 28)22.73 (15.71, 34.53)14.9286 ± 4.01782 BBS Sitting to standing #VR Group (n = 28)4.00 (3.00, 4.00)4 (4, 4)0.9410.103Control Group (n = 28)4.00 (3.00, 4.00)4 (3, 4)Standing unsupported #VR Group (n = 28)4.00 (3.50, 4.00)4 (4, 4)0.6570.211Control Group (n = 28)4.00 (3.00, 4.00)4 (3.25, 4)Sitting with back unsupported #VR Group (n = 28)4.00 (4.00, 4.00)4 (4, 4)0.174 0.045 Control Group (n = 28)4.00 (4.00, 4.00)4 (3.25, 4)Standing to sitting #VR Group (n = 28)3.00 (3.00, 4.00)4 (3.25, 4)0.4170.087Control Group (n = 28)4.00 (3.00, 4.00)4 (3, 4)Transfers #VR Group (n = 28)4.00 (3.00, 4.00)4 (4, 4)0.310 0.039 Control Group (n = 28)4.00 (3.00, 4.00)3 (3, 4)Standing with eyes closed #VR Group (n = 28)3.00 (3.00, 4.00)4 (3, 4)0.944 0.046 Control Group (n = 28)3.00 (2.50, 4.00)3 (2.25, 4)Standing with feet together #VR Group (n = 28)3.00 (3.00, 3.50)4 (3, 4)0.746 0.001 Control Group (n = 28)3.00 (2.00, 4.00)3 (2.25, 4)Reaching forward with outstretched arm #VR Group (n = 28)3.00 (3.00, 3.00)3 (3, 4)0.3450.093Control Group (n = 28)3.00 (2.00, 3.00)3 (2, 3)Retrieving object from floor #VR Group (n = 28)3.00 (2.50, 3.00)4 (3, 4)0.726 0.006 Control Group (n = 28)3.00 (2.00, 3.00)3 (2, 4)Turning to look behind #VR Group (n = 28)3.00 (3.00, 4.00)4 (4, 4)0.979 0.001 Control Group (n = 28)3.00 (2.00, 4.00)3 (2, 4)Turning 360 degrees #VR Group (n = 28)3.00 (2.00, 3.00)4 (3, 4)0.931 0.001 Control Group (n = 28)3.00 (2.00, 4.00)3 (2, 3.75)Placing alternate foot on stool #VR Group (n = 28)3.00 (2.00, 3.00)4 (3, 4)0.610 0.014 Control Group (n = 28)3.00 (1.00, 4.00)3 (1, 4)Standing with one foot in front #VR Group (n = 28)3.00 (2.00, 3.00)3 (3, 4)0.5580.059Control Group (n = 28)3.00 (2.00, 3.00)3 (1.25, 3)Standing on one leg #VR Group (n = 28)3.00 (2.00, 3.00)3 (3, 3)0.061 0.011 Control Group (n = 28)3.00 (1.00, 3.00)2 (1, 3)BBS Total Score #VR Group (n = 28)44.50 (40.00, 48.50)52 (47.5, 54)0.883 < 0.001 Control Group (n = 28)45.00 (35.00, 52.00)44 (33.5, 51.5) FMA-LE Reflex activity of flexors #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)> 0.999> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Reflex activity of extensors #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)> 0.999> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Flexion synergy: Hip #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)> 0.999> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Flexion synergy: Knee #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.642> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Flexion synergy: Ankle #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 2)0.9170.317Control Group (n = 28)2.00 (1.00, 2.00)2 (1.25, 2)Extension synergy: Hip #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.6420.418Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Extension synergy: Hip adduction #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.6420.304Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Extension synergy: Knee #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.6420.427Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Extension synergy: Ankle plantarflexion #VR Group (n = 28)2.00 (1.50, 2.00)2 (2, 2)0.2230.434Control Group (n = 28)2.00 (1.00, 2.00)2 (1, 2)Voluntary movement within flexion synergy: Knee flexion #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 2)0.8380.583Control Group (n = 28)2.00 (1.00, 2.00)2 (1, 2)Voluntary movement within flexion synergy: Ankle dorsiflexion #VR Group (n = 28)1.00 (0.00, 2.00)2 (1, 2)0.6730.993Control Group (n = 28)1.00 (0.00, 2.00)2 (1, 2)Voluntary movement with little synergy: Knee flexion to 90° #VR Group (n = 28)1.00 (0.00, 1.00)1 (1, 1.75)0.4560.467Control Group (n = 28)1.00 (0.00, 2.00)1.5 (0.25, 2)Voluntary movement with little synergy: Ankle dorsiflexion #VR Group (n = 28)0.00 (0.00, 1.00)1 (0.25, 1)0.5720.462Control Group (n = 28)0.00 (0.00, 1.00)1 (0, 1)Normal reflex activity #VR Group (n = 28)0.00 (0.00, 0.00)0 (0, 0)0.7310.971Control Group (n = 28)0.00 (0.00, 0.50)0 (0, 0)Coordination/Speed: Tremor #VR Group (n = 28)0.00 (0.00, 1.00)0.5 (0, 2)0.1840.356Control Group (n = 28)1.00 (0.00, 1.00)1 (0, 2)Coordination/Speed: Dysmetria #VR Group (n = 28)0.00 (0.00, 1.00)0.5 (0, 2)0.2350.343Control Group (n = 28)1.00 (0.00, 1.00)1 (0, 2)Coordination/Speed: Time #VR Group (n = 28)0.00 (0.00, 0.50)0 (0, 1)0.0740.807Control Group (n = 28)0.50 (0.00, 1.00)1 (0, 1)FMA-LE Total ScoreVR Group (n = 28)22.36 ± 5.3325.00 ± 5.312460.7360.221Control Group (n = 28)22.89 ± 6.4424.8571 ± 5.85450#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 2Comparison of DGI scores between the two groups before and after intervention (X̄ ± S / M (P₂₅, P₇₅))Outcome measuresGroupBaseline4 Weeks post-interventionBaseline between-group P-valueBetween-group P-value (Change) DGI DGI-1 Level walking #VR Group (n = 28)2.00 (2.00, 3.00)3 (3, 3)0.5080.243Control Group (n = 28)2.00 (1.50, 3.00)3 (2, 3)DGI-2 Walking with horizontal head turns #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 3)0.4900.901Control Group (n = 28)2.00 (1.00, 2.00)2 (2, 3)DGI-3 Walking with vertical head turns #VR Group (n = 28)1.00 (1.00, 2.00)2 (2, 2.75)0.523 0.022 Control Group (n = 28)1.00 (1.00, 2.00)2 (1, 2)DGI-4 Walking while performing a cognitive task #VR Group (n = 28)1.50 (1.00, 2.00)2 (2, 2)0.971 0.030 Control Group (n = 28)1.00 (1.00, 2.00)2 (1, 2)DGI-5 Walking and stopping suddenly #VR Group (n = 28)1.00 (1.00, 2.00)2 (2, 2)0.2450.065Control Group (n = 28)1.00 (1.00, 1.00)1 (1, 2)DGI-6 Walking and starting suddenly #VR Group (n = 28)1.00 (1.00, 2.00)2 (1, 2)0.191 0.021 Control Group (n = 28)1.00 (1.00, 1.50)1 (1, 2)DGI-7 Walking and stepping over an obstacle #VR Group (n = 28)1.00 (1.50, 2.00)2 (2, 2)0.823 0.022 Control Group (n = 28)1.00 (1.00, 2.00)1.5 (1, 2)DGI-8 Walking and changing direction #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 3)0.9420.119Control Group (n = 28)2.00 (1.00, 2.00)2 (1.25, 2.75)DGI Total ScoreVR Group (n = 28)12.46 ± 4.0017.6429 ± 2.959160.643 0.010 Control Group (n = 28)11.96 ± 4.0314.9286 ± 4.01782 TUGT TUGT Total Score #VR Group (n = 28)21.77 (16.57, 30.31)17.6429 ± 2.959160.876 0.015 Control Group (n = 28)22.73 (15.71, 34.53)14.9286 ± 4.01782 BBS Sitting to standing #VR Group (n = 28)4.00 (3.00, 4.00)4 (4, 4)0.9410.103Control Group (n = 28)4.00 (3.00, 4.00)4 (3, 4)Standing unsupported #VR Group (n = 28)4.00 (3.50, 4.00)4 (4, 4)0.6570.211Control Group (n = 28)4.00 (3.00, 4.00)4 (3.25, 4)Sitting with back unsupported #VR Group (n = 28)4.00 (4.00, 4.00)4 (4, 4)0.174 0.045 Control Group (n = 28)4.00 (4.00, 4.00)4 (3.25, 4)Standing to sitting #VR Group (n = 28)3.00 (3.00, 4.00)4 (3.25, 4)0.4170.087Control Group (n = 28)4.00 (3.00, 4.00)4 (3, 4)Transfers #VR Group (n = 28)4.00 (3.00, 4.00)4 (4, 4)0.310 0.039 Control Group (n = 28)4.00 (3.00, 4.00)3 (3, 4)Standing with eyes closed #VR Group (n = 28)3.00 (3.00, 4.00)4 (3, 4)0.944 0.046 Control Group (n = 28)3.00 (2.50, 4.00)3 (2.25, 4)Standing with feet together #VR Group (n = 28)3.00 (3.00, 3.50)4 (3, 4)0.746 0.001 Control Group (n = 28)3.00 (2.00, 4.00)3 (2.25, 4)Reaching forward with outstretched arm #VR Group (n = 28)3.00 (3.00, 3.00)3 (3, 4)0.3450.093Control Group (n = 28)3.00 (2.00, 3.00)3 (2, 3)Retrieving object from floor #VR Group (n = 28)3.00 (2.50, 3.00)4 (3, 4)0.726 0.006 Control Group (n = 28)3.00 (2.00, 3.00)3 (2, 4)Turning to look behind #VR Group (n = 28)3.00 (3.00, 4.00)4 (4, 4)0.979 0.001 Control Group (n = 28)3.00 (2.00, 4.00)3 (2, 4)Turning 360 degrees #VR Group (n = 28)3.00 (2.00, 3.00)4 (3, 4)0.931 0.001 Control Group (n = 28)3.00 (2.00, 4.00)3 (2, 3.75)Placing alternate foot on stool #VR Group (n = 28)3.00 (2.00, 3.00)4 (3, 4)0.610 0.014 Control Group (n = 28)3.00 (1.00, 4.00)3 (1, 4)Standing with one foot in front #VR Group (n = 28)3.00 (2.00, 3.00)3 (3, 4)0.5580.059Control Group (n = 28)3.00 (2.00, 3.00)3 (1.25, 3)Standing on one leg #VR Group (n = 28)3.00 (2.00, 3.00)3 (3, 3)0.061 0.011 Control Group (n = 28)3.00 (1.00, 3.00)2 (1, 3)BBS Total Score #VR Group (n = 28)44.50 (40.00, 48.50)52 (47.5, 54)0.883 < 0.001 Control Group (n = 28)45.00 (35.00, 52.00)44 (33.5, 51.5) FMA-LE Reflex activity of flexors #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)> 0.999> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Reflex activity of extensors #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)> 0.999> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Flexion synergy: Hip #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)> 0.999> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Flexion synergy: Knee #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.642> 0.999Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Flexion synergy: Ankle #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 2)0.9170.317Control Group (n = 28)2.00 (1.00, 2.00)2 (1.25, 2)Extension synergy: Hip #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.6420.418Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Extension synergy: Hip adduction #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.6420.304Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Extension synergy: Knee #VR Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)0.6420.427Control Group (n = 28)2.00 (2.00, 2.00)2 (2, 2)Extension synergy: Ankle plantarflexion #VR Group (n = 28)2.00 (1.50, 2.00)2 (2, 2)0.2230.434Control Group (n = 28)2.00 (1.00, 2.00)2 (1, 2)Voluntary movement within flexion synergy: Knee flexion #VR Group (n = 28)2.00 (1.00, 2.00)2 (2, 2)0.8380.583Control Group (n = 28)2.00 (1.00, 2.00)2 (1, 2)Voluntary movement within flexion synergy: Ankle dorsiflexion #VR Group (n = 28)1.00 (0.00, 2.00)2 (1, 2)0.6730.993Control Group (n = 28)1.00 (0.00, 2.00)2 (1, 2)Voluntary movement with little synergy: Knee flexion to 90° #VR Group (n = 28)1.00 (0.00, 1.00)1 (1, 1.75)0.4560.467Control Group (n = 28)1.00 (0.00, 2.00)1.5 (0.25, 2)Voluntary movement with little synergy: Ankle dorsiflexion #VR Group (n = 28)0.00 (0.00, 1.00)1 (0.25, 1)0.5720.462Control Group (n = 28)0.00 (0.00, 1.00)1 (0, 1)Normal reflex activity #VR Group (n = 28)0.00 (0.00, 0.00)0 (0, 0)0.7310.971Control Group (n = 28)0.00 (0.00, 0.50)0 (0, 0)Coordination/Speed: Tremor #VR Group (n = 28)0.00 (0.00, 1.00)0.5 (0, 2)0.1840.356Control Group (n = 28)1.00 (0.00, 1.00)1 (0, 2)Coordination/Speed: Dysmetria #VR Group (n = 28)0.00 (0.00, 1.00)0.5 (0, 2)0.2350.343Control Group (n = 28)1.00 (0.00, 1.00)1 (0, 2)Coordination/Speed: Time #VR Group (n = 28)0.00 (0.00, 0.50)0 (0, 1)0.0740.807Control Group (n = 28)0.50 (0.00, 1.00)1 (0, 1)FMA-LE Total ScoreVR Group (n = 28)22.36 ± 5.3325.00 ± 5.312460.7360.221Control Group (n = 28)22.89 ± 6.4424.8571 ± 5.85450#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Comparison of DGI scores between the two groups before and after intervention (X̄ ± S / M (P₂₅, P₇₅)) #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 3Between-group comparison of changes in brain activation (β-values) during the stepping-in-place task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #− 0.002907 (− 0.011090, 0.004348)0.001106 (− 0.005815, 0.005011)− 1.1600.246Right S1 #0.001855 (− 0.005507, 0.006188)0.000526 (− 0.002799, 0.005545)0.1730.863Left M1 #0.000036 (− 0.007237, 0.006318)0.001684 (− 0.009250, 0.005612)− 0.0170.986Right M1 #0.003093 (− 0.001903, 0.005109)0.000928 (− 0.004568, 0.004401)1.2980.194Left PPC #0.003245 (− 0.002006, 0.005293)0.001793 (− 0.003362, 0.006407)0.2250.822Right PPC #0.005190 (− 0.003090, 0.012114)− 0.003734 (− 0.012162, 0.005765)2.181 0.029 LeftSMA #0.001578 (− 0.001293, 0.008837)0.000372 (− 0.001784, 0.005336)− 1.0040.315RightSMA #0.003629 (0.001268, 0.011338)0.000741 (− 0.006748, 0.003712)2.147 0.032 Left PMC#0.002170 (− 0.001033, 0.008040)− 0.000532 (− 0.008286, 0.005506)1.7830.075Right PMC#0.001971 (− 0.005997, 0.004285)0.001544 (− 0.003951, 0.010058)0.4670.640Left aPFC #0.000853 (− 0.003036, 0.003416)− 0.001102 (− 0.002580, 0.001762)1.1770.239Right aPFC #0.003687 (− 0.001136, 0.006159)0.001173 (− 0.001758, 0.005577)0.8310.406LeftDLPFC#0.002933 (− 0.000505, 0.006469)0.001035 (− 0.004364, 0.003419)1.2980.194RightDLPFC#0.000341 (− 0.005711, 0.004267)− 0.000789 (− 0.007400, 0.003876)0.4850.628#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 3Between-group comparison of changes in brain activation (β-values) during the stepping-in-place task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #− 0.002907 (− 0.011090, 0.004348)0.001106 (− 0.005815, 0.005011)− 1.1600.246Right S1 #0.001855 (− 0.005507, 0.006188)0.000526 (− 0.002799, 0.005545)0.1730.863Left M1 #0.000036 (− 0.007237, 0.006318)0.001684 (− 0.009250, 0.005612)− 0.0170.986Right M1 #0.003093 (− 0.001903, 0.005109)0.000928 (− 0.004568, 0.004401)1.2980.194Left PPC #0.003245 (− 0.002006, 0.005293)0.001793 (− 0.003362, 0.006407)0.2250.822Right PPC #0.005190 (− 0.003090, 0.012114)− 0.003734 (− 0.012162, 0.005765)2.181 0.029 LeftSMA #0.001578 (− 0.001293, 0.008837)0.000372 (− 0.001784, 0.005336)− 1.0040.315RightSMA #0.003629 (0.001268, 