This systematic review was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [36], the PRISMA-S (Extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews) guidelines [37], and the GRADE (Grading of Recommendations Assessment, Development, and Evaluation) approach [38]. The study protocol was preregistered in PROSPERO (CRD42024592600).

A systematic search of multiple databases was conducted (HS) on July 24, 2024, to identify randomized controlled trials (RCTs) of immersive VR-based interventions for stress reduction in the general population. The information sources searched included the databases MEDLINE PubMed, CINAHL EBSCOhost, the Cochrane Controlled Trials Register (CENTRAL), PsycInfo EBSCOhost, and Web of Science. Study registries, citation searching, or further resources were not considered. The search strings combined different keywords and text terms related to VR and stress, completed by validated filters [39] for RCTs. There were no further limits applied, for example, for date or language. Search strategies were peer reviewed by a second researcher (KM) regarding PRESS (Peer Review of Electronic Search Strategies) guidelines [40]. The search string used for MEDLINE is provided as an example in Table 1, and the complete search strategies for each database are provided in Multimedia Appendix 1. The search was updated on February 13, 2026, using the same strategy with publication date filters applied.

Inclusion was limited to RCTs that used either a within-subjects or between-subjects design and were published in English or German. Study protocols were excluded. Included studies were required to focus on human participants aged 18 years or older without any psychiatric diagnoses (eg, anxiety disorder, posttraumatic stress disorder, or depression) due to special demands regarding mental health. No further restrictions were applied regarding demographics such as gender, age, income, or intervention setting (eg, workplace or university, clinical or private setting). This review exclusively encompassed interventions based on VR with a primary focus on stress reduction, delivered through immersive technologies such as cave automatic virtual environments (CAVE) or head-mounted displays (HMD). In this context, a CAVE was defined as an enclosed room with virtual images projected onto the walls, floor, and ceiling. An HMD, in contrast, is a headset worn on the head, providing virtual images and sounds while shielding external stimuli. There was no further restriction regarding intervention content, for example, nature exposure, meditation task, or game. VR-based interventions with other aims than reducing PSL, such as stress inoculation training, pain management, physiological recovery, skill training, or the treatment of psychiatric disorders, were excluded. Furthermore, studies that used VR interventions for anxiety reduction during medical procedures, such as chemotherapy or dental surgery, were excluded to prevent potential biases caused by pain or medication effects. However, studies conducted in a clinical setting or among adults with a physical health condition (eg, cancer) were included if the intervention aimed at a reduction of PSL in general, independent of medical treatment, and no psychiatric diagnosis was involved. Further, inclusion required that the study design incorporated at least one neutral control condition or group, such as no intervention, treatment as usual, waitlist control, or a neutral environment (eg, urban or empty VR) as a placebo. To focus on the primary outcome of this review, only studies that provided PSLs, including arousal or negative affect, were considered. PSL had to be assessed using a validated questionnaire, a visual analog scale (VAS), or a numeric rating scale, with measurements taken at least postintervention. Secondary outcomes encompassed PSL in comparison to an active control group, as well as physiological stress parameters such as heart rate (HR), heart rate variability (HRV), skin conductance level (SCL), and systolic blood pressure (SYS). Secondary psychological outcomes included anxiety, depression, feelings of restoration, and positive emotions. Studies that were limited to physical stress outcomes were excluded.

All records identified through the literature search were first screened for duplicates using the online tool Rayyan [41], followed by a manual verification by one reviewer (HS) to ensure accuracy. After the removal of duplicates, two researchers (HS and IS) independently screened the titles and abstracts to exclude clearly irrelevant studies. Any disagreements were resolved through discussion. Next, full texts of the remaining articles were retrieved and assessed for eligibility based on the above-mentioned predefined eligibility criteria. This process was conducted independently by two researchers (HS and IS). Disagreements were resolved through discussion, and if consensus could not be reached, a third reviewer (KM) was consulted. If relevant information was missing to clearly apply the eligibility criteria, the authors of the respective studies were contacted to provide further details.

