A protocol for this review was registered on the Open Science Framework [10], a widely used, publicly available platform that permits the open sharing of research data and material. This analysis is part of a larger evidence-synthesis project focusing on XR interventions for health and procedural anxiety.

A set of systematic reviews was identified, which included meta-analyses of randomized controlled trials relating to patient-directed XR interventions for health and procedural anxiety outcomes published from 2013 onward. In line with the panoramic meta-analysis approach, these reviews informed the sample of trials and effect sizes that were subsequently examined in generalized linear models (see the Statistical Methods section below). Specific inclusion and exclusion criteria are detailed in Table 1. Procedural anxiety was defined as anxiety both related and proximal to specific medical procedures, generally as a measure of state anxiety [11]. Health anxiety was defined as anxiety relating to broader health conditions, generally as a measure of trait anxiety [12]. Studies that based their analyses on physiological measures of anxiety were excluded (eg, blood pressure and heart rate), and data were only extracted that compared intervention outcomes against no-treatment or treatment-as-usual controls. Eligibility criteria did not impose any further stipulations for the trial-level data, and the extraction of effect sizes was not restricted to studies from any given language, date, or region.

Searches were carried out on May 30, 2023, in MEDLINE and Embase (on the Ovid platform), APA PsycINFO (in Ovid), and Epistemonikos. This variety of databases was selected to cover general medical literature, specialist psychology sources, and a collection of systematic reviews. Searches used appropriate subject headings and keywords for the intervention (virtual reality), outcome (health anxiety), and study design (systematic reviews), with full terms and methods specified in Multimedia Appendix 1. Records were screened independently and in duplicate, both at the title and abstract stage and at full-text stage, recording reasons for exclusion at full-text and involving a third reviewer in case of disagreement.

Informed by included reviews and relevant trials within them, a set of high-level indications were inductively generated as distinct “use cases” of XR. Different classifications were developed for procedural anxiety and general health anxiety. Within each indication and outcome, the highest quality, the most recent review was prioritized for effect size extraction from included forest plots, defining quality using the results of appraisal with the AMSTAR-2 (A Measurement Tool to Assess Systematic Reviews, Revised Instrument) [13]. The analysis then proceeded through the remaining reviews sequentially, organizing reviews by publication year and, within year, by quality, until no additional trials were identified, extracting evidence from randomized controlled trials only and preferring effect estimates for the same trial and outcome from higher-priority reviews. Where available, trial-level risk of bias (RoB) assessments were extracted, from included reviews providing sufficient evidence. Extraction items are provided in the protocol. The original trial was consulted when (1) extracted effect sizes were inappropriate or possibly inaccurate (eg, a standardized mean difference greater than 2) or (2) effect sizes were not presented as standardized mean differences and relevant standard deviations were not presented in included plots. One reviewer led extraction with auditing by a second.

All effect estimates were expressed as standardized mean differences, where negative values are positive (ie, greater reduction in anxiety). For each outcome, a 1-step meta-analysis model was estimated using a 3-level generalized linear model, where level 1 (patients within trials) is implied, level 2 is the study-level effect estimate, and level 3 is the indication. All models were estimated using restricted maximum likelihood and random effects at level 2 and level 3, with a common level 2 variance component across all indications to stabilize estimation. Heterogeneity was quantified using a standard variance partitioning method, generating a between-study, within-indication I², and a between-indication I² based on the Higgins and Thompson method. Statistical analyses were undertaken on Stata (version 18; StataCorp).

The registered protocol stated that for any instances containing too few indications in an outcome (fewer than 10), sensitivity analysis would be undertaken with a 2-level model, using fixed effects for indication at level 2. This model was compared with an unconditional meta-analysis model using a likelihood ratio test as an additional assessment of the value of indication as a grouping variable. Additionally, protocol-specified meta-regressions by age group (adult vs children), system immersion (immersive vs nonimmersive), and mechanism of action (relaxation or distraction vs exposure or education) were undertaken, where sufficient variation was present. Finally, funnel plots were drawn as an exploratory analysis, and an Egger test was estimated to examine small-study bias where the number of trials for any one indication and outcome was greater than 10.

