The meta-analysis was conducted according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement [23]. The completed PRISMA 2020 checklists for the main manuscript (Checklist 1) and the abstract (Checklist 2) are provided. The study protocol was registered in the PROSPERO database (CRD420251113000).

Ethics approval was not required for this systematic review because it did not involve the recruitment of patients.

The search strategy was developed de novo and was not adapted from prior reviews. It was reported in accordance with the PRISMA-S (Preferred Reporting Items for Systematic reviews and Meta-Analyses literature search extension) guideline [24]. The PRISMA-S checklist is provided as Checklist 3.

We systematically searched 12 databases and 1 gray literature source, including PubMed (NLM); Web of Science (Clarivate Analytics); Institute of Electrical and Electronics Engineers (IEEE) Xplore; Scopus (Elsevier); CINAHL (EBSCOhost); PsycINFO (EBSCOhost); EMBASE (Elsevier); Cochrane Central Register of Controlled Trials (CENTRAL; Cochrane Library); and Chinese databases, encompassing China National Knowledge Infrastructure (CNKI), Wanfang Chinese Database, China Science and Technology Journal Database (VIP), and China Biology Medicine disc (CBM). No simultaneous multidatabase searches were performed on any platform, and each database was searched separately. The CENTRAL database was searched as a study registry for ongoing and completed randomized controlled trials (RCTs). All databases were initially searched from inception to August 15, 2025. An updated search was conducted on April 30, 2026, to identify any additional eligible studies published in the interim. The updated search identified 2 additional eligible studies, which were incorporated into the final synthesis. In addition to database searches, gray literature was searched via ProQuest Dissertations and Theses Global (ProQuest), and the reference lists of all included studies were manually screened to identify any other eligible studies. No online resources or print sources were browsed.

The search strategy was developed by FC in consultation with XG, using a combination of MeSH terms and free-text keywords, related to gastrointestinal cancer (eg, gastric/stomach, colorectal/colon/rectal), IVR (eg, virtual reality and head-mounted display), and RCTs. The search strategy was peer-reviewed by XX using the PRESS (Peer Review of Electronic Search Strategies) checklist [25] prior to execution. The search strategy was first developed in PubMed and subsequently adapted for each database, accounting for differences in syntax, controlled vocabulary, and search operators across platforms. No database-level date or language limits were applied during the searches. Language eligibility was assessed during the screening phase in accordance with the inclusion and exclusion criteria. The full search strategies for all databases are provided in Multimedia Appendix 1.

Inclusion criteria were formulated based on the PICOS framework:

The exclusion criteria were as follows: (1) studies for which the full text could not be obtained or that were conference abstracts, editorials, or protocols and (2) studies not published in Chinese or English.

All search results were imported into NoteExpress (version 4.1.0.10021; Beijing Aegean Software) for data management. After all duplicate articles were removed, 2 researchers (FC and XG) independently conducted a primary screening by examining the titles and abstracts of the studies. Then, they reviewed the full text of each paper according to the inclusion and exclusion criteria and performed data extraction independently, including author, study year, country, study design, sample size, participants (age and sex), time point, intervention types, intervention setting, provider, intervention methods, intervention content, intervention duration, control conditions, outcomes, and instruments. The data extraction form was developed and pilot-tested based on the Cochrane Handbook for Systematic Reviews of Interventions [28]. Meanwhile, they cross-checked the screening results. Any discrepancies were resolved by discussion or, if necessary, inclusion of a third researcher (XX). In the interest of transparency, the PRISMA checklist has been included as Checklist 1. This checklist details our adherence to reporting standards, facilitating the assessment and reproducibility of our research.

The risk of bias was assessed by 2 researchers (FC and XG) independently according to the Revised Cochrane risk-of-bias tool for randomized trials (RoB 2) [29]. Disagreements were resolved either by consensus or by a third partner (XX). Five domains were considered in the evaluation process: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Each domain was assessed for risk of bias, and the overall bias of studies was then identified to be “low risk,” “some concerns,” and “high risk.”

The Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) approach was used by 2 independent researchers (RW and YL) to assess the certainty of the evidence. The GRADE approach is a widely recognized tool for grading evidence certainty in systematic reviews and clinical practice guidelines. It categorizes evidence certainty into 4 levels, “very low,” “low,” “moderate,” or “high,” each determined by specific criteria [30]. Factors that may diminish evidence certainty (eg, risk of bias, inconsistency, indirectness, imprecision, and publication bias) as well as those that may enhance it (eg, large effect, plausible confounding, and dose-response) were evaluated [31]. In the event of discrepancies in the certainty ratings among the reviewers, a third reviewer (XX) was consulted.

For continuous variables, with the exception of vital signs where most studies reported the mean difference (MD) and SD, all other indicators were expressed as mean (SD) post intervention. For studies that measured outcomes at multiple time points, the first postintervention measurement was selected for the meta-analysis of the main effects. Furthermore, we calculated the MD or standardized mean difference (SMD) with the corresponding 95% CI, based on whether the outcomes were measured by the same scales or not. For categorical variables, we calculated the risk ratio with corresponding 95% CI. Statistical significance was determined using a 2-sided P value threshold of <.05. The effect sizes were classified as small (0.20‐0.49), moderate (0.50‐0.79), or large (≥0.80) [32]. If the study did not provide mean or SD, we converted available data into mean (SD), median (95% CI), or median (IQR), according to the Cochrane Handbook for Systematic Reviews of Interventions [28].

Where outcome data were incompletely reported in included studies (eg, missing SDs or only median reported without dispersion measures), we contacted the corresponding authors via email to request the missing information. A maximum of 2 email attempts were made, spaced 2 weeks apart. For studies where authors did not respond or data remained unavailable, the study was included in the systematic review narrative but excluded from the specific quantitative meta-analysis for that outcome. This approach was documented in the study characteristics table and noted as a limitation.

To address the anticipated heterogeneity arising from variations in clinical settings and populations across studies, random-effects models were preferred to the usual statistical tests [33]. In addition, the Hartung-Knapp-Sidik-Jonkman method was used to reduce the production of false positives inherent to the DerSimonian-Laird method and to obtain more robust estimates of variance [34]. Heterogeneity was quantified using the τ² statistic [35], while the I² statistic was reported for descriptive reference (with values of 25%, 50%, and 75% commonly indicating low, moderate, and high levels, respectively) [36]. To better assess the practical implications of heterogeneity, a 95% prediction interval was also calculated [37] to estimate the range within which the true effect of a future, similar study would be expected to fall, thereby providing a more clinically meaningful interpretation of heterogeneity. If meta-analyses showed high heterogeneity, we performed subgroup analyses and meta-regression to look into any potential study-level factor influencing the treatment effect. We conducted sensitivity analyses on the outcome measures using the leave-one-out method to assess the robustness of the combined results.

A total of 3054 records were identified through database searches, with an additional 14 records retrieved via other methods, including citation searching and ProQuest Dissertations & Theses Global. After removing 1204 duplicates, 1850 articles underwent title and abstract screening, of which 1745 were excluded as irrelevant. Full texts of 105 articles were retrieved for eligibility assessment, and 93 were excluded due to ineligible populations (n=36), intervention (n=12), outcomes (n=22), and study design (n=23). Of the 14 records identified through other methods, 12 were excluded owing to ineligible population (n=8) or ineligible study design (n=4). Ultimately, 14 studies were included in this review (see Figure 1).

The characteristics of the included studies are summarized in Table 1. Among the 14 included RCTs, 78.6% (n=11) of the studies enrolled older patients with an average age of ≥60 years [38]. Twelve studies were 2-arm parallel RCTs [11,12,16,17,20,39-45], one study used a 3-arm design [13], and another used a 4-arm design [46]. Geographically, studies were conducted between 2020 and 2026, including China (n=6) [11,17,20,39-41], Germany (n=2) [13,43], South Korea (n=2) [16,45], Japan (n=1) [12], Spain (n=1) [44], Australia (n=1) [42], and England (n=1) [46]. A total of 837 patients were included across all studies, with sample sizes ranging from 9 to 126 participants, encompassing individuals with colorectal cancer (n=7) [11,13,20,42-44,46], liver cancer (n=3) [16,17,41] and esophageal cancer (n=1) [39]. Additionally, 3 studies included mixed populations of patients with gastrointestinal cancer, including colorectal, esophageal, gastric, pancreatic, biliary tract, and liver cancers [12,40,45].