Relevant study information and data were extracted from eligible studies by 2 of 3 independent researchers (HS, IS, and KM) using a standardized spreadsheet form. To ensure accuracy, extracted data were compared and any discrepancies during the data extraction process were resolved through discussion. Extracted data items included (1) bibliographic information (first author, year, and country) and study design aspects (within-subjects or between-subjects, sample size, setting, and target population); (2) participant demographics, including age, gender, educational level, family status, race, health conditions, and prior VR experience; (3) numbers of participants in the intervention and control groups, as well as dropout numbers; (4) details of the VR intervention, such as conducted procedure, treatment duration, number and length of VR sessions, VR content and environment (provided relaxation technique, environment style and realism, content motion, technical aspects of the output device [device type, refresh rate, and field of view]), and level of user interactivity; (5) details of passive and, if applicable, active control conditions (procedure, treatment duration, intervention content); (6) details of measurements and analyses; and (7) reported outcome effects. The primary outcome variable of PSL, including negative affect or arousal, was addressed using a predefined hierarchical structure based on use frequency in the included studies to ensure standardized data extraction: Negative Affect Scale of Positive and Negative Affect Schedule (PANAS-NA), Profile of Mood States (POMS), State Scale of State-Trait Anxiety Inventory (STAI-S), Stress Scale of Depression, Anxiety, and Stress Scale (DASS-S), VAS for stress (VAS-S), Perceived Stress Scale (PSS), Warr’s Mood Scale (Warr’s) and miscellaneous measures. Secondary outcomes included physiological parameters of SCL, HR, HRV (high frequency component [HF-HRV], root mean square of successive differences [rMSSD], low frequency to high frequency ratio [LF/HF], and HRV response fraction), and SYS, as well as measures of restoration (Perceived Restorativeness Scale [PRS], Restoration Outcome Scale [ROS], and Recovery Experience Questionnaire [REQ]), positive emotion (positive affect scale of Positive and Negative Affect Schedule [PANAS-PA] and Zuckerman Inventory of Personal Reactions [ZIPERS]), anxiety (STAI-S, anxiety scale of Depression, Anxiety, and Stress Scale [DASS-A], anxiety subscale of Hospital Anxiety and Depression Scale [HADS-A], Generalized Anxiety Disorder [GAD] scale, VAS for anxiety [VAS-A], trait subscale of State-Trait Anxiety Inventory [STAI-T]), and depression (Self-Rating Depression Scale [SDS], depression subscale of Hospital Anxiety and Depression Scale [HADS-D], depression scale of Depression, Anxiety, and Stress Scale [DASS-D], depression subscale of Profile of Mood States [POMS-D], depression subscale of Warr’s Mood Scale [Warr’s-D], depression scale of Patient-Reported Outcomes Measurement Information System [PROMIS-D]). Additionally, measurements of perceived presence (IGroup Presence Questionnaire [IPQ], Presence Questionnaire [PQ], ITC-Sense of Presence Inventory [ITC-SOPI]), cybersickness (Simulator Cybersickness Questionnaire [SSQ]), and user experience with the VR intervention were considered. All effect measures were primarily extracted for postintervention time point in the form of mean and SD. If studies reported only standard errors, CIs, or IQRs, we calculated the SD according to methods described in the Cochrane Handbook [42]. In case of multisession interventions, the measure after the first session was extracted when reported. For between-subjects design studies that encompassed multiple relevant VR interventions (eg, an interactive VR game and a noninteractive VR exposure to nature), analyses were based on aggregated means and SDs across VR intervention groups to prevent double-counting of the neutral control. For within-subjects design studies, the VR intervention with the highest “dose,” such as the greatest level of biodiversity or stimulus intensity, was selected for inclusion. If relevant information on outcomes or intervention design was missing, the authors of the respective studies were contacted to obtain missing data. In case of nonresponse, studies with incomplete data on the PSL measure were excluded from the meta-analysis. Studies providing data on the primary outcome but limited information on intervention design or intervention characteristics were included in the main analyses. Further analyses only included studies providing the necessary information.