0.011338)0.000741 (− 0.006748, 0.003712)2.147 0.032 Left PMC#0.002170 (− 0.001033, 0.008040)− 0.000532 (− 0.008286, 0.005506)1.7830.075Right PMC#0.001971 (− 0.005997, 0.004285)0.001544 (− 0.003951, 0.010058)0.4670.640Left aPFC #0.000853 (− 0.003036, 0.003416)− 0.001102 (− 0.002580, 0.001762)1.1770.239Right aPFC #0.003687 (− 0.001136, 0.006159)0.001173 (− 0.001758, 0.005577)0.8310.406LeftDLPFC#0.002933 (− 0.000505, 0.006469)0.001035 (− 0.004364, 0.003419)1.2980.194RightDLPFC#0.000341 (− 0.005711, 0.004267)− 0.000789 (− 0.007400, 0.003876)0.4850.628#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Between-group comparison of changes in brain activation (β-values) during the stepping-in-place task (M (P₂₅, P₇₅)) #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 4Between-group comparison of changes in brain activation (β-values) during the affected-side obstacle-crossing task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #− 0.000930 (− 0.005652, 0.002812)− 0.005349 (− 0.012970, 0.000840)1.2290.219Right S1 #− 0.000216 (− 0.009587, 0.007860)0.000456 (− 0.004325, 0.005327)− 0.3810.703Left M1 #0.000558 (− 0.009634, 0.010255)0.000522 (− 0.007786, 0.003798)0.5020.616Right M1 #0.000686 (− 0.003463, 0.004204)− 0.000565 (− 0.005905, 0.003096)0.4850.628Left PPC #− 0.002048 (− 0.005873, 0.002647)− 0.000033 (− 0.005118, 0.005203)− 0.8660.387RightPPC #0.001038 (− 0.004294, 0.006539)0.002540 (− 0.006035, 0.006676)− 0.2770.782LeftSMA #− 0.000682 (− 0.005713, 0.004304)0.000902 (− 0.002318, 0.006780)− 1.3850.166RightSMA #0.002198 (− 0.002226, 0.007264)− 0.001834 (− 0.006016, 0.004408)1.4200.156Left PMC#0.003371 (− 0.003910, 0.008520)− 0.000189 (− 0.007508, 0.006835)0.8480.396Right PMC#0.001421 (− 0.001732, 0.004698)− 0.002435 (− 0.007606, 0.001040)2.285 0.022 Left aPFC #0.000477 (− 0.003478, 0.003921)0.000240 (− 0.005062, 0.002728)0.9520.341Right aPFC #0.000064 (− 0.002356, 0.002420)− 0.002345 (− 0.005387, 0.001989)1.5930.111LeftDLPFC#0.001200 (− 0.003783, 0.005676)− 0.000404 (− 0.007680, 0.006692)0.6750.500RightDLPFC#− 0.000061 (− 0.004034, 0.006399)− 0.000083 (− 0.006586, 0.006009)0.0170.986#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 4Between-group comparison of changes in brain activation (β-values) during the affected-side obstacle-crossing task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #− 0.000930 (− 0.005652, 0.002812)− 0.005349 (− 0.012970, 0.000840)1.2290.219Right S1 #− 0.000216 (− 0.009587, 0.007860)0.000456 (− 0.004325, 0.005327)− 0.3810.703Left M1 #0.000558 (− 0.009634, 0.010255)0.000522 (− 0.007786, 0.003798)0.5020.616Right M1 #0.000686 (− 0.003463, 0.004204)− 0.000565 (− 0.005905, 0.003096)0.4850.628Left PPC #− 0.002048 (− 0.005873, 0.002647)− 0.000033 (− 0.005118, 0.005203)− 0.8660.387RightPPC #0.001038 (− 0.004294, 0.006539)0.002540 (− 0.006035, 0.006676)− 0.2770.782LeftSMA #− 0.000682 (− 0.005713, 0.004304)0.000902 (− 0.002318, 0.006780)− 1.3850.166RightSMA #0.002198 (− 0.002226, 0.007264)− 0.001834 (− 0.006016, 0.004408)1.4200.156Left PMC#0.003371 (− 0.003910, 0.008520)− 0.000189 (− 0.007508, 0.006835)0.8480.396Right PMC#0.001421 (− 0.001732, 0.004698)− 0.002435 (− 0.007606, 0.001040)2.285 0.022 Left aPFC #0.000477 (− 0.003478, 0.003921)0.000240 (− 0.005062, 0.002728)0.9520.341Right aPFC #0.000064 (− 0.002356, 0.002420)− 0.002345 (− 0.005387, 0.001989)1.5930.111LeftDLPFC#0.001200 (− 0.003783, 0.005676)− 0.000404 (− 0.007680, 0.006692)0.6750.500RightDLPFC#− 0.000061 (− 0.004034, 0.006399)− 0.000083 (− 0.006586, 0.006009)0.0170.986#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Between-group comparison of changes in brain activation (β-values) during the affected-side obstacle-crossing task (M (P₂₅, P₇₅)) Right PPC # #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 5Between-group comparison of changes in brain activation (β-values) during the unaffected-side obstacle-crossing task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #0.002864 (− 0.004134, 0.007817)0.000416 (− 0.005505, 0.004008)1.8000.072Right S1 #0.000958 (− 0.004717, 0.006020)− 0.001410 (− 0.005950, 0.004227)0.6060.545Left M1 #− 0.001392 (− 0.014618, 0.004041)− 0.001433 (− 0.004950, 0.010129)− 1.2290.219Right M1 #0.000755 (− 0.003503, 0.004654)− 0.001433 (− 0.004950, 0.010129)− 0.2080.835LeftPPC #− 0.000607 (− 0.003198, 0.005243)0.001286 (− 0.004136, 0.004617)− 0.4850.628RightPPC #0.002870 (− 0.002004, 0.018127)0.000930 (− 0.007512, 0.007165)1.6790.093LeftSMA #− 0.001008 (− 0.011540, 0.005292)− 0.000208 (− 0.003974, 0.003036)− 0.8310.406RightSMA #0.003538 (− 0.006653, 0.011950)− 0.001547 (− 0.006172, 0.003004)1.974 0.048 Left PMC#0.002638 (− 0.002623, 0.006725)0.000088 (− 0.007849, 0.006532)1.0210.307Right PMC#0.002807 (− 0.000655, 0.007362)− 0.003559 (− 0.008883, 0.002356)3.289 0.001 Left aPFC #0.001553 (− 0.000979, 0.007367)− 0.000969 (− 0.007876, 0.004229)1.7310.083Right aPFC #0.002185 (− 0.002806, 0.004711)− 0.002981 (− 0.007577, 0.000147)3.082 0.002 LeftDLPFC#0.003200 (− 0.002005, 0.008646)− 0.002000 (− 0.004920, 0.004777)1.8350.066RightDLPFC#0.000971 (− 0.003859, 0.007108)− 0.001121 (− 0.009547, 0.005471)1.1950.232#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 5Between-group comparison of changes in brain activation (β-values) during the unaffected-side obstacle-crossing task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #0.002864 (− 0.004134, 0.007817)0.000416 (− 0.005505, 0.004008)1.8000.072Right S1 #0.000958 (− 0.004717, 0.006020)− 0.001410 (− 0.005950, 0.004227)0.6060.545Left M1 #− 0.001392 (− 0.014618, 0.004041)− 0.001433 (− 0.004950, 0.010129)− 1.2290.219Right M1 #0.000755 (− 0.003503, 0.004654)− 0.001433 (− 0.004950, 0.010129)− 0.2080.835LeftPPC #− 0.000607 (− 0.003198, 0.005243)0.001286 (− 0.004136, 0.004617)− 0.4850.628RightPPC #0.002870 (− 0.002004, 0.018127)0.000930 (− 0.007512, 0.007165)1.6790.093LeftSMA #− 0.001008 (− 0.011540, 0.005292)− 0.000208 (− 0.003974, 0.003036)− 0.8310.406RightSMA #0.003538 (− 0.006653, 0.011950)− 0.001547 (− 0.006172, 0.003004)1.974 0.048 Left PMC#0.002638 (− 0.002623, 0.006725)0.000088 (− 0.007849, 0.006532)1.0210.307Right PMC#0.002807 (− 0.000655, 0.007362)− 0.003559 (− 0.008883, 0.002356)3.289 0.001 Left aPFC #0.001553 (− 0.000979, 0.007367)− 0.000969 (− 0.007876, 0.004229)1.7310.083Right aPFC #0.002185 (− 0.002806, 0.004711)− 0.002981 (− 0.007577, 0.000147)3.082 0.002 LeftDLPFC#0.003200 (− 0.002005, 0.008646)− 0.002000 (− 0.004920, 0.004777)1.8350.066RightDLPFC#0.000971 (− 0.003859, 0.007108)− 0.001121 (− 0.009547, 