A total of 18 meta-analyses informed the sample of trials, of which 10 provided procedural anxiety outcomes and 8 provided general health anxiety outcomes. Across these meta-analyses, 53 trial effects for procedural health anxiety outcomes and 30 trial effects for general health anxiety outcomes were examined. Two studies [14,15] reported data from multiple XR interventions, which were examined separately. Analyses excluded 15 trials that other meta-analyses included: 14 did not meet the specified definitions of XR, and one used physiological measures only. Study selection processes are outlined in Figure 1, via a PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flowchart (see full list of included reviews in Multimedia Appendix 2 [16-33]).

Details of study-level extraction, measures used, and interventions are provided in Multimedia Appendix 3. Reviews examined anxiety indications relating to either a specific medical procedure (n=8) or those additionally relating to general health conditions and prolonged treatment pathways (n=10). The average number of trials examined within each review was 6.67 (SD 3.38), although 18 overlapping trials were detected in multiple meta-analyses. Trial-specific RoB assessments or quality appraisals were reported in all but 3 of the included studies and are presented in Multimedia Appendix 4. These methodological evaluations followed inconsistent assessment criteria (eg, the Cochrane RoB, ROBINS-I, and Delphi List tools); however, only 8 trials were considered to have low RoB overall and 49 were deemed to have a high risk or some concerns. Methods for assessing anxiety notably varied between trials and were dependent on clinical indications. The most common measures were the Modified Dental Anxiety Scale (for dental studies), the State-Trait Anxiety Inventory (for cancer treatment, cardiac rehabilitation), the Modified Yale Preoperative Anxiety Scale (for nondental surgical operations), the Self-rating Anxiety Scale (for maternity), the Observational Scale of Behavioral Distress for radiographic procedures (for imaging), the Pain Anxiety Symptoms Scale (for chronic pain conditions), and simple visual analog scales (for wound care, needle-related anxieties).

Most of the included trial analyses (44/83, 53%) evaluated anxiety in pediatric patients only, while 36 trials focused on adults only, and 3 trials involved mixed-age samples (ie, children and adults). From an intervention perspective, the vast majority of procedural anxiety interventions (46/53, 55%) included just a single XR-based therapy session, the duration of which varied from 1 minute [34] to 90 minutes [35] (based on the context and population). Interventions for general health anxieties were more heterogeneous in design, ranging from single sessions to repeated activities over a period of months (Multimedia Appendix 3). Immersive forms of XR were adopted in 37 study interventions, whereas 46 interventions used nonimmersive methods. Interventions incorporated a range of XR technologies, including standalone virtual reality systems, portable glasses, mobile phone head-mounted displays, and monitor-based simulation devices.

In total, 66 interventions were classed as distraction-based or relaxation-based. These included natural scenery experiences (28 trial interventions), adventure or rollercoaster games (22 trial interventions), cartoons (12 trial interventions), mindfulness applications (1 trial intervention), and 3D images/videos (2 trial interventions). Conversely, 16 interventions were classified as education- or exposure-based. Separations between education- and exposure-based methods were not made, as numerous interventions incorporated elements of both components, and studies rarely provided sufficient details or rationale about XR procedures. This was also the case for distraction- and relaxation-based approaches (although it is also worth noting that all distraction experiences could be considered somewhat relaxing relative to anxiety-inducing medical stimuli).

Using an innovative statistical method, this analysis demonstrated that the effectiveness of XR interventions for both procedural and health anxiety is heterogeneous but conclusive; however, the evidence base for procedural anxiety is limited by publication bias, and effectiveness for procedural anxiety is moderated by age of participant and mechanism of effect. As a distinctive feature of panoramic meta-analysis, it was demonstrated that effects in procedural anxiety generalize across currently tested indications, while indication explains a relatively low amount of variance in general health anxiety outcomes; but that generalizability in a future trial is highly uncertain. Put otherwise, the evidence for the effectiveness of XR interventions in currently assessed indications cannot be used to assert that XR interventions would likely be effective for any given indication.