The risk of bias (RoB) was assessed by two independent researchers (HS and KM) for each included study using the Cochrane Collaboration’s Risk of Bias tool for randomized trials [42]. Each study was evaluated as having a low, high, or unclear RoB across the following domains: random sequence generation (selection bias), allocation concealment (selection bias), blinding of participants and personnel (performance bias), blinding of outcome assessment (detection bias), incomplete outcome data (attrition bias), and selective reporting (reporting bias). Any discrepancies between reviewers were resolved through discussion. An overall RoB judgment was not conducted.

The quality of evidence was assessed by two independent researchers (HS and KM) following the GRADE approach [38]. Therefore, the certainty of evidence was rated for each analyzed outcome (PSL, positive emotion, anxiety, depression, restoration, HR, HRV, SCL, and SYS) and categorized as high, moderate, low, or very low for the domains RoB, inconsistency, indirectness, imprecision, and publication bias. The overall certainty of evidence was determined by the GRADE classification system as high, moderate, low, or very low.

To investigate potential predictors and differences in stress-reducing effects based on variations in VR interventions, studies were categorized for the domain setting, intervention type, environment realism, environment style, content motion, and user interactivity. Descriptions of the domains and the applied labels are provided in Multimedia Appendix 2. Ladakis et al [14] synthesize in a scoping review that most studies on VR interventions for stress reduction focus on the following application fields: the general population, students or working adults, groups with high daily stress, and people with special demands in a clinical setting. Therefore, labels for the domain setting were determined as general population, workplace/university, and clinical setting. Environmental realism was poorly described in most studies. Therefore, a binary labeling in real-world recording or computer simulation was conducted. As a dichotomous division (yes or no) did not adequately capture the variety of the included studies in content motion as well as in user interactivity, an intermediate stage was implemented. Consequently, there were studies with no (simple 3D photos or simulations without animated elements), low (a static scene from a consistent perspective with some moving events like leaves in the wind), or high (a moving scene with changing perspective like on a pathway) content motion. User interactivity was also divided into 3 levels, based on Zhang et al [43] and Gras [44]: low control (only watching and exploring scene by head movement), medium control (option to change perspective while moving in the environment by controller or body movement), or high control (option to interact with elements due to an interactive task or game). The assignment of the included studies to the different labels of domains is provided in the Results section. If a study evaluated the characteristics of multiple relevant VR interventions, the effect sizes were pooled and included as one intervention in subgroup analyses. If intervention arms differed in methodological or technical aspects and categorization was not clearly applicable for a domain, the study was excluded from analyses to prevent double counting.

The primary meta-analysis was conducted and visualized by forest plots using Cochrane's Review Manager (RevMan), version 9.11.0. As PSL was measured differently by the included studies, for each study, standardized mean differences (SMDs) between intervention and control groups were calculated for the primary outcome of PSL after the intervention, along with corresponding 95% CIs. Due to substantial variation in intervention design, a random-effects model following the restricted maximum likelihood method was applied to pool the results. As the approach described by Hartung, Knapp, Sidik, and Jonkman (HKSJ) is proven to perform better when trials of similar size are combined and to result in more adequate error rates, the HKSJ correction was applied to calculate CIs. Secondary outcomes were also analyzed using random-effects meta-analyses of SMDs. Following Cohen’s recommendations [45], SMDs can be categorized as none (SMD<0.2), small (0.2≤SMD<0.5), medium (0.50≤SMD<0.80), or large (SMD≥0.80) effect.

