This review was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) guideline. The systematic review was registered with the International Platform of Registered Systematic Review and Meta-Analysis Protocols (INPLASY; registration number INPLASY2024110101).

A comprehensive search with no language restrictions was conducted through 6 electronic databases, including MEDLINE, Embase, PubMed, Cochrane Central Register of Controlled Trials (CENTRAL), Web of Science, and PsycINFO. Relevant studies published from database inception to July 2024 were included. Details of the search strategy are provided in Multimedia Appendix 1. Studies were included if they met the following criteria: (1) study design was randomized controlled trials (RCTs), (2) participants were adults (≥18 y) diagnosed with stroke, (3) interventions involved treatment using VR technology, including IVR, SIVR, and NIVR, (4) comparisons were made against routine rehabilitation (conventional therapy), and (5) outcomes assessed lower limb motor function, balance, gait, walking, or mobility. Studies were excluded if they lacked detailed intervention descriptions or outcome indicators, or did not report pre- and postintervention comparisons. In addition, studies combining VR interventions with external neurostimulation techniques were excluded. The detailed criteria are summarized in Table 1.

The outcomes of interest were functional recovery of the lower extremities, categorized into primary and secondary outcomes. The primary outcomes included balance, assessed using the Berg Balance Scale (BBS), and mobility, measured by the Timed Up and Go Test (TUG) and the 10-Meter Walk Test (10-MWT). Secondary outcomes included stride and step length, the Functional Reach Test (FRT), the Falls Efficacy Scale (FES), and the Dynamic Gait Index (DGI).

Two reviewers (WL and MS) independently screened the titles and abstracts, removing duplicates and excluding irrelevant studies. The remaining studies were then assessed in full based on predefined inclusion and exclusion criteria. Any disagreements were resolved through discussion, with a third arbiter (YW) involved when necessary. Extracted data included study design, participant characteristics (age, sex, stroke type, and disease duration), intervention details (type, duration, and frequency), and outcome measures (balance, gait speed, and step length). To compare pre- and postintervention data, the mean change (Δmean) was calculated by subtracting the postintervention mean from the preintervention mean. In addition, the change in SD (SD_difference) was computed using the formula from the Cochrane Handbook [22].

The Cochrane Risk of Bias (RoB) tool was used to assess the quality of RCTs [23], focusing on sequence generation, allocation concealment, blinding, and outcome reporting. The updated RoB2 tool [24] was used to evaluate five domains: randomization process, deviations from intended interventions, outcome measurement, missing outcome data, and selective reporting. Each domain was rated as low, high, or unclear risk of bias, contributing to an overall judgment for interpreting meta-analysis results.

We performed a meta-analysis using a random-effects model for all continuous variables. Mean differences (MD) and 95% CIs were calculated for primary and secondary outcomes, with P<.05 considered statistically significant. I² was used to assess heterogeneity. Subgroup analyses were performed based on factors such as time after stroke (≤7 d, >7 d ≤6 mo and >6 mo), VR type (IVR, SIVR, and NIVR) and total frequency of intervention (<20 sessions and ≥20 sessions). We also performed a sensitivity analysis using the latest assessment in each study. RevMan 5.4 (The Cochrane Collaboration) software was used for all analyses and forest plots illustrating the pooled results. Both RevMan 5.4 software and Stata 17.0 (StataCorp) were used for funnel plots and Egger test to assess publication bias.

A total of 2145 records were identified through database searches. After removing 477 duplicates, 1668 records were screened based on titles and abstracts, leaving 426 studies for full-text review. Following this review, 24 studies [25-48] were ultimately included in the meta-analysis. The flowchart is presented in Figure 1.

A total of 24 RCTs [25-48] comprising 768 participants were included in the analysis. The studies focused on patients with stroke across different recovery stages, with 473 (61.6%) participants targeting individuals more than 6 months poststroke [28] and the remainder focusing on the subacute stage (7 d to 6 mo) [46]. Most studies (16/24, 67%) [25-32,35,37,39-43,48] used a single-blind design, while 3 (13%) studies [36,40,41] used IVR intervention, 7 (29%) studies [29,31,33-35,46,48] used SIVR intervention and 14 (58%) studies [25-28,30,32,37-39,42-45,47] used NIVR intervention. Intervention durations ranged from 3 to 12 weeks, with 10‐40 total sessions and an average of 18‐19 sessions. The most common frequencies were 12 sessions (6 studies [28,29,39,40,43,47]) and 20 sessions (5 studies [31,32,34,35,37]). The baseline characteristics of each study are detailed in Multimedia Appendix 2.

The meta-analysis results are summarized in Table 2. Three of the outcomes were statistically different.

Meta-analysis of 16 studies (496 participants) showed VR therapy significantly improved lower limb mobility in patients with stroke compared to conventional therapy, with a mean TUG score difference of −1.67 (95% CI, −2.89 to −0.46; P=.007) seconds. Heterogeneity was low (I²=0%; P=1.00), indicating consistent and reliable results favoring VR therapy (Figure 4).