IVR is primarily used in perioperative and chemotherapy settings for distraction, exposure therapy, rehabilitation training, knowledge education, and cognitive behavioral therapy. To implement the interventions, 12 studies (85.7%) used VR immersive scenes (eg, nature and ocean) [11-13,16,17,39-45], 2 used VR interactive games [13,43], 3 made use of VR anatomical models [16,42,46], and 2 incorporated VR videos [16,44]. One study delivered a structured VR-based cognitive behavioral therapy program integrating cognitive reconstruction and behavioral techniques [20]. Moreover, 6 of the 14 (42.9%) studies combined 2 or more VR interventions [13,16,20,42-44]. Interventions were administered across diverse clinical settings: 8 of the 14 (57.1%) studies were conducted in hospital wards [13,16,39,40,43-46]; interventions were administered during minor procedures under local anesthesia in operating theaters, such as implantable port placement, in 3 studies [12,17,41]; in an outpatient department in 1 study [42]; in a dedicated VR experience room in 1 study [11]; and without restriction on setting in 1 study [20]. These interventions were delivered by health care providers, with 11 studies (78.6%) involving physicians [12,13,16,17,40-46] and 3 (21.4%) involving nurses [11,20,39]. Interventions were primarily implemented at 3 time points: before treatment (eg, surgery, radiotherapy, and chemotherapy; 7/14, 50%) [11,13,16,40,42,44,46], during treatment (4/14, 28.6%) [12,17,39,41], and after treatment (2/14, 14.3%) [43,45], and 1 study spanned the entire perioperative period from presurgery to 6 weeks post discharge [20]. The duration and frequency of interventions exhibited significant differences, ranging from 7 studies involving single treatment sessions [12,16,17,41,42,45,46] to 7 studies using systematic programs lasting up to 2 months, with multiple weekly sessions [11,13,20,39,40,43,44].

The results of the risk of bias assessment are presented in Figure 2. Among the 14 included studies, 3 (21.4%) demonstrated a high overall risk of bias [12,39,44], 8 (57.1%) were rated as having some concerns [11,13,16,17,20,40,41,45], and 3 (21.4%) studies had a low overall risk of bias [42,43,46]. Six of the 14 (42.9%) studies had some concerns regarding the randomization process (domain 1) because information on allocation concealment was not provided [11,17,39-41,44]. Owing to the nature of the interventions, multiple studies did not specify whether blinding was applied to either the intervention providers or participants. Two studies implemented blinding for staff responsible for patient treatment and care, as well as for those assessing outcomes [20,43]. Another study had the control group view placebo scenes through VR headsets [40]. One study excluded intervention providers from VR usage, with rules being conveyed solely through standard verbal instructions [42]. The participant attrition in 3 studies might raise some concerns about the risk of bias due to potential deviation from the intended intervention (domain 2) [13,40], but 2 of them used an intention-to-treat analysis to address the missing data [13,20]. Additionally, 1 study did not mention whether deviations from the intended intervention occurred due to factors inherent to the experimental environment, raising some concerns about the risk of bias due to potential deviation from the intended intervention (domain 2) [39]. Three studies presented some concern about the risk of bias in outcome measurement (domain 4) because some self-reporting tools were administered by health care professionals [16,17,41]. Three studies presented some concern about the risk of bias in outcome measurement (domain 4) because they used self-report measures, unlike the remaining studies that used objective measures [12,39,44].

The overall certainty of evidence for each outcome, evaluated using the GRADE framework, is summarized in Table 2. Evidence of moderate certainty was found for the effects on knowledge, length of stay, systolic blood pressure, and heart rate. Ratings were downgraded primarily due to serious imprecision (wide CIs and/or limited sample sizes). Evidence of low certainty was identified for anxiety, pain, and diastolic blood pressure, with downgrades due to serious risk of bias, serious inconsistency, or serious imprecision. Evidence for quality of life was of very low certainty, reflecting serious concerns in risk of bias, inconsistency, and imprecision.