0.005471)1.1950.232#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Between-group comparison of changes in brain activation (β-values) during the unaffected-side obstacle-crossing task (M (P₂₅, P₇₅)) Left PPC # Right PPC # #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 6Between-group comparison of changes in brain activation (β-values) during the sit-to-stand task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #0.002864 (− 0.004134, 0.007817)0.000416 (− 0.005505, 0.004008)1.9220.055Right S1 #0.000958 (− 0.004717, 0.006020)− 0.001410 (− 0.005950, 0.004227)0.6410.522Left M1 #− 0.000607 (− 0.003198, 0.005243)0.001286 (− 0.004136, 0.004617)− 0.9000.368Right M1 #0.002870 (− 0.002004, 0.018127)0.000930 (− 0.007512, 0.007165)1.1250.260LeftPPC #− 0.001008 (− 0.011540, 0.005292)− 0.000208 (− 0.003974, 0.003036)− 0.7270.467RightPPC #0.003538 (− 0.006653, 0.011950)− 0.001547 (− 0.006172, 0.003004)1.1950.232LeftSMA #0.002638 (− 0.002623, 0.006725)0.000088 (− 0.007849, 0.006532)1.2460.213RightSMA #0.002807 (− 0.000655, 0.007362)− 0.003559 (− 0.008883, 0.002356)1.1770.239Left PMC#0.001553 (− 0.000979, 0.007367)− 0.000969 (− 0.007876, 0.004229)0.5370.591Right PMC#0.002185 (− 0.002806, 0.004711)− 0.002981 (− 0.007577, 0.000147)0.4670.640Left aPFC #0.003200 (− 0.002005, 0.008646)− 0.002000 (− 0.004920, 0.004777)0.3290.742Right aPFC #0.000971 (− 0.003859, 0.007108)− 0.001121 (− 0.009547, 0.005471)0.3120.755LeftDLPFC#0.003200 (− 0.002005, 0.008646)− 0.002000 (− 0.004920, 0.004777)1.8350.066RightDLPFC#0.000971 (− 0.003859, 0.007108)− 0.001121 (− 0.009547, 0.005471)1.1950.232#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 6Between-group comparison of changes in brain activation (β-values) during the sit-to-stand task (M (P₂₅, P₇₅))Brain regionVR groupControl groupZ-valueP-valueLeft S1 #0.002864 (− 0.004134, 0.007817)0.000416 (− 0.005505, 0.004008)1.9220.055Right S1 #0.000958 (− 0.004717, 0.006020)− 0.001410 (− 0.005950, 0.004227)0.6410.522Left M1 #− 0.000607 (− 0.003198, 0.005243)0.001286 (− 0.004136, 0.004617)− 0.9000.368Right M1 #0.002870 (− 0.002004, 0.018127)0.000930 (− 0.007512, 0.007165)1.1250.260LeftPPC #− 0.001008 (− 0.011540, 0.005292)− 0.000208 (− 0.003974, 0.003036)− 0.7270.467RightPPC #0.003538 (− 0.006653, 0.011950)− 0.001547 (− 0.006172, 0.003004)1.1950.232LeftSMA #0.002638 (− 0.002623, 0.006725)0.000088 (− 0.007849, 0.006532)1.2460.213RightSMA #0.002807 (− 0.000655, 0.007362)− 0.003559 (− 0.008883, 0.002356)1.1770.239Left PMC#0.001553 (− 0.000979, 0.007367)− 0.000969 (− 0.007876, 0.004229)0.5370.591Right PMC#0.002185 (− 0.002806, 0.004711)− 0.002981 (− 0.007577, 0.000147)0.4670.640Left aPFC #0.003200 (− 0.002005, 0.008646)− 0.002000 (− 0.004920, 0.004777)0.3290.742Right aPFC #0.000971 (− 0.003859, 0.007108)− 0.001121 (− 0.009547, 0.005471)0.3120.755LeftDLPFC#0.003200 (− 0.002005, 0.008646)− 0.002000 (− 0.004920, 0.004777)1.8350.066RightDLPFC#0.000971 (− 0.003859, 0.007108)− 0.001121 (− 0.009547, 0.005471)1.1950.232#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Between-group comparison of changes in brain activation (β-values) during the sit-to-stand task (M (P₂₅, P₇₅)) Left PPC # Right PPC # #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 7Between-group comparison of changes in resting-state functional connectivity (M (P₂₅, P₇₅))Brain regionsVR groupControl groupZ-valueP-valueS1.L — M1.R #0.2426 (− 0.1094, 0.5015)− 0.0679 (− 0.3116, 0.1809)2.1470.032SMA.L — aPFC.R #0.1552 (− 0.0324, 0.4161)− 0.1127 (− 0.4567, 0.1288)2.7530.006PPC.R — DLPFC.L #0.1307 (− 0.1212, 0.3096)− 0.0861 (− 0.4774, 0.1583)1.9560.048#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 7Between-group comparison of changes in resting-state functional connectivity (M (P₂₅, P₇₅))Brain regionsVR groupControl groupZ-valueP-valueS1.L — M1.R #0.2426 (− 0.1094, 0.5015)− 0.0679 (− 0.3116, 0.1809)2.1470.032SMA.L — aPFC.R #0.1552 (− 0.0324, 0.4161)− 0.1127 (− 0.4567, 0.1288)2.7530.006PPC.R — DLPFC.L #0.1307 (− 0.1212, 0.3096)− 0.0861 (− 0.4774, 0.1583)1.9560.048#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Between-group comparison of changes in resting-state functional connectivity (M (P₂₅, P₇₅)) #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 8Between-group comparison of changes in resting-state HbO₂ concentration (M (P₂₅, P₇₅))Brain regionVR group (n = 28)Control group (n = 26)Z-valueP-valueLeft S1 #− 0.000189 (− 0.000696, 0.000771)− 0.000049 (− 0.000918, 0.000864)− 0.2770.782Right S1 #0.000154 (− 0.000284, 0.000500)0.000062 (− 0.000205, 0.000413)0.3120.755Left M1− 0.000164 (− 0.000888, 0.000343)0.000118 (− 0.000187, 0.001053)0.9000.368Right M10.000062 (− 0.000286, 0.000259)0.000020 (− 0.000185, 0.000245)0.0520.959LeftPPC #− 0.000070 (− 0.000336, 0.000420)− 0.000096 (− 0.000453, 0.000140)0.8310.406RightPPC #− 0.000094 (− 0.000830, 0.000323)− 0.000059 (− 0.001312, 0.000359)− 0.1380.890LeftSMA #0.000064 (− 0.000452, 0.000703)− 0.000054 (− 0.000649, 0.000135)0.8660.387RightSMA #− 0.000050 (− 0.001243, 0.000619)0.000007 (− 0.000499, 0.000592)− 0.5190.604Left PMC#0.000042 (− 0.000200, 0.001153)0.000112 (− 0.000944, 0.000368)− 0.7620.446Right PMC#− 0.000006 (− 0.000330, 0.000240)− 0.000159 (− 0.000575, 0.000036)1.1430.253Left aPFC #0.000037 (− 0.000250, 0.000278)0.000074 (− 0.000161, 0.000225)− 0.3120.755Right aPFC #− 0.000015 (− 0.000236, 0.000222)− 0.000043 (− 0.000285, 0.000208)0.3640.716LeftDLPFC#− 0.000089 (− 0.000521, 0.000479)0.000139 (− 0.000188, 0.000665)− 1.2460.213RightDLPFC#− 0.000086 (− 0.000407, 0.000311)− 0.000011 (− 0.000381, 0.000552)− 0.3640.716#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 8Between-group comparison of changes in resting-state HbO₂ concentration (M (P₂₅, P₇₅))Brain regionVR group (n = 28)Control group (n = 26)Z-valueP-valueLeft S1 #− 0.000189 (− 0.000696, 0.000771)− 0.000049 (− 0.000918, 0.000864)− 0.2770.782Right S1 #0.000154 (− 0.000284, 0.000500)0.000062 (− 0.000205, 0.000413)0.3120.755Left M1− 0.000164 (− 0.000888, 0.000343)0.000118 (− 0.000187, 0.001053)0.9000.368Right M10.000062 (− 0.000286, 0.000259)0.000020 (− 0.000185, 0.000245)0.0520.959LeftPPC #− 0.000070 (− 0.000336, 0.000420)− 0.000096 (− 0.000453, 0.000140)0.8310.406RightPPC #− 0.000094 (− 0.000830, 0.000323)− 0.000059 (− 0.001312, 0.000359)− 0.1380.890LeftSMA #0.000064 (− 0.000452, 0.000703)− 0.000054 (− 0.000649, 0.000135)0.8660.387RightSMA #− 0.000050 (− 0.001243, 0.000619)0.000007 (− 0.000499, 0.000592)− 0.5190.604Left PMC#0.000042 (− 0.000200, 0.001153)0.000112 (− 0.000944, 0.000368)− 0.7620.446Right PMC#− 0.000006 (− 0.000330, 