VR therapy significantly improved TUG scores in patients >6 months poststroke (MD −1.53 s, 95% CI −3.08 to −0.02; P=.05). In contrast, no significant improvement was observed in patients 7 days to 6 months poststroke (MD −1.62 s, 95% CI −3.85 to 0.61; P=.15). These findings underscore the greater effectiveness of VR therapy in improving mobility in patients with chronic stroke.

NIVR significantly improved TUG scores compared to conventional therapy (MD −1.67 s, 95% CI −3.10 to −0.23; P=.02) with no heterogeneity. In contrast, neither IVR (MD −1.13 s, 95% CI −5.04 to 2.78; P=.57) nor SIVR (MD −1.96 s, 95% CI −4.75 to 0.82; P=.17) showed significant effects, highlighting the superior efficacy of nonimmersive VR in improving functional mobility in patients with stroke.

A total frequency of ≥20 sessions significantly improved TUG performance compared to conventional therapy (MD −1.98 s, 95% CI −3.33 to −0.63; P=.004) with no heterogeneity. In contrast, <20 sessions showed no significant effect (MD –0.39 s, 95% CI −3.16 to 2.37; P=.78). Emphasizing the enhanced efficacy of higher-frequency VR therapy in promoting lower limb mobility improvement. The above results are demonstrated in Figure 5.

Seven studies [33-35,37,40,43,44] (167 participants) showed no significant difference in gait speed improvement between VR therapy and conventional treatment, with a mean 10-MWT difference of −0.91 (95% CI −3.33 to 1.50; P=.46) seconds. Heterogeneity was low (I²=0%; P=.98), indicating high consistency across studies, but the overall results did not favor VR therapy (Figure 6). Meanwhile, subgroup analysis revealed that VR therapy did not significantly improve gait speed across different poststroke durations, VR types, or treatment frequencies, indicating limited efficacy in enhancing 10-MWT outcomes. Forest plots for these results are shown in Figures S1-S3 in Multimedia Appendix 4.

More than 50% (14/24) of the studies [25-27,29-32,35,37,38,41,42,44,46] had a low risk of bias in random sequence generation, indicating the use of appropriate randomization methods. However, approximately 50% (13/24) of the studies [25,28,31,33,34,36,39,40,43-46,48] had unclear allocation concealment, and over 10% (3/24) of the studies [30,38,47] had a high risk, suggesting potential issues with randomization. In contrast, 75% (18/24) of the studies [26-28,30-33,37-43,45-48] had a low risk of bias in blinding of outcome assessors. In addition, 95% (23/24) of the studies [25-45,47,48] demonstrated a low risk of bias related to incomplete outcome data, though a small number had an unclear risk. Over 80% (20/24) of the studies [25,26,28-37,40-46,48] showed a low risk of selective reporting, with no evidence of reporting bias. Overall, the risk of other biases was minimal. In conclusion, while concerns remain regarding allocation concealment and blinding, approximately 70% (17/24) of the studies [25,26,29,31-37,40-43,45,46,48] had a low risk of bias, supporting the credibility of the meta-analysis findings (Figures 8 and 9). The funnel plot (Multimedia Appendix 5) showed a symmetrical distribution, with most studies clustered at the top, suggesting potential small-sample effects, as negative or nonsignificant findings are less likely to be published. The Egger test did not indicate the presence of publication bias (P>.05). The funnel plots are shown in agreement with statistical tests because they lack significant asymmetry.

This study included 24 RCTs [25-48] with a total of 768 participants. Our findings support the initial hypothesis that VR-based interventions significantly improve lower limb function compared to conventional therapy. The observed positive effects on lower limb rehabilitation, particularly in BBS, TUG, and step length, underscore the potential of VR therapy in advancing rehabilitation approaches and improving patient outcomes. VR technology has proven effective in facilitating motor learning through multisensory integration and task-oriented training [49,50]. VR environments provide real-time visual, auditory, and proprioceptive feedback, which can enhance neuroplasticity and promote motor function recovery [51]. The immersive multisensory interaction in VR also improves patient attention, motivation, and engagement during rehabilitation [52]. In addition, task-oriented training in VR allows patients to engage in repetitive, goal-directed movements that mimic real-life activities, reinforcing motor control and improving balance and gait performance [53]. This approach motivates patients to practice essential movements repetitively, leading to improved outcomes. These mechanisms collectively highlight the potential of VR in enhancing rehabilitation outcomes, providing a foundation for further analysis in the subsequent subgroup results. Subgroup analyses showed that VR therapy was more effective in patients more than 6 months after stroke and less effective between 7 days and 6 months after stroke in BBS. According to recommendations from the Mayo Clinic [54], stroke rehabilitation typically begins within 24-48 hours after stroke onset, with the highest rates of recovery generally occurring within weeks and months following the stroke. However, our study found that VR therapy demonstrated statistically different effectiveness in patients with stroke more than 6 months after stroke. This seems to represent a potential rehabilitation benefit from VR therapy even in patients who have had a stroke more than 6 months after the stroke. Regarding the impact of different VR types on outcomes, studies have shown that NIVR was effective in improving TUG test results in patients with stroke, whereas SIVR and IVR were not statistically different. Most importantly, our study demonstrated that for stroke survivors, 20 or more sessions of VR treatment resulted in statistically significant improvements in balance and mobility, while less than 20 sessions did not. Therefore, this study suggests that higher treatment frequency, especially more than 20 sessions, is more helpful in improving balance and mobility functions. A study [55] conducted a subgroup analysis on the effect of VR on upper limb function recovery in patients with stroke based on intervention frequency, which found that when VR intervention frequency exceeded 18 sessions, upper limb motor function outcomes showed statistical differences compared to fewer than 18 sessions. This is similar to our study, where lower limb functions (BBS and TUG) began to show statistical differences when the total intervention frequency reached ≥20 sessions.