Statistical precision of the average intervention effect was assessed by I² statistic. Further, heterogeneity and the practical robustness of observed intervention effects were explored by 95% prediction intervals (PIs). As the PI for the primary outcome of PSL was not significant, sensitivity analyses were conducted to identify possible subgroups with significant PIs and thus more robust results. For this purpose, studies were subdivided based on study design, including gender ratio (men or women dominance), age group (<40 years or ≥40 years), setting (clinical setting or general population), target group, and intervention duration (single- or multisession intervention). Further subgroups were formed based on features of the VR intervention, including the used technical device (HMD or CAVE), intervention type, environment style and realism, content motion, and level of user interactivity. The presence of reporting bias was explored by examining funnel plot asymmetry and conducting the Egger test in IBM SPSS Statistics, version 30.0. To assess the robustness of the findings and to evaluate the potential influence of the study design, a series of sensitivity analyses were conducted.

Potential predictors of the effectiveness of VR-based interventions for stress reduction were examined through meta-regressions, conducted in IBM SPSS Statistics, version 30.0, using the inverse-variance method. As the number of studies providing all data for an overall multivariate regression model was inadequate for the number of observed predictors, predictors were purpose-guided and allocated to two subsets defined by study design and technical or content-related VR characteristics for analyses. The assignment of the separate predictors for different meta-regressions is provided in Table 2 as well as the total number of included studies and participants. Only studies providing information on all observed predictors in each subset were included in analyses. Subsequently, all ordinal and continuous explanatory variables regarding study design (mean age of participants, percentage of men, intervention duration in days, number of VR sessions, single session length) were separately tested in univariate linear regression analyses (32 studies). As the chi-square statistic (Multimedia Appendix 3) revealed that a higher age of participants and multisession interventions were more likely in the clinical setting, the setting was dummy coded and integrated in regression analyses (clinical setting=yes/no). Further, multivariate regression analyses examining technical and content-related VR characteristics were conducted in a subset of 17 studies that provided sufficient data. As statistical power was low, the number of investigated predictors had to be limited. The chi-square test revealed a high correlation of the intervention type with the level of content motion and user interactivity, wherefor this categorial variable was not considered. Meta-regressions were visualized using bubble plots, generated in R Studio (version 2026.01.00). For all statistical tests, a P value of <.05 was considered the threshold for statistical significance.

The PRISMA flow diagram (Figure 1) provides an overview of the study selection process. The database search yielded 13,187 results, of which 4971 were identified as duplicates and removed. Following the screening of the titles and abstracts of the remaining 8217 papers, 296 articles were retrieved and screened in full text. In total, 55 studies met the predefined and in methods section described inclusion criteria and were included in the systematic review. Of these, 16 publications did not provide complete data (ie, mean and SD postintervention for both intervention and control group) for the primary outcome of PSL and were excluded from the meta-analysis (Table 3, MA=“no”), resulting in 39 studies included in the quantitative synthesis. Studies that seemed to fit at first glance but were excluded during full-text screening are provided in Multimedia Appendix 4.

Table 3 provides an overview of the main characteristics of the 55 studies included in this review [46-52,54-100]. Most studies were published within the past 5 years, while 8 were conducted between 2017 and 2019, and 3 prior to 2011. Among them, 43 studies used a between-subjects and 12 a within-subjects design. The 55 studies included one of the following neutral control conditions: 20 trials built/urban VR environment, 6 trials empty/blank VR control, 3 trials waitlist control, 23 trials treatment as usual/no intervention, and 2 trials other nonrelaxing tasks without VR. A total of 17 studies included an additional active control condition: 5 trials digital 2D simulations (2D), 3 trials real-world nature exposure (RWE), 6 trials traditional stress management (TSM) techniques, and 5 trials an alternative VR intervention that was not part of the main intervention focus. About 18 trials focused on the general population [49,53,56,61,64,65,68-70,79-81,84,86-88,92,96], 27 trials were conducted in work-related settings (2 in workplace, 23 in university, and 2 combined) [46-48,50,51,54,57,59,60,66,67,71,74-76,78,82,83,85,89-91,93,94,97,98,100], and 10 trials targeted subgroups with specific demands in clinical settings (eg, pregnant women, older adults, or clinical residents) [52,55,58,62,63,72,73,77,95,99].