0.000240)− 0.000159 (− 0.000575, 0.000036)1.1430.253Left aPFC #0.000037 (− 0.000250, 0.000278)0.000074 (− 0.000161, 0.000225)− 0.3120.755Right aPFC #− 0.000015 (− 0.000236, 0.000222)− 0.000043 (− 0.000285, 0.000208)0.3640.716LeftDLPFC#− 0.000089 (− 0.000521, 0.000479)0.000139 (− 0.000188, 0.000665)− 1.2460.213RightDLPFC#− 0.000086 (− 0.000407, 0.000311)− 0.000011 (− 0.000381, 0.000552)− 0.3640.716#: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Between-group comparison of changes in resting-state HbO₂ concentration (M (P₂₅, P₇₅)) Left PPC # Right PPC # #: Data not conforming to a normal distribution are presented as Median (P₂₅, P₇₅) and analyzed using the Mann-Whitney U test Table 9Correlations between changes in brain activation during different tasks and changes in walking function scores in the VR group (median (P₂₅, P₇₅))Brain activation (task)Scale r P RightPPC (during stepping)TUGT− 0.3730.049RightSMA (during unaffected-side obstacle crossing)TUGT0.5900.001 Table 9Correlations between changes in brain activation during different tasks and changes in walking function scores in the VR group (median (P₂₅, P₇₅))Brain activation (task)Scale r P RightPPC (during stepping)TUGT− 0.3730.049RightSMA (during unaffected-side obstacle crossing)TUGT0.5900.001 Correlations between changes in brain activation during different tasks and changes in walking function scores in the VR group (median (P₂₅, P₇₅)) Table 10Correlations between changes in functional connectivity and changes in walking function scores in the VR group (median (P₂₅, P₇₅))Functional connectivity (brain regions)Scale r P PPC.R—DLPFC.LDGI0.4030.041S1.L—M1.RDGI-0.3780.048 Table 10Correlations between changes in functional connectivity and changes in walking function scores in the VR group (median (P₂₅, P₇₅))Functional connectivity (brain regions)Scale r P PPC.R—DLPFC.LDGI0.4030.041S1.L—M1.RDGI-0.3780.048 Correlations between changes in functional connectivity and changes in walking function scores in the VR group (median (P₂₅, P₇₅)) Affected-limb obstacle-crossing task For the affected-limb obstacle-crossing task, the VR group demonstrated a significantly greater increase in activation in the right PMC than the control group (P = 0.022). Differences in other brain regions were not statistically significant (Table 4; Fig. 5). Unaffected-limb obstacle-crossing task During the unaffected-limb obstacle-crossing task, the VR group showed significantly greater increases in activation in the right SMA (P = 0.048), the right PMC(P = 0.001), and the right anterior prefrontal cortex(aPFC) (P = 0.002) compared to the control group. No other regions showed statistically significant between-group differences (Table 5; Fig. 5). Sit-to-stand task The between-group comparison of activation changes during the sit-to-stand task revealed no statistically significant differences in any of the analyzed SMN regions (Table 6). Fig. 5Brain activation maps. (A, B) Brain activation maps during the stepping-in-place task. (C, D) Brain activation maps during the affected-side obstacle-crossing task. (E, F) Brain activation maps during the unaffected-side obstacle-crossing task Brain activation maps. (A, B) Brain activation maps during the stepping-in-place task. (C, D) Brain activation maps during the affected-side obstacle-crossing task. (E, F) Brain activation maps during the unaffected-side obstacle-crossing task Resting-state functional connectivity changes within SMN After the 4-week intervention, the between-group comparison of changes in resting-state functional connectivity within SMN revealed that the VR group showed a significantly greater increase in connectivity compared to the control group for three specific region pairs: between the left primary somatosensory cortex(S1) and the right M1 (P = 0.032), between the left SMA and the right aPFC (P = 0.006), and between the right PPC and the left dorsolateral prefrontal cortex (DLPFC) (P = 0.048). No statistically significant between-group differences were found for the remaining connections (Table 7; Fig. 6). Fig. 6Functional connectivity within the SMN and correlation analyses. (A, B) Functional connectivity between brain regions within the SMN. (C, D) Correlation plots of right PPC activation during the stepping task and right SMA activation during unaffected-side obstacle crossing with TUGT scores. (E, F) Correlation plots of connectivity strength (S1.L-M1.R and PPC.R-DLPFC.L) with DGI scores Functional connectivity within the SMN and correlation analyses. (A, B) Functional connectivity between brain regions within the SMN. (C, D) Correlation plots of right PPC activation during the stepping task and right SMA activation during unaffected-side obstacle crossing with TUGT scores. (E, F) Correlation plots of connectivity strength (S1.L-M1.R and PPC.R-DLPFC.L) with DGI scores Exploratory analysis: resting-state HbO₂ concentration changes As an exploratory analysis, we also examined changes in mean resting-state HbO₂ concentration within SMN regions. After the 4-week intervention, between-group comparison revealed no statistically significant differences for any of the analyzed brain regions (see Table 8). Given that the mean HbO₂ concentration is susceptible to systemic physiological noise and is not a standard resting-state fNIRS metric, these exploratory findings should be interpreted with caution. Correlation analyses between brain function and walking performance Correlation between brain activation and walking function After controlling for age, gender, years of education, BMI, time since stroke onset, and hemiplegic side, partial correlation analyses were conducted between the change scores (post-minus-pre-intervention) in brain regions showing significant between-group activation differences during fNIRS tasks and the change scores in DGI, TUGT, and BBS, which also showed significant between-group differences. The results indicated that within the VR group, the change in activation (β-value) in the right PPC during the stepping-in-place task was negatively correlated with the change in TUGT score (r = − 0.373, P = 0.049). Conversely, the change in activation in the right SMA during the unaffected-limb obstacle-crossing task was positively correlated with the change in TUGT score (r = 0.590, P = 0.001) (Table 9; Fig. 6). Correlation between functional connectivity and walking function Similarly, after controlling for the same covariates, partial correlation analyses were performed between the change scores in resting-state