This systematic review will include studies that meet the following inclusion and exclusion criteria:

(1) Studies recruiting participants with social anxiety disorder were eligible for inclusion (No restrictions were imposed on age, gender, ethnicity, nationality, educational attainment, or socioeconomic background). (2) Studies were included if participants’ social anxiety was assessed using recognised and validated social anxiety instruments, such as the Liebowitz Social Anxiety Scale (LSAS), Public Speaking Anxiety Scale (PSAS), or Social Interaction Anxiety Scale (SIAS). (3) Studies may be included if recruited participants possess the physical capacity to complete interventions and assessments. Furthermore, participants with severe organic pathology, functional impairment, comorbid mental disorders, or serious physical illnesses rendering them unable to continue participation, or those with incomplete clinical data, shall be excluded.

This systematic review focuses on SGVRET as the intervention of interest. SGVRET is operationally defined as a non-pharmacological digital intervention grounded in CBT and exposure principles, in which exposure exercises are delivered through virtual reality systems without real-time, in-session therapist guidance, and is typically self-administered individually, with eligibility determined by the absence of real-time therapist involvement (34). SGVRET is typically delivered via immersive or semi-immersive virtual reality systems that present automated or pre-programmed exposure scenarios, with exposure progression controlled by the system or the participant. Participant interaction is primarily mediated through automated prompts or predefined task structures. Minimal therapist involvement outside the exposure process (e.g., initial guidance, technical setup, or post-session feedback) is permitted, provided that no real-time therapeutic guidance is delivered during the exposure sessions. Accordingly, this review will include studies in which SGVRET constitutes the primary intervention and exclude studies that do not employ a self-guided virtual reality exposure framework as defined above.

This systematic review will include studies comparing SGVRET with non-SGVRET conditions (such as VRET or CBT). Additionally, studies comparing combined interventions (such as SGVRET combined with conventional CBT) with alternative therapies used alone (such as VRET alone) will also be included.

This systematic review will include studies reporting the following primary and secondary outcomes. Primary outcome measures primarily concern changes in the severity of social anxiety symptoms before and after the intervention. These encompass measurements using validated continuous self-report scales (e.g., LSAS, Libowitz Social Anxiety Scale; PSAS, Public Speaking Anxiety Scale; BFNES, Brief Fear of Negative Evaluation Scale). Additionally, secondary results will encompass changes from baseline to post-intervention and follow-up time points. These alterations include specific domains of social anxiety, such as alongside measurements from scales including the Hospital Anxiety and Depression Scale (HADS) and Patient Health Questionnaire-9 (PHQ-9).

This systematic review will only include peer-reviewed randomised controlled trials (RCTs). Non-randomised controlled trials, such as observational studies, cross-sectional studies, case–control studies, and qualitative studies, will be excluded.

This study implemented a comprehensive literature screening and evaluation process to ensure research objectivity and accuracy. The process will be conducted by three independent reviewers: YYL, JFS, and GHH. Initially, YYL will be responsible for downloading and conducting an initial screening of the literature, excluding studies irrelevant to the research topic. Subsequently, the preliminarily selected relevant literature will be submitted to JFS for a more detailed eligibility assessment. The literature ultimately included in the analysis will be reviewed by both YYL and JFS. Where the two reviewers disagreed, the third reviewer (GHH) made the final decision. Only literature unanimously approved by all three reviewers will be included in the final analysis.

This review will assess the full texts of potentially eligible studies to determine the final studies for inclusion. The study screening process will be recorded using a PRISMA 2020 flow diagram. The PRISMA flow diagram is provided in a planned format to illustrate the intended identification, screening, and inclusion process, and study counts will be completed following the literature search and screening (35). The study flowchart is shown in Figure 2.

This figure illustrates the intended study identification, screening, and inclusion process.

Study counts will be populated after completion of the literature search and screening.

Two independent reviewers will use a standardised data extraction form to extract data. A predefined and standardised codebook will be developed to guide the data extraction process, clearly defining each variable and corresponding coding rules to ensure consistency and reproducibility across reviewers (Table 2). During the data extraction process, researchers YYL and JFS will follow the PICO (S) framework, a structured approach for identifying and evaluating the components of clinical evidence in research articles. Specifically, researchers will compile a summary table of study characteristics detailing key information from all included studies, structured as shown in Table 3. This table will encompass the following fundamental elements: (1) Study details (first author, year of publication, country/region of study), (2) Population characteristics (mean age, sample size), (3) Intervention and control (type of intervention, duration of intervention, type of control group), (4) Primary outcomes (measurement tools and primary research findings). In addition to the study characteristics summary, researchers will plan a separate outcome measures summary table to systematically outline all predefined primary and secondary outcomes, their corresponding measurement instruments, and expected assessment time points (Supplementary material 2). When multiple validated social anxiety measures are reported within the same study, selective outcome extraction will be minimised through the application of predefined and consistently applied prioritisation rules. Prioritisation is guided by construct relevance to SGVRET and established psychometric robustness. Specifically, the PSAS will be prioritised in studies primarily targeting public speaking anxiety, given its comprehensive coverage of cognitive, behavioural, and physiological anxiety components and strong validity (36). For overall social anxiety severity, the LSAS will be prioritised due to its direct assessment of fear and avoidance across socially evaluative situations that closely align with exposure-based and virtual reality intervention mechanisms (37). When measures primarily reflect interaction-based or evaluative social anxiety, the Brief Fear of Negative Evaluation Scale (BFNES) or the Social Interaction Anxiety Scale (SIAS) will be selected based on their established reliability and construct validity (38). Where more than one eligible scale is reported, a single outcome measure will be selected according to these predefined criteria to avoid selective outcome extraction bias.