Most studies (37 trials) implemented single-session interventions (Table 3). The remaining 18 studies involved multisession interventions lasting from 2 days to 6 months (1 trial 2 days, 1 trial 3 days, 1 trial 4 days, 2 trial 5 days, 1 trial 6 days, 3 trials 1 week, 1 trial 8 days, 3 trials 2 weeks, 1 trials 4 weeks, 1 trial 5 weeks, 1 trial 6 weeks, 1 trial 3 months, and 1 trial 6 months). The number of VR sessions ranged from 1 to 42 (37 trials 1 session, 9 trials 2‐5 sessions, 7 trials 6‐8 sessions, 2 trials more than 26 sessions). The duration of individual VR sessions ranged from 1 to 60 minutes (29 trials <10 min, 25 trials 10‐30 min, 1 trial>30 min).

A total of 31 studies investigated the effects of pure exposure to natural environments in VR (eg, forest, beach, or park) on stress reduction [46,47,49,52,53,56,58,61,64,65,70-73,77,79,80,82-84,87-90,92,93,97-100]. About 12 studies combined natural VR environments with a meditation task [51,55,57,59,62,63,78,85,86,94,95,100]. Another 9 studies studied the effect of biophilic design elements integrated into environments such as offices, streets, or courtyards in VR [50,54,66,68,73,75,76,81,96]. Five studies focused on games or interactive tasks as the main VR content [60,65,67,89,91], and 1 study [48] investigated miscellaneous content in VR. About 17 studies included more than one VR intervention to examine differences in factors such as environment type and biodiversity [49,53,54,66,68,69,72,73,76,80,82,92,96,97], meditation guidance [100], or level of user interactivity [65,89]. In nearly all studies (n=53), the VR intervention was delivered via HMD, while only 2 studies used a CAVE as an output device [49,61]. Around 28 studies used real-world recordings (photos or videos) as VR content [46-49,52,53,56,58,61,66,70-73,79,80,82-85,87,89,90,92,95,97-99], whereas 25 used a computer-generated simulation [50,51,54,55,57,59,60,62-65,67-69,75-78,81,86,88,93,94,96]. Two studies did not provide any technical details about the VR intervention.

This systematic review synthesizes evidence on the effectiveness of various VR-based interventions for stress reduction across a broad range of application areas and identifies characteristics that may predict this effectiveness. The meta-analyses demonstrated that VR-based interventions led to significantly greater reductions in PSL, anxiety, and depression as well as increases in positive affect, compared to neutral control conditions. However, the nonsignificant PIs for reductions in negative affect pointed to substantial heterogeneity, supporting the notion that VR interventions do not improve psychological stress in all individuals similarly. In contrast, significant PIs for positive affect and depressive symptoms suggest more consistent beneficial effects in these domains. Regarding physiological stress markers, no significant effects were observed for most parameters (eg, HR, HRV, or SCL). Nevertheless, a significant reduction was found for SYS, with significant PIs indicating a robust and consistent intervention effect. Beyond establishing the overall effectiveness of VR interventions in reducing PSL, this meta-analysis also identified several potential predictors of enhanced intervention outcomes. Specifically, multisession interventions, higher image refresh rates, and increased environmental realism, particularly through real-world recordings, were positively associated with stronger stress-reducing effects of VR interventions in univariate regression analyses.