functional connectivity for the region pairs showing significant between-group differences and the change scores in DGI, TUGT, and BBS. The results showed that within the VR group, the change in functional connectivity between the left PPC and the right DLPFC was positively correlated with the change in DGI score (r = 0.403, P = 0.041). In contrast, the change in connectivity between the left S1 and the right M1 was negatively correlated with the change in DGI score (r = − 0.378, P = 0.048) (Table 10; Fig. 6). Adverse events Throughout the 4-week intervention period, all participants were monitored for adverse events by the supervising therapist during each training session, and any symptoms or discomfort were recorded in real time using a standardized case report form. No adverse events (including falls, dizziness, cybersickness, fatigue, or any other discomfort) were observed or reported by any participant in either group throughout the 4-week intervention period. The VR training was well-tolerated, supporting its safety for use in patients with stroke. Discussion This randomized controlled trial represents the first investigation to systematically employ both clinical functional assessment and fNIRS to explore the effects of multisensory-integrated VR gait training on gait adaptability and cortical activity in patients with stroke during the recovery phase. The main findings can be summarized across three levels: Behaviorally, the training was associated with significantly improved patients’ walking and balance abilities in complex environments. Neurally, it was associated with enhanced activation in the right PPC and SMA. At the brain network level, it was associated with changes in functional connectivity within SMN and between SMN and cognitive control networks. These improvements in behavioral and neural indices were closely interrelated, collectively revealing a potential rehabilitation pathway. This pathway leverages multi-sensory integration and task-specific training to facilitate functional changes in the brain, thereby improving gait adaptability. Clinical functional improvement The primary conclusion of this study is that multisensory-integrated VR training was associated with a potential advantage in improving gait adaptability and balance in patients with stroke. Compared with conventional training, the VR group showed significantly greater improvement on DGI, TUGT and BBS, which directly reflect adaptability in complex environments, functional mobility and balance [9]. This finding is consistent with the conclusions of previous meta-analyses, which reveal that VR training can effectively improve parameters such as cadence and step length [7], and further extends these benefits to more complex, daily-life task scenarios. However, it should be noted that sensitivity analysis using a worst-case scenario for missing data showed that these significant differences were not sustained under the most conservative imputation assumptions (Supplementary Table 2). Therefore, these findings should be interpreted with caution. Particularly insightful is the VR group’s superior performance on gait tasks incorporating cognitive challenges, despite no between-group difference in basic lower-limb motor function. This suggests that the core therapeutic effect of our training protocol lies in enhancing neural integration and motor control strategies, rather than merely increasing muscle strength [28]. This result aligns perfectly with the core principles of “task-oriented” training and “multi-sensory stimulation” in modern neurorehabilitation [29–31]. It also shares the same rehabilitative rationale as VR studies focusing on upper limb function [32, 33], thereby providing new empirical evidence to support clinical guidelines recommending such technology-enhanced rehabilitation approaches [8]. The observed improvements in gait adaptability, mobility, and balance in the VR group are consistent with a growing body of evidence on dual-task gait training in stroke populations. Recent meta-analyses have demonstrated that cognitive-motor dual-task training significantly improves walking performance (WMD = 3.19, 95% CI: 2.26 to 4.12), lower limb motor function (WMD = 2.78, 95% CI: 1.38 to 4.18), and cognitive function (WMD = 2.93, 95% CI: 0.95 to 4.91) in patients with stroke [34]. Similarly, network meta-analysis findings indicate that motor-walking dual-task training is efficient for improving walking speed, step length, and BBS scores, while cognitive-balance dual-task training yields the most significant improvements in TUGT performance [5]. Our finding of significant enhancements in DGI and TUGT aligns with these meta-analytic conclusions and yields effect sizes comparable to those reported in the literature. For instance, Huimeng et al. [35] reported that four weeks of cognitive or motor dual-task gait training effectively improved lower-limb coordination in subacute patients with stroke, with benefits observed across both training types. Importantly, unlike conventional dual-task paradigms that often use separate cognitive and motor tasks (e.g., walking while counting backwards), our VR training seamlessly integrated cognitive demands into the motor tasks through gamified scenarios that required response inhibition, visuospatial processing, and attention switching. This integration may enhance ecological validity and patient engagement, potentially leading to better transfer to real-world walking situations. The environmental relevance of our approach is supported by studies demonstrating that dual-task abilities during activities representative of daily living are significantly impaired in stroke survivors [36], and that VR paradigms can effectively assess and train these ecologically valid dual-task scenarios [37]. The specific cognitive components engaged in our VR tasks—response inhibition (Go/No-Go), visuospatial processing, and attention switching—align with the mental demands of real-world ambulation. Recent fNIRS studies have shown that dual-task walking in patients with stroke elicits increased prefrontal cortex activation [38, 39], and that different cognitive task types (e.g., arithmetic vs. speech tasks) produce distinct patterns of dual-task interference [40]. Our observed changes in brain activation across regions such as the prefrontal cortex (involved in cognitive control) and the PPC (involved in visuospatial processing) provide neurophysiological support for the dual-task nature of the