The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 declaration was followed in this systematic review and was checked against the PRISMA 2020 expanded checklist in Checklist 1 [26]. The study protocol was prospectively recorded in the International Prospective Register of Systematic Reviews (PROSPERO: 2025 CRD420251131116).

This study adhered to the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) checklist (Checklist 2) [26] extension to systematically document the literature search strategy and process. Information sources included the following databases and their respective platforms: PubMed (via NCBI), Cochrane Library (via Wiley), Web of Science Core Collection (via Clarivate), Embase (via Ovid), CINAHL (via EBSCOhost), and Scopus (via Elsevier). Backward and forward citation tracking was performed using reference lists and the citation index in Web of Science. Furthermore, we have traced and screened the studies included in relevant previous reviews to supplement the present review. Experts and corresponding authors were contacted by email to identify unpublished or ongoing studies. No additional sources or search methods were used.

Detailed search strategies combining subject headings and free-text terms were developed for each database; the full strategies are provided in Multimedia Appendix 1. Searches were limited to records from database inception to January 27, 2026, and to English-language publications. Published search filters for RCTs were adapted where applicable. The search approach was informed and updated from previously validated strategies in related reviews [22]. All strategies underwent peer review by an experienced information specialist and were refined accordingly. Literature collection and screening were conducted through EndNote (version 20; Clarivate).

Inclusion and exclusion criteria were formulated according to the population, intervention, comparison, outcomes, and study design principles [27]：

Double-blind screening was carried out independently by 2 reviewers (LW and XL) who first selected papers by looking at titles, abstracts, and keywords. Full-text screening against inclusion or exclusion criteria was performed on papers that satisfied the first criterion. To guarantee thorough literature coverage, secondary screening required obtaining reverse references. A third reviewer was consulted to address disagreements.

Two independent researchers extracted pertinent publication data (eg, authors, title, and year), study design (eg, RCTs and number of assessment time points), patient numbers, patient characteristics (eg, age, sex, and duration of pain), intervention measures, available outcome measures, and results. For all pre- and postintervention measures, outcome data were displayed as mean (SD). We translated data that were presented as discrete measures, such as the median, to mean (SD) [28].

Based on quality evaluation criteria, the Cochrane Library’s Risk of Bias 2 tool was used to evaluate the quality of the included studies, primarily assessing RCTs across 5 domains: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing outcome data, bias in measurement of the outcome, and bias in selection of the reported result [29]. The κ coefficient was used to measure methodological quality and risk of bias assessment consistency between the 2 independent reviewers. Any item-level disagreements arising during this process were resolved by arbitration from a third reviewer.

The credibility of the cumulative evidence was evaluated using the Confidence in Network Meta-Analysis (CINeMA) framework [30,31] to categorize the evidence for each intervention across outcome measures into 4 levels: high, moderate, low, and very low [32,33]. Two independent reviewers (LW and XL) evaluated each study. Discrepancies were resolved through discussion, and a third reviewer (TY) adjudicated any unresolved disagreements.

A Bayesian network meta-analysis was performed using R software (v4.5.1; R Foundation for Statistical Computing) and RStudio (Posit Software, PBC) within a Markov Chain Monte Carlo framework [34]. Network diagrams illustrated direct comparisons between VR therapies, and heterogeneity was assessed using chi-square tests [35]. To maintain alignment with other indicators, a random-effects model was always used, regardless of whether significant heterogeneity was found. The standardized mean difference (SMD) and 95% credible intervals (CrIs) were used to report pooled effect sizes and precision. Results were presented in forest plots and effect estimate tables. Global consistency was evaluated by comparing deviance bar and deviance information criterion between consistent and inconsistent models, with a deviance information criterion difference <5 favoring consistency [36].

Model convergence was assessed via diagnostic plots (Figure S6, Multimedia Appendix 2); studies with shrinkage factors >1.02 were excluded via sensitivity analysis [37]. To explore heterogeneity sources, prespecified covariates—including types of VR interventions, intervention complexity, patient age, intervention duration, region, sample size, study year, and control type—were incorporated using Bayesian network meta-regression. A covariate was considered significant if the 95% CI of its regression coefficient excluded 0 [38]. Node-splitting analysis evaluated local inconsistency by testing differences between direct and indirect evidence. Treatments were ranked using surface under the cumulative ranking curve (SUCRA) and average rankings, with higher SUCRA values indicating better efficacy [39]. Publication bias was assessed using adjusted funnel plots and the Egger test (for outcomes with ≥10 studies); P<.05 and funnel plot asymmetry indicated potential bias, with the trim-and-fill method applied if small-study effects were detected [40].