Given the considerable heterogeneity in the effects of IVR interventions on anxiety and pain among patients with gastrointestinal cancer, a series of subgroup analyses and meta-regressions were conducted based on intervention design (Table 3). Regarding anxiety, the time point of the intervention significantly moderated the effect size (between-subgroup P=.001; meta-regression P=.03). Studies delivering the intervention during treatment yielded a large and statistically significant reduction in anxiety (SMD −0.95, 95% CI −1.32 to −0.46; P=.65 within subgroup; I²=0%), whereas those conducted before treatment showed a smaller, nonsignificant effect (SMD −0.26, 95% CI −0.91 to 0.39; P=.20; I²=38%). A similar situation emerged for intervention duration. Longer interventions lasting ≥20 minutes were associated with a greater anxiolytic effect (SMD −0.95, 95% CI −1.53 to −0.37; I²=0%) compared with shorter interventions (SMD −0.38, 95% CI −0.92 to 0.16; I²=58.5%; between-subgroup P=.008; meta-regression P=.02). Furthermore, the clinical setting also influenced anxiety outcomes (between-subgroup P=.02), with the most pronounced benefit observed in operating room settings (SMD −0.95, 95% CI −1.53 to −0.37; I²=0%). In contrast, the number of intervention sessions did not significantly moderate the effect on anxiety (between-subgroup P=.64; meta-regression P=.75). However, with regard to pain and diastolic blood pressure, none of the interventions included in the analysis showed a significant difference in outcomes (P>.05). The bubble plots are presented in Figures 9 and 10, and the forest plots are provided in Multimedia Appendix 2.

Heterogeneity among the included studies may have influenced the overall specificity and sensitivity of IVR-based interventions within the internal validation dataset. Using the leave-one-out method, we found the studies to exhibit relatively good stability. We explored sources of heterogeneity through subgroup analyses and meta-regression, ultimately identifying intervention timing and duration as potential factors influencing IVR’s efficacy in alleviating patient anxiety. Sensitivity analyses, conducted by sequentially omitting individual studies, revealed that the pooled effects across most iterations remained statistically significant, demonstrating relatively high degrees of robustness. For further details, please refer to Multimedia Appendix 2.

This systematic review and meta-analysis synthesized evidence from 14 RCTs involving 837 patients to evaluate the effectiveness of IVR interventions for symptom management in patients with gastrointestinal cancer. The pooled analyses demonstrated that IVR significantly reduced anxiety and pain, which were 2 core symptoms targeted by symptom management interventions. In addition, IVR was associated with a significant reduction in length of hospital stay. No statistically significant effects were observed for quality of life, symptom management knowledge, or vital signs. However, the interpretation of these findings requires careful consideration of the underlying heterogeneity, risk of bias, and certainty of evidence. Specifically, while the 95% CIs indicated that the average effect of IVR on anxiety, pain, and length of stay was statistically significant, the corresponding 95% prediction intervals were wide and included the null value for all 3 outcomes. This discrepancy underscores that the true effect of IVR in a future clinical setting may vary substantially, depending on contextual factors such as intervention timing, session duration, and patient characteristics. Furthermore, the GRADE certainty of evidence was rated as low for anxiety and pain and very low for quality of life, with downgrades primarily attributable to serious risk of bias, inconsistency, and imprecision. Consequently, these findings should be interpreted as promising but preliminary.

The significant reductions in anxiety and pain observed in this review are broadly consistent with prior systematic reviews examining VR-based interventions in cancer populations [18,19]. For instance, Wu et al [18] reported moderate effects of VR on both anxiety and pain across various cancer types. The present review extends these findings by demonstrating that IVR remains effective specifically in the gastrointestinal cancer population, a group that faces unique symptom challenges including visceral pain and heightened situational anxiety during procedures such as transarterial chemoembolization [47,48].