Although positive effects emerged across a range of general and procedural health anxieties, various indications remain poorly established. Indeed, most meta-analyses are based on context-dependent intervention data, which relates to specific medical procedures (eg, needle insertions and dental surgery) or long-term conditions (eg, cancer and cardiovascular disease). Panoramic meta-analysis methodologies were specifically undertaken to synthesize evidence from diverse clinical contexts and to establish the exchangeability of effects between different intervention types, technologies, populations, and indications. However, key questions remain about the effectiveness of XR within broader domains. For instance, studies are notably lacking in relation to imaging and diagnostic screening procedures (eg, endoscopies and mammography), where anxiety can have well-documented impacts on the quality of care [36-38]. Moreover, most current trials concern pediatric patient groups, despite health-related anxieties showing significant (and potentially growing) prevalence among adults [39,40]. Given the uncertain generalizability of study data that was highlighted in the present analyses, research needs to ascertain the best-suited clinical contexts for future XR interventions.

For the management of procedural anxieties, meta-regressions showed XR to be more effective in children (vs adults), and when using nonimmersive (as opposed to immersive) technologies. These findings imply that age and technology-related variables could influence the success of future XR programs. However, these findings did not replicate general health anxieties, indicating a role for broader, psychopathological factors in shaping effectiveness. For instance, positive outcomes may be more easily achieved when targeting relatively minor forms of anxiety (eg, a child’s worries about needles), compared with more complex and longstanding dispositions (eg, anxieties relating to major surgery or lifelong phobias). Relatedly, nonimmersive XR may be used for implementing simple, distraction-based interventions, whereas immersive devices may be required when attempting to administer more sophisticated or wide-ranging therapeutic solutions that are challenging to deliver successfully. When developing future applications, practitioners must consider the role of these interrelated components, since the value of key technological features (eg, fidelity and face validity) will depend on an intervention’s wider situational context.

Future research is needed to advance our mechanistic understanding of the empirical data. Here, studies should not just explore population- and technology-specific variables (eg, the age of patients, system immersion, and types of hardware), but also broader methodological components relating to how the intervention is delivered and how it actually “works”. Notably, the present results found no significant differences in the management of procedural anxiety between interventions using relaxation or distraction and interventions using education or exposure. However, there were a limited number of trials that examined exposure- and education-based XR methods in the evaluated review evidence, and there were insufficient data for analyzing general health anxiety outcomes. In addition, the included studies rarely provided sufficient details or rationale about their XR procedures and any cointerventions that were present, which limited capabilities for comparing different mechanisms of action. It is important that the contrasting psychophysiological processes implicated by different XR methods are considered during prospective assessments, as they are likely to impact on system requirements. For instance, a high degree of affective fidelity may be important for exposure therapies aiming to provoke “lifelike” stressors and psychophysiological responses; whereas aspects of user presence and task engagement may be more pertinent in simple distraction-based applications (see [41] for related framework). Nonetheless, without further empirical attention in this area, the relationships between different intervention types, psychophysiological mechanisms, and system features remain somewhat unclear. Overall, it is therefore vital that future work aims to address these “gaps” in theoretical understanding, so that the prospective design of XR programs can be sufficiently optimized for clinical implementation.

The use of panoramic meta-analysis meant that it was not possible to examine questions not generally addressed in single-indication meta-analyses. While the use of an overview of reviews was efficient in this analysis (and in keeping with the original implementation of this method), it is possible that very recently published trials not captured in existing systematic reviews were missed. In addition, this analysis was reliant on the quality of underpinning systematic reviews including search, extraction, and appraisal. None of the available reviews were rated as high-quality on AMSTAR-2, and many reviews were rated as critically low. Most trials were not rated as being at low RoB in their “parent” reviews. Relatedly, both the trials and reviews contained inconsistencies and inaccuracies in the reporting of study interventions (eg, in terms of describing levels of system immersion, or conceptualizing procedural and general health anxieties). Though included reviews were able to be traced to “saturation”, ensuring the identification of all relevant trials identified across reviews, it is possible that relevant trials were still missed.

While XR interventions are promising for both procedural and health anxiety, a number of unanswered questions remain. Future research should broaden the indications in which XR interventions are used, and should continue to develop emerging approaches, such as immersive technology interventions, to optimize their effectiveness while continuing to refine approaches for which effectiveness is well established.