Regarding the high heterogeneity of BBS results, we have conducted subgroup analyses based on the time after stroke, VR type, and total frequency. The results showed that time after stroke and differences within the NIVR group may be the cause of heterogeneity. At the same time, we found that the heterogeneity of the group with a total frequency greater than or equal to 20 times was high, which may be due to the different durations and weekly times of different study schemes. In addition, we sorted out VR devices and intervention content (Multimedia Appendix 2) and conducted a sensitivity analysis (Multimedia Appendix 3). The results showed that the main source of heterogeneity came from studies using Nintendo Wii, with a value as high as 90%. The intervention content differed in each study, which prevented us from performing a subgroup analysis by intervention content. We added the intervention content and VR devices of each study in the table named “Baseline Characteristics of Included Studies” (Multimedia Appendix 2). In addition, we conducted a sensitivity analysis by the leave-one-out method and found that one study [25] was an important source of heterogeneity, ie, the overall heterogeneity of the studies changed from 81% to 35% after removing this study, but the overall heterogeneity remained 78%‐83% after removing the other studies in turn, but unfortunately, we did not find a significant source of heterogeneity for this study.

Currently, the American Stroke Association’s recommended rehabilitation measures for poststroke recovery include physical therapy, audiology, recreational therapy, and rehabilitation nursing. As a form of audiology recreational therapy, VR treatment can create immersive VR environments that generate an illusion of bodily movement. This immersive experience can increase the activation of brain areas related to movement, thereby boosting neuroplasticity and facilitating the reorganization of synaptic connections within the nervous system. By impacting the central nervous system directly, VR therapy plays a key role in reshaping and aiding the recovery of neural structures after a stroke, contributing significantly to rehabilitation and the restoration of motor functions [56]. When considering VR as a vital component of stroke rehabilitation, this study points out that the following four factors should be taken into account:

A key limitation of this study is the relatively small sample size due to the lack of large-scale population studies, which may affect the statistical power and generalizability of the results. The study also assumed that data were normally distributed, and the conversion of medians to means in some studies may introduce errors [58], especially if the data deviate from normality. Furthermore, the calculation of changes in SD assumed a correlation coefficient (r) of 0.5, without conducting sensitivity analyses to explore the impact of different r values, which could affect the accuracy of SD calculations. In addition, the study did not evaluate the safety of VR therapy, and although it is generally considered safe, some studies report minor adverse effects like dizziness or headaches [59]. Besides, since the patients in the included studies did not differentiate whether the patients had a previous hemorrhagic stroke or ischemic stroke, one of the limitations of our study was that we could not further give separate personalized VR protocols. Finally, our limitation is the high heterogeneity of our study, the source of which may be due to the fact that the interventions in the control group were not uniform, for example, balance training in some studies and treadmill exercise in others. Therefore, we addressed this issue through subgroup analysis and sensitivity analysis. Future research should address these limitations by using larger sample sizes, obtaining original mean and SD data, verifying assumptions on correlation, and systematically evaluating the safety of VR-based therapy.

A strength of our study is the inclusion of a larger number of studies in the meta-analysis. This increased sample size allowed for more extensive subgroup analyses compared to previous studies, which provided valuable insights into factors influencing the effectiveness of VR-based therapy. In addition, our study revealed that higher intervention frequency may play a crucial role in stroke rehabilitation, offering guidance for future treatment strategies.

Meta-analysis of RCTs demonstrates that VR therapy is effective in improving lower limb function in stroke survivors, with higher treatment frequency (particularly ≥20 sessions) yielding superior outcomes. These findings highlight the importance of session intensity in maximizing rehabilitation benefits and provide strong evidence to support the integration of VR-based therapies into standard rehabilitation protocols for patients with stroke. Compared to earlier studies, this review emphasizes the critical role of treatment frequency, demonstrating that increased VR therapy sessions significantly enhance recovery, particularly in balance and functional mobility. Based on these results, we recommend that ≥20 sessions of VR interventions be incorporated into clinical guidelines for chronic stroke rehabilitation to optimize functional outcomes. Furthermore, the accessibility and cost-effectiveness of VR interventions should be considered by policy makers for potential inclusion in medical insurance coverage, which could improve patient access and reduce long-term health care costs. However, limitations such as the relatively small sample size and the need for more comprehensive safety evaluations highlight the necessity for further investigation. Future research should examine the effects of various VR technologies and environments on stroke recovery to refine therapeutic strategies and enhance clinical outcomes.