The subgroup analyses indicates that the time point and duration of intervention are critical factors influencing anxiety levels. Interventions implemented during treatment, such as local anesthesia surgery [12] or chemotherapy [17,41], were associated with a large and statistically significant reduction in anxiety. This situation aligns with the attentional mechanisms proposed by Usman et al [49

] and may be particularly relevant for patients undergoing transarterial chemoembolization. During transarterial chemoembolization, patients remain conscious throughout a protracted sequence of invasive maneuvers, such as vascular puncture, which collectively drive anxiety to peak levels during the procedure itself [47,48]. In this context, highly immersive VR engages multiple sensory channels simultaneously, diverting attentional resources away from noxious stimuli and psychologically transporting the patient from the stressful treatment environment into a controlled, calming virtual space [19]. By directly interrupting the ongoing stressors, IVR exerts a more immediate and potent emotional regulatory effect than interventions delivered prior to treatment [50]. Pretreatment applications, such as simulated procedural walkthroughs [44] or knowledge-based education [16], primarily address the cognitive dimension of anticipatory anxiety. They struggle to fully replicate the complex multisensory stimuli of the actual treatment setting, such as equipment noise, physical discomfort, and antiseptic odors, and the resulting discrepancy between expectation and reality may paradoxically heighten anxiety once the procedure commences [51]. With respect to intervention duration, sessions lasting 20 minutes or longer produced significantly greater anxiety reduction than shorter sessions. This finding is consistent with theoretical models positing a neural adaptation window of approximately 20 minutes for transitioning from an aroused sympathetic state to a parasympathetically mediated relaxed state [6,49,52]. A longer duration permits patients to fully acclimatize to the virtual environment, establish a sense of presence, and engage deeply with the intervention content, which is particularly crucial for adult patients [53]. Nevertheless, the benefits of prolonged exposure appear subject to diminishing returns. Evidence suggests that beyond 30 minutes per session, additional anxiolytic gains plateau and may even reverse due to visual fatigue or physical discomfort associated with prolonged headset wear [52,54]. Although this systematic review could not determine the optimal session frequency, related evidence suggests that training 3 to 5 times weekly may enhance patients’ ability to apply relaxation techniques learned in virtual environments, thereby promoting skill consolidation and sustained symptom self-management [52,54]. This hypothesis warrants prospective validation in future trials.

With respect to pain, neither subgroup analysis nor meta-regression identified statistically significant moderators of the IVR effect. Nevertheless, a sensitivity analysis pointed to a likely source of the observed heterogeneity. Exclusion of the study by Zhao et al [39] reduced τ² from 0.16 to 0.01 and I² from 76.2% to 9%. This study differed notably from the others in both intervention duration and pain phenotype. Zhao et al [39] evaluated a 2-month multimodal IVR program designed to manage chronic pain and associated symptoms during chemotherapy for esophageal cancer. In contrast, the remaining studies examined single-session, short-term IVR protocols targeting acute procedural pain during surgery or localized interventions [12,17,41]. This situation suggests that IVR may operate through distinct mechanisms in acute versus chronic pain contexts. Acute procedural pain relief likely depends on immediate attentional distraction, whereas chronic cancer pain, which frequently involves central sensitization and psychological comorbidity [55], may require extended multimodal interventions that foster sustained emotional regulation and enhanced self-efficacy. Whether these differential effects account for the apparent inconsistency with prior reviews reporting VR benefits predominantly in acute pain settings [56] warrants further investigation.

With respect to symptom management knowledge, the pooled analysis did not reveal a statistically significant benefit of IVR, which contrasts with studies of mobile health and telemedicine interventions that have demonstrated significant improvements in disease-related knowledge among patients with cancer [57,58]. This discrepancy may reflect a difference in intervention design focus. Unlike digital health platforms specifically engineered for structured knowledge transfer, the IVR interventions included in this review primarily emphasized immersive relaxation and attention distraction [51]. The experiential, rather than instructional, orientation of these applications may limit their ability to convey factual information about symptom identification, self-monitoring, and coping strategies [59]. Nevertheless, the wide CI and moderate GRADE certainty suggest that a potential benefit cannot be excluded. Whether integrating educational modules within immersive symptom management environments enhances knowledge acquisition merits further study.