The initial yield from each database was: PubMed (n=124), Embase (n=229), Cochrane Library (n=159), Web of Science (n=267), Scopus (n=208), CINHAL (n=76), and other sources (n=5). The initial search identified 1063 records from databases and 5 from other sources. After removing duplicates, 545 records were screened by title and abstract. Of these, 414 were excluded for being abstracts only, involving animal studies, conference proceedings, or not being RCTs. Full texts of the remaining 104 papers were assessed. Subsequently, 61 papers were excluded due to unavailability of full texts or trial registrations, and a further 23 were excluded for reasons including unsuitable interventions, missing relevant outcomes or participants, being secondary or follow-up analyses, or unavailable data. Finally, 25 studies [41-65] were included in the network meta-analysis (Figure 1).

The 25 RCTs included in this study involved a total of 2183 patients with CLBP. The sample size of individual studies ranged from 18 to 1067 participants, with a mean age between 20 and 68 years and a generally balanced sex distribution. The interventions were categorized into 3 technical types: fully immersive VR (using head-mounted displays for complete visual immersion, 7 studies), semi-immersive systems (eg, large-screen projection combined with partial body interaction, 1 study), and nonimmersive desktop-based VR (monitor-based, 17 studies). Table 1 lists the author-assigned group names and definitions that were used to define the specific intervention protocols and control groups. In terms of dose parameters, the intervention duration varied from a single acute session to long-term programs lasting up to 12 weeks. The treatment frequency ranged from 2 to 5 sessions per week, with each session lasting 15 to 60 minutes, yielding a cumulative intervention time of 30 minutes to 36 hours. These dose parameters will be used for subsequent heterogeneity analysis and subgroup comparisons. Details are shown in Table 2.

The methodological quality of the included RCTs was assessed using the Cochrane Risk of Bias 2 tool. Among the 25 studies, 2 were rated as low risk, 15 as moderate risk, and 8 as high risk (Figure 2). Overall, the methodological quality of the included studies was judged as moderate. The interrater agreement κ values for each domain ranged from 0.61 to 0.84, reflecting substantial consistency among assessors.

The quality of evidence across the outcomes varied. For pain intensity (visual analog scale), it ranged from moderate to very low, with most comparisons rated as low or very low. All comparisons for disability and kinesiophobia were rated as very low. The downgrading of evidence was primarily due to within-study bias, imprecision, and heterogeneity or incoherence. Indirectness and reporting bias were not reasons for downgrading in any comparison. The results of all comparisons are provided in Table 3.

Consistent with previous systematic reviews, the present network meta-analysis demonstrated that shooting games and VR-based equestrian training significantly reduced pain intensity in patients with CLBP. No VR intervention yielded a statistically significant improvement in disability. Shooting games significantly outperformed traditional exercise controls in reducing kinesiophobia, although no significant difference was observed relative to nonexercise or placebo controls. Based on the GRADE (Grading of Recommendations Assessment, Development and Evaluation) criteria, the overall quality of evidence ranged from low to very low, primarily due to high heterogeneity, risk of bias, and imprecision. Network meta-regression revealed no significant moderating effects of covariates. Despite significant heterogeneity, the good consistency of the model supports the reference value of the present findings [66].

The superior performance of shooting games and VR-based equestrian training in pain intensity may be explained by their integration of core stability training with cognitive-behavioral elements within an immersive, interactive environment [67]. The shooting game intervention was not blinded to participants or outcome assessors throughout implementation, which could have resulted in an overestimation of the treatment effect; this conclusion should be regarded cautiously [68]. Exercise therapy and cognitive behavioral treatment are 2 examples of how pain science is moving toward a biopsychosocial model. By strengthening movement confidence, encouraging graded exposure to movement, and improving attentional distraction, these modalities probably address both the nociceptive and emotional aspects of pain [69]. The combination of physical and psychological involvement may be essential for therapeutic success, as seen by the lower rankings for pain outcomes of therapies that mostly rely on cognitive restructuring or passive distraction, such as VR-based cognitive behavioral therapy. None of the VR therapies significantly improved impairment. Moreover, the certainty of evidence for all disability comparisons was rated as very low, due to serious imprecision, inconsistency, and high risk of bias. There is a possibility that excessive clinical heterogeneity between studies may have led to nonsignificant results [70]; therefore, we searched for another disability assessment tool for comparison. The studies by Kim et al [60] and Zadro et al [59] used the Roland-Morris Questionnaire for disability assessment, and both reported no statistically significant differences between groups, which further supports the conclusions of this study. The analysis of Li et al [22] showed notable gains in the disability index, most likely as a result of selecting possibly more successful trials by establishing a criterion of ≥8 interventions. In contrast, traditional exercise therapy leads to more significant improvement in disability [71]. While VR interventions frequently concentrate on improving engagement and adherence without clearly separating the “medium effect” from the “therapeutic component effect,” traditional interventions benefit from well-validated mechanisms that directly target functional capacity. This could help to explain why the evidence for disability outcomes is still far less certain than that for pain intensity [72]. We found a decrease in kinesiophobia that was comparable to the improvement trend in pain scores when comparing the shooting game to conventional exercise therapy. This discrepancy may be caused by the lack of diversion, which increases kinesiophobia levels during treatment [10]. However, this comparison is based on data from a single study and has very poor-quality evidence. We summarized the characteristics of the intervention in this study, where participants were exclusively soccer athletes with an average age below 30 years. The unique physical activity levels, beliefs, and attitudes toward exercise within this population may introduce confounding factors affecting the reliability. This population’s particular levels of physical activity, attitudes toward exercise, and ideas about it could introduce confounding factors that alter the reliability [54,65].