Most of the included studies were published within the past 5 years, reflecting the rapidly growing interest in the use of VR for managing psychological stress. Among the various intervention types identified, virtual exposure to natural environments emerged as the most frequently investigated approach. These interventions were predominantly passive in nature, offering limited interactivity, although some incorporated additional elements such as guided meditation or interactive mini-games to enhance their stress-reducing effects. The majority of studies used HMDs as a delivery method for VR content, were conducted in university or workplace settings, and evaluated single-session interventions. In contrast, RCTs evaluating long-term interventions or targeting specific populations with specific needs, such as pregnant women or older adults, were less common and warrant further investigation. Overall, the findings of this systematic review highlight the considerable diversity of VR-based interventions for stress reduction, characterized by substantial heterogeneity in target populations, methodological and technical design, as well as in reported levels of immersion, presence, and user experience.

While previous qualitative literature reviews highlighted the general effectiveness of VR interventions for stress reduction [14,24,26-28,35], the present review extends those findings by quantifying intervention effects for stress reduction and statistically identifying possible predictors of intervention success. Chen et al [34] noted that existing meta-analyses have primarily focused on clinical populations, which motivated their investigation of virtual nature exposure and 2D images on anxiety, stress, and depression in healthy adults. In contrast, our analysis included a broader range of VR-based approaches, especially those delivered through immersive technologies. The meta-analysis of 39 RCTs confirmed that VR interventions can effectively reduce psychological stress in the general population. However, the wide and nonsignificant PIs indicate substantial variability in intervention effects and suggest that not all subgroups may benefit equally in practice. One possible explanation for this heterogeneity is the occurrence of cybersickness [101], which has been associated with increases in physiological stress markers (eg, cortisol levels and HR) [102], elevated psychological stress [103], and diminished intervention effects (eg, on anxiety) [104]. As cybersickness is likely influenced by characteristics of the VR content [105], it may contribute, together with the heterogeneity of intervention types, to the observed variability in PSL reduction. However, as only 4 of the included studies assessed cybersickness, its impact on intervention outcomes could not be systematically evaluated.

Univariate regression analyses further indicated that multisession interventions were associated with greater effectiveness compared to single-session interventions. This finding is consistent with previous research suggesting that longer-term interventions may be more beneficial for stress reduction [106]. One possible explanation is that studies using multisession designs more frequently targeted populations with elevated baseline stress levels, such as clinical samples [52,55,58,62,72,77] or individuals in workplace and university settings [46,48,57,59,74,100]. In contrast, single-session studies often involved experimentally induced, transient stress in younger participants [56,64,69,79,83,93]. This distinction is relevant, as existing evidence indicates that individuals with chronic conditions [107,108] and university students [109-111] are particularly susceptible to higher and more persistent stress levels. Moreover, while chronic stress typically requires repeated intervention exposure, acute stress may respond to brief relaxation techniques [112,113]. Accordingly, the optimal number, duration, and frequency of VR sessions likely depend on the nature of stress (chronic vs acute) and remain to be clarified by future research.

Notably, age and intervention setting were not identified as significant predictors in our regression analyses. Similarly, Li et al [29] reported no significant influence of age or session duration on the effectiveness of VR-based nature exposure on positive affect, while highlighting an overrepresentation of younger adults in the existing literature. Nevertheless, other studies suggest that older adults may represent a particularly promising target group for VR-based stress reduction [114], potentially benefiting even more than college students [115].