This meta-analysis indicates that perioperative IVR interventions exert no significant immediate effect on vital signs in patients with gastrointestinal cancer, consistent with findings from prior studies [60]. Notably, stable vital signs are frequently regarded as a physiological reflection of the patient’s psychological equilibrium. Multiple studies indicate that IVR aids in maintaining intraoperative vital sign stability, reflecting both its role in emotional regulation [6,61] and providing evidence for its safety as an adjunctive intraoperative intervention [17]. Consequently, despite failing to achieve statistical significance, IVR’s performance in sustaining vital sign stability supports its feasibility and safety for perioperative application.

The pooled reduction in length of hospital stay observed in this review is consistent with the hypothesis that effective symptom management may accelerate postoperative recovery and facilitate earlier discharge [46,55]. Comparable reductions in hospitalization have been reported in reviews of VR-based rehabilitation in surgical populations [43]. However, this finding must be interpreted with considerable caution. The 95% prediction interval was exceptionally wide, reflecting substantial uncertainty about the expected effect in different health care settings or patient populations. Moreover, the GRADE certainty of evidence was downgraded to moderate due to serious imprecision arising from the limited cumulative sample size. While the observed effect is clinically promising, it should be viewed as hypothesis generating rather than definitive.

This meta-analysis indicates that the efficacy of IVR interventions in improving patient quality of life remains inconclusive. This uncertainty likely stems primarily from methodological limitations in the design and reporting of end point measures across existing studies. Patient quality of life constitutes a multidimensional long-term end point, with significant improvements typically requiring cumulative temporal effects [62]. However, the studies included in this analysis predominantly used single-session or short-term intervention protocols and systematically lacked comprehensive long-term follow-up data. Consequently, the existing evidence base may be insufficient to capture the potential long-term impact of IVR on quality of life. These findings echo those of Hao et al [15], whose umbrella review noted that evidence for VR effects on quality of life in cancer populations is sparse and methodologically weak.

Regarding safety, our research indicates that IVR was well tolerated by patients, contrasting with concerns raised by Tian et al [63]. This discrepancy may stem from our included studies primarily relying on patient self-reported adverse reactions. The majority of enrolled patients were aged over 60 years, and susceptibility to motion sickness typically diminishes with advancing age [18]. One study analyzed patients’ motion sickness responses by measuring blood pressure and heart rate [17], echoing findings by Chattha et al [64] and implying a potential correlation between physiological variables and motion sickness.

Several limitations should be considered in this meta-analysis. First, the number of included studies was relatively small for certain outcomes (eg, knowledge), which may limit the statistical power and generalizability of these findings. Second, the overall risk of bias was rated as high or of some concern in more than half of the included studies, particularly in the domains of randomization and blinding, which may affect the validity of the pooled estimates. Third, certain clinical and methodological heterogeneity across studies, along with the low to moderate certainty of evidence for most outcomes as assessed by GRADE, suggests that the results should be interpreted with caution and that firmer conclusions await more consistent and precise data. Finally, as only articles published in English and Chinese were included, language bias cannot be entirely ruled out. These limitations highlight the need for larger, more rigorously designed RCTs with standardized intervention protocols and outcome measures to further validate and extend our findings.

This systematic review and meta-analysis provides evidence that IVR is an effective nonpharmacological adjunct for symptom management in patients with gastrointestinal cancer, significantly reducing anxiety and pain as 2 core symptoms in this population and potentially shortening hospital stay. Subgroup analyses indicate that these benefits are maximized when IVR is administered during active treatment and for session durations of at least 20 minutes, offering actionable parameters for clinical implementation. However, the GRADE certainty of evidence was rated as low for anxiety and pain, and the wide 95% prediction intervals underscore substantial uncertainty about effect magnitudes across different clinical settings. Consequently, these findings should be interpreted as promising but preliminary. Future research should prioritize well-powered, multicenter trials with standardized protocols and longer-term follow-up to confirm the durability of symptom management benefits and validate optimal intervention parameters.