intervention. These findings are consistent with research demonstrating that frontoparietal regions are critically involved in dual-task walking post-stroke [21], and that cognitive-motor dual-task training can promote neuroplasticity and strengthen functional connectivity in key brain regions [41]. Collectively, these findings suggest that multisensory-integrated VR training may be a practical dual-task training approach, improving both motor and cognitive aspects of gait adaptability. Activation of key brain regions The fNIRS data provide direct neural correlates for the observed clinical functional improvements, precisely identifying two core brain regions associated with the training. These clinical functional improvements carry significant clinical importance, as post-stroke gait abnormalities, particularly spatiotemporal asymmetry and high variability, have been strongly linked to a substantially increased risk of falls [42]. Therefore, investigating the underlying mechanisms, from behavior to neural function, is particularly crucial. PPC: strengthening the hub for visuomotor integration We found that after VR training, patients exhibited significantly enhanced activation in the right PPC during the stepping task, and this activation was negatively correlated with faster walking speed (shorter TUGT time). The PPC is a key node for integrating visuospatial information, online movement planning, and error monitoring [43]. Its role is crucial for visual guidance during interactive gait tasks [44], and its functional integrity is fundamental for successful motor adaptation [20]. The findings of this study suggest that repetitive obstacle avoidance and terrain adaptation training within the VR environment may constitute to functional cahnges in the PPC’s “environmental perception-motor planning” circuit [22]. The enhanced PPC activation observed here reflects improved regional efficiency in visuomotor processing, which may contribute to greater behavioral-level gait adaptability. Importantly, as β-values represent regional activation rather than network-level coordination, these findings should be interpreted as evidence of local functional enhancement within this key node. The observed negative correlation between right PPC activation during the stepping task and TUGT performance warrants further clarification. The stepping task used in our fNIRS assessment—requiring participants to lift their feet in place alternately—serves as a simplified model of the rhythmic lower limb control that is fundamental to walking. Although the TUGT is a composite test involving standing, walking, turning, and sitting, its core component—timed walking over a short distance—reliably relies on efficient visuomotor integration and rapid gait adjustments. Enhanced PPC activation during stepping reflects improved neural efficiency in processing visuospatial information and in generating appropriate motor commands [43]. This neural efficiency likely translates into smoother and faster performance during the walking phase of the TUGT, explaining why greater PPC activation correlates with shorter (better) TUGT times. Thus, the stepping task captures a critical sensorimotor process that underpins performance in more complex functional tasks, such as the TUGT. SMA: optimizing the efficiency of motor planning and initiation Simultaneously, during obstacle-crossing tasks requiring precise gait control, patients showed significantly increased activation in the right SMA. The SMA is responsible for movement preparation, initiation, sequence coordination, and rhythm control [45]. Its enhanced activation suggests that the training may have strengthened the patients’ ability to plan and execute complex movement sequences. Consistent with the interpretation of β-values as regional activation, this finding suggests improved local processing efficiency in the SMA, which may be central to motor planning for complex gait tasks [15]. This finding is supported by other studies employing robotic-assisted training [10] or neurofeedback [16] to modulate the SMA to improve gait directly. The association between SMA activation and functional improvement demonstrates consistency across different training modalities. For instance, Enzinger et al. also found that increased SMA activation in patients with stroke following treadmill training was positively correlated with improvements in motor endurance [46]. Furthermore, longitudinal studies indicate that post-stroke gait training can induce increased SMA activation [47], providing direct neuroplastic evidence for the training effects observed in this study. Additionally, in movement disorders such as Parkinson’s disease, neuromodulation targeting the SMA (e.g., repetitive transcranial magnetic stimulation) has been shown to improve motor function [48, 49]. Changes in network functional connectivity The most profound finding of this study is that VR training was not only associated with enhanced activity in specific brain regions but also with systematic alterations in collaborative patterns within brain networks. Enhanced collaboration between sensorimotor and cognitive-motor circuits Resting-state functional connectivity analysis revealed enhanced connectivity between the left S1 and the right M1 after training. This may indicate improved efficiency in the cross-hemispheric sensorimotor integration pathway [50], which is crucial for rebuilding coordinated and stable support phases following hemiplegia. More importantly, the cognitive-motor control circuit was significantly strengthened, as evidenced by enhanced functional connectivity between the left SMA and the right aPFC, and between the right PPC and the left DLPFC. The DLPFC/aPFC are core regions for higher-order cognitive control functions such as attention and executive function [21]. Given the number of connectivity comparisons performed and the exploratory nature of this analysis, these findings should be considered preliminary and interpreted with caution. The strengthening of these connections may reflect that VR training was associated with improved capacity of the brain’s cognitive system for online monitoring and adaptive regulation of motor behavior. This finding provides a precise neural mechanism that could explain the significant improvement observed in the VR group on the “walking while performing a cognitive task” sub-item of the DGI. By reinforcing the “cognitive-motor” interface, the training enhanced patients’ dual-task processing ability [21]. Interpretation of prefrontal activation patterns: an efficiency-prioritized compensatory strategy A phenomenon worthy of in-depth discussion is that while the local activation of the DLPFC itself did not reach significant between-group differences, its functional connectivity with the PPC was significantly enhanced. As with the other connectivity findings, this result should be considered exploratory given the multiple comparisons. This may reflect an efficient “network-based compensation” strategy [14, 21]. In this training focused on gait adaptability, the brain might have prioritized optimizing the communication bandwidth and efficiency between the key sensorimotor region (PPC) and the cognitive control hub (DLPFC), rather than merely increasing the local metabolic load of the DLPFC. This pattern of “connectivity preceding activation” is not uncommon in neuroplasticity research and may represent a specific stage in the process of functional adaptation. This suggests that future studies aiming to elicit stronger local DLPFC activity might need to incorporate dedicated training modules with greater cognitive complexity (e.g., working memory, conflict resolution). In summary, the findings of this study support the following integrated explanation: Multisensory-integrated VR training, by providing a task environment rich in multi-sensory stimuli and continuously progressive challenges, persistently challenged the patients’ gait control system. This process may have engaged the brain in high-intensity online sensory integration (relying on the PPC) and motor sequence planning (relying on the SMA), leading to enhanced activation in these two key nodes [10, 16]. Repeated training may have further prompted the brain to optimize its resource allocation strategy, associated with changes in functional connections between brain regions responsible for sensory integration, motor execution, and cognitive monitoring [30], particularly by establishing more efficient “cognitive-motor” collaborative pathways [21]. Ultimately, these neural changes, from the local to the network level, were associated with superior gait adaptability in patients in complex real-world environments. Limitations This study has several limitations. First, the 17.6% attrition rate and the finding that dropouts had significantly lower baseline DGI scores than completers indicate that attrition was not completely random. This non-random attrition may have introduced bias, potentially influencing the observed outcomes. The worst-case scenario sensitivity analysis further showed that the significant improvements in DGI, TUGT, and BBS observed in the per-protocol analysis were not sustained under the most conservative assumptions about missing data (all P > 0.05, Supplementary Table 2). Although the FMA-LE outcome remained non-significant across both the per-protocol and worst-case scenario analyses, the sensitivity of the primary outcomes to missing data suggests that the findings should be interpreted with caution. Second, although we reported means and standard deviations for all outcomes, we did not provide effect sizes (mean differences) and 95% confidence intervals for between-group comparisons, as recommended by the CONSORT statement. This limits the ability to assess the magnitude of treatment effects directly and to compare them with those of other studies. Readers should interpret the findings with caution, focusing on the direction and consistency of effects rather than precise estimates. Third, the inherent detection characteristics of fNIRS technology limited our ability to assess changes in deep brain structures such as the basal ganglia and cerebellum, which are also crucial components of the gait network. Fourth, the relatively small sample size and single-center design limit the generalizability of the findings, and the exploratory nature of this study meant that corrections for multiple comparisons were not applied, increasing the risk of Type I error. Therefore, all findings should be considered preliminary. Fifth, the lack of long-term follow-up data prevents confirmation of the durability of the training-induced neurobehavioral improvements. Sixth, although the study protocol was submitted for registration on time, the public release of the registration record was delayed by administrative procedures, constituting a procedural limitation. Seventh, the absence of unilateral spatial neglect and hemianopia was determined based on medical records and clinical examination rather than standardized screening tools, which may have missed subclinical cases. Future research should address these limitations through multi-center trials with larger samples, long-term follow-up, and pre-specified primary outcomes with appropriate corrections for multiple comparisons, and ensure that all CONSORT-recommended statistics are fully reported. Conclusion This study suggests that multisensory-integrated VR gait rehabilitation training may improve gait adaptability and balance in patients with stroke during the recovery phase, although these findings should be considered preliminary given the sensitivity of the primary outcomes to missing data. The observed effects were associated with multilevel neuroplastic changes: behaviorally, improvements in walking and dual-tasking abilities in complex environments; neurally, increased activation in the PPC and SMA, along with changes in functional connectivity within the SMN and between the SMN and cognitive control networks. These findings provide preliminary evidence, encompassing both clinical benefits and neural mechanisms, that may support the potential integration of this training modality into stroke rehabilitation protocols. However, given the non-random attrition and the finding that the significant improvements in DGI, TUGT, and BBS were not sustained under the most conservative assumptions about missing data, these results should be considered preliminary and require confirmation in future studies with lower attrition rates and more robust designs. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2. Supplementary Material 3. Supplementary Material 3. Supplementary Material 4. Supplementary Material 4. Supplementary Material 5. Supplementary Material 5. Supplementary Material 6. Supplementary Material 6. Supplementary Material 7. Supplementary Material 7. Supplementary Material 8. Supplementary Material 8. Supplementary Material 9. Supplementary Material 9. Supplementary Material 10. Supplementary Material 10.