We also examined methodological and technical characteristics of VR content as potential predictors of VR intervention success. Our analyses indicated that the type of environment (eg, blue vs green space) did not significantly influence stress-reducing effects of VR interventions, consistent with the findings of the narrative review by Li et al [116]. In contrast, univariate regression analyses identified environmental realism (eg, computer simulation vs real-world video recordings) as a significant predictor, with stronger stress reduction observed for real-world recordings. However, this effect did not persist in multivariate analyses, possibly due to interdependencies with other factors such as intervention type and user interactivity. Although user interactivity did not emerge as a significant predictor, the observed effect size suggests that a more passive user experience may be associated with greater reductions in PSL. This is in line with findings by Reese et al [117], who reported stronger stress-reducing effects for passive VR nature exposure compared to active engagement within the VR environment. At the same time, the high heterogeneity of effects in studies involving active tasks, along with the limited representation of highly interactive interventions in our dataset, may have reduced statistical power and contributed to the lack of significance [118]. Moreover, previous research indicates that higher interactivity may enhance positive affect [97], suggesting that its impact may differ depending on the outcome domain (eg, negative vs positive emotions) as well as individual user preferences. Accordingly, the role of user interactivity and environmental realism warrants further investigation. The level of content motion was not identified as a significant predictor in our analyses, although individual studies have reported beneficial effects of dynamic content on physiological outcomes such as heart rate [119], as well as higher perceived presence for video-based compared to slideshow-based VR environments [120,121]. Moreover, this meta-analysis identified refresh rate as a promising predictor of stress-reducing effects of VR. However, the refresh rate considered in this meta-analysis reflects only the technically available frequency of image updates of the used HMD, whereas the actual rate at which images appear on the display (framerate) depends on the rendering of raw recordings into the VR content [122], for which insufficient information was available. As a higher refresh rate is generally associated with a better user experience, which is closely linked to interactivity and content motion [95,98,99], the observed interrelations between these variables are to be expected. It is therefore plausible that a smoother and more coherent user experience plays a crucial role for the effectiveness of VR-based stress interventions. Overall, given the substantial heterogeneity across studies, further research is needed to better understand predictors of intervention effects, particularly with regard to potential differences between chronic stress and acute mental load. Although the multivariate model showed significant predictive power, it did not fully account for the variability in outcomes, indicating that additional, as yet unidentified factors are likely to influence intervention success.

Significant effects of VR interventions were also observed for secondary outcomes, particularly increased positive affect and reductions in anxiety and depressive symptoms. Notably, the PIs for positive emotion and depression were significant, suggesting that these beneficial effects are relatively consistent across individuals. In contrast, nonsignificant PIs for other negative emotional outcomes (eg, PSL or anxiety) indicate that improvements in these domains may depend on additional factors. VR interventions were further associated with a significant reduction in SYS, with significant PIs indicating consistent effects across participants. These findings are in line with previous research, such as Yao et al [13], which demonstrated stress-reducing effects of VR interventions on both psychological and physiological parameters. Moreover, the beneficial impact of (virtual) nature exposure on blood pressure has been summarized in prior reviews [123,124]. While Song et al [124] also reported positive effects on HR and HRV, no significant effects on these or other physiological stress markers were observed in our analyses. Several factors may explain this discrepancy. First, the number of studies reporting physiological outcomes was relatively small, potentially limiting statistical power. Second, multisession interventions, which tend to yield stronger effects, were underrepresented in the available data. Third, substantial variability in the assessment and reporting of physiological measures, as well as potential limitations in measurement sensitivity, may have obscured detectable effects. The considerable heterogeneity in physiological outcomes is also reflected in previous studies [125,126], which attribute inconsistencies to differences in study design and population characteristics [126]. Overall, our findings are consistent with prior meta-analyses reporting reductions in PSL and increases in positive affect following VR interventions, alongside nonsignificant effects on cardiovascular parameters such as HR and HRV when compared to neutral control conditions [29,115].

Consistent with our finding of no substantial differences between VR interventions and active control conditions, such as 2D simulations or real-world nature exposure, previous studies suggest that VR-based nature exposure may be as effective as real-life exposure. For instance, Reese et al [127] reported similar improvements in well-being and PSL following both digital forest bathing and in vivo nature exposure. Likewise, the meta-analysis by Fan and Baharum [128] confirmed VR-based nature exposure as a viable and nearly equivalent alternative to real-world nature experience for stress reduction.