The present findings are consistent with and extend previous evidence syntheses. Huang et al [16] reported no significant analgesic effect of VR for CLBP, but their analysis included only 2 RCTs with limited statistical power. Kumar et al [24] and Choi et al [25] also concluded that VR treatment can reduce pain; they did not perform subgroup analyses based on VR game type, level of immersion, or degree of physical involvement, nor did they explore how different game contents might influence outcomes. Li et al [22] confirmed that VR training improves pain, kinesiophobia, and disability immediately after the intervention, but did not distinguish specific VR game types. Lo et al [18] suggested that nonimmersive VR is more effective for low back pain, while Henríquez-Jurado et al [23] found no significant difference between immersive and nonimmersive VR. These studies indicate that hardware-based classification fails to capture therapeutic differences.

Brea-Gómez et al [11] conducted a meta-analysis stratified by VR devices. By performing subgroup analyses based on 3 distinct types of VR devices (Nintendo consoles, horse-riding simulators, and ProKin system), the study found that horse-riding simulators and the ProKin system significantly reduced pain compared with control conditions, whereas Nintendo consoles did not show a significant effect. Regarding kinesiophobia, the ProKin system was effective, while Nintendo consoles even yielded unfavorable results. These findings represent an important early effort to explore the differential efficacy of various VR game-like devices, clearly indicating that hardware characteristics and interaction modalities including sensor precision, immersion level, and task type may critically influence treatment outcomes. Rather than classifying VR interventions based on hardware or immersion level, this study extends research on efficacy differences across VR categories by refining task‐specific typologies and directly comparing 7 distinct VR content modalities. We confirmed that VR interventions combining dynamic full-body movement and cognitive engagement (shooting games and equestrian training) are superior to passive or single-task VR modalities. Consistent with previous studies, no VR intervention significantly improved disability, providing more precise network meta-analytic evidence.

This study represents the first network meta-analysis to systematically compare and rank the efficacy of distinct VR-based training modalities on pain intensity, disability, and kinesiophobia in patients with CLBP, which moves beyond aggregate categorizations of VR and enables direct and indirect comparisons across different intervention types. By identifying the relative ranking of interventions, this work contributes to the field by clarifying which VR modalities may offer the greatest therapeutic potential and by revealing critical gaps in the existing evidence base, such as high heterogeneity and limited evidence for disability outcomes. However, several limitations warrant consideration. The analysis was constrained by the inclusion of only 25 studies and 7 intervention types. Data were particularly scarce for kinesiophobia, potentially affecting the robustness of network meta-analysis. Furthermore, interventions like pelvic flexibility games, VR-based walking training, and VR-based yoga training are likely undervalued due to limited studies, small samples, wide CIs, and poor network connectivity, reliant on indirect comparisons [73]. The occurrence of VR-related adverse effects and participant dropouts in some studies may limit engagement levels and could affect the interpretation of overall outcomes, even where reported attrition is low.

In clinical practice, for the treatment of pain and kinesiophobia using VR, we recommend shooting games, while taking into account the patient’s individual circumstances, or using VR as an adjunctive therapy in combination with other methods. Given the current evidence profile, none of the VR-mediated interventions can be recommended, as determined by substantial heterogeneity, high risk of bias, and low certainty of evidence. Future research could focus on the following. First, to obtain higher-level comparison effect estimates, future research should concentrate on performing more head-to-head RCTs that directly compare various VR exposure regimens. Second, to minimize heterogeneity and improve the reproducibility of outcomes, the hardware components, software components, dosage of the intervention, and length of treatment should be precisely specified. Third, long-term follow-up should be included in studies to evaluate the sustainability of the intervention and investigate its long-term benefits for relapse prevention and improvement of disability.

Overall, certain VR-based trainings show promise for CLBP. The shooting game demonstrated a relative advantage over the also effective simulated horse-riding training in improving pain and alleviating kinesiophobia, ranking higher in pain control probability. However, definitive conclusions remain limited by insufficient evidence and considerable clinical heterogeneity. More high-quality RCTs are needed to evaluate their short- and long-term benefits and to provide robust evidence.