Several limitations should be considered when interpreting the results of this meta-analysis. First, no information scientist was involved in developing the search strategy, and relevant studies may therefore have been missed. However, the search followed a systematic and transparent strategy based on the Cochrane guidelines and the PICO framework, and the final search string was pretested and refined according to the PRESS guidelines by 2 independent researchers. Second, due to the nature of VR intervention studies, blinding of group allocation is rarely feasible, and only 3 studies implemented blinded control conditions. As a result, potential biases related to placebo effects or response bias cannot be fully excluded. However, evidence from VR-based pain management suggests that expectation effects may be limited in this context [129]. Third, many included studies exhibited a high risk of selection, performance, and detection bias, highlighting methodological limitations within the existing literature. Fourth, although strict inclusion criteria were applied to reduce heterogeneity, considerable variability between studies remained. Analyses could not identify more homogeneous subsets of studies with more robust effects regarding stress reduction through VR interventions. Prediction intervals were often nonsignificant, and I² values ranged from moderate to high, limiting the generalizability of the findings across populations. Fifth, the chi-square test indicated associations between several predictors, suggesting potential confounding. For example, multisession interventions were predominantly conducted in clinical settings and involved participants over 40 years of age. These interdependencies should be taken into account when interpreting the predictive value of related factors such as intervention duration in univariate analyses. Sixth, interventions involving active user engagement, such as interactive tasks or the use of controllers, remain underrepresented in the current literature, with the majority of interventions offering a passive VR experience [14]. Consequently, the stress-reducing potential of more interactive VR interventions cannot yet be adequately assessed. Seventh, limited reporting of participant characteristics (eg, prior VR experience) and insufficient detail on intervention features restricted the feasibility of conducting more robust subgroup and regression analyses. As a result, the statistical power of predictor analyses was limited and should be considered when interpreting these findings. Eighth, data on potential predictors were inconsistently reported, hampering a more comprehensive examination of their influence. It is therefore possible that additional unmeasured factors contribute to intervention effectiveness, which may explain the relatively low proportion of explained variance observed in the analyses.

While previous reviews have demonstrated the general effectiveness of VR interventions for stress reduction without quantifying their effects [14,24,26-28,35], or have focused on specific subgroups [32,130,131] or intervention types [29,30,128], this review highlights the considerable potential of VR-based interventions to reduce psychological stress in the general population. By quantifying effects across a broad range of approaches and, for the first time, directly comparing diverse intervention formats, this meta-analysis provides a more comprehensive understanding of their effectiveness. Given that VR-based interventions appear comparable in effectiveness to established stress management techniques, they represent a promising alternative, particularly in contexts where access to real-life nature or conventional interventions is limited or resource-intensive. While the findings suggest reliable improvements in positive affect and reductions in depressive symptoms, effects on PSLs were more variable, with some individuals experiencing limited benefit. This variability was observed in nearly all subgroups, suggesting that individual characteristics and contextual factors play a critical role in shaping intervention outcomes. Multisession interventions generally yielded stronger effects than single-session designs, whereas intervention duration and total number of VR sessions did not emerge as significant predictors. Thus, the optimal number, duration, and frequency of VR sessions remain unclear and may differ depending on whether stress is acute or chronic. Future studies with larger and more diverse samples are needed to directly compare effects across target groups and application contexts, including short-term stress induction and everyday stress experiences.

This meta-analysis further identified characteristics associated with a more calming user experience, such as higher refresh rates and lower levels of user interactivity, as potential predictors of stronger stress-reducing effects. These findings may inform the design of future VR approaches for stress reduction by highlighting the importance of a high-quality and coherent user experience. However, individual preferences regarding content motion and user interactivity in VR suggest that customizable VR environments may further enhance effectiveness. Advances in VR hardware and content development could facilitate such personalization, raising important questions about how to implement these as accessible, low-threshold tools for everyday stress management.

Despite these insights, a substantial proportion of the variance in intervention outcomes remains unexplained, indicating that the stress-reducing effects of VR are shaped by multiple interacting factors, including participant characteristics, intervention design, and technical features. Future RCTs should therefore systematically assess and report these variables to validate current findings, explore interdependencies, and establish evidence-based guidelines for optimizing VR-based stress reduction across diverse populations.