During data extraction, if outcome data required for synthesis are missing, unclear, or inconsistently reported, the corresponding authors will be contacted to request clarification or additional information.

Only data explicitly reported in the original publications or obtained directly from study authors will be included in the quantitative synthesis. Studies for which essential outcome data cannot be obtained will be retained for qualitative description only. This approach enables us to systematically organise and evaluate the relevant information extracted from the included documents, thereby enhancing the clarity and conciseness of our analysis. Additionally, any discrepancies that arise during the data extraction process will be resolved through discussion, and a third reviewer (GHH) will be consulted if necessary. The final extracted dataset will be securely stored for analysis.

To ensure the quality of the selected literature, this review will employ the critical appraisal tools of the Joanna Briggs Institute (JBI), the Cochrane Risk of Bias 2.0 tool, and the Grading of Recommendations, Assessment, Development and Evaluation (GRADE) system to assess the overall quality of the evidence (39–41).

The methodological quality of included randomised controlled trials will be assessed using the JBI Critical Appraisal Checklist for Randomised Controlled Trials. In the JBI evaluation, two reviewers (YYL and JFS) independently assessed the quality of the studies and resolved any differences by making consensus-based decisions or, if necessary, by discussion with a third reviewer (GHH). The scoring criteria will adopt the scoring system provided by the JBI guidelines to assess the quality of the included literature. The system assigns 1 point to each criterion that is fully met, with 1 point awarded for a “yes” answer and 0 points for a “no” or “unclear” answer. More than 80 percent of the total score is considered high quality, between 50 and 80 percent is considered medium quality and less than 50 percent is considered low quality. The appraisal results will be used to identify potential sources of bias and to inform the interpretation of findings. This systematic evaluation method ensures the reliability and scientific rigour of the review conclusions.

In the assessment of risk of bias, two reviewers (YYL and JFS) will independently evaluate each included study using the Cochrane Risk of Bias 2.0 tool. Discrepancies will be resolved through discussion, with consultation of a third reviewer (GHH) when necessary. Each study will be assessed across the five Risk of Bias domains, and judgements will be classified as “low risk of bias,” “some concerns,” or “high risk of bias” (42). An overall risk-of-bias judgement will be derived based on the domain-level assessments, in accordance with Risk of Bias guidance. Outcome measures will not be judged as high or low risk of bias solely on the basis of being subjective or objective. Instead, risk of bias in outcome measurement will be determined by factors such as blinding, susceptibility to influence by knowledge of the intervention, and consistency of outcome assessment across groups.

In the GRADE assessment, a comprehensive judgement is made from five dimensions: research bias, inconsistency, indirectness, imprecision and publication bias. If there are limitations, the evidence will be downgraded; if the results are consistent or the effect is significant, it can be upgraded. Ultimately, the evidence will be classified into four grades: high, medium, low, and very low (43). Two researchers (YYL and JFS) independently completed the scoring and calculated the K value to assess the consistency of the scoring, based on the Landis and Koch criteria: 0.00–0.20 indicates mild consistency, 0.21–0.40 indicates general consistency, 0.41–0.60 indicates moderate consistency, 0.61–0.80 indicates high consistency, and 0.81–1.00 indicates nearly perfect consistency. In case of any discrepancy, it shall be determined by the third researcher (GHH).

Importantly, studies scoring lower in JBI or GRADE assessments will not be excluded. We will explicitly highlight the methodological limitations of these studies within the systematic review and discuss their potential implications for the review’s conclusions.

The statistical analysis of this study will be conducted using Stata 18.0 software (44). For continuous variables, the weighted mean difference (WMD) will be calculated when measurement methods were consistent across studies, or the standardised mean difference (SMD) when different measurement scales were used. For dichotomous variables, odds ratios (ORs) with 95% confidence intervals (CIs) will be computed. Heterogeneity among studies was assessed using the I² statistic and Cochran’s Q test. Depending on the degree of heterogeneity, either a fixed-effect or random-effects model will be applied. Potential publication bias will be evaluated by Egger’s test and visual inspection of funnel plots. A p-value < 0.05 will be considered statistically significant.

The I² statistic will be used to evaluate the degree of heterogeneity among studies. According to the Cochrane Manual, when I² > 50%, significant heterogeneity is considered to exist (45). When p > 0.1 and I² < 50%, the fixed effects model is adopted; When p < 0.1 and I² > 50%, the random effects model will be adopted. If the data are sufficiently comparable, a meta-analysis will be conducted. If not comparable, the results will be reported using descriptive analysis methods.

In addition to statistical heterogeneity, potential sources of heterogeneity at the intervention and participant levels will be systematically examined. The following characteristics will be coded to capture intervention- and participant-level heterogeneity: (1) intervention duration, including total treatment length and cumulative exposure time; (2) type of VR system, categorised according to the technological configuration of the platform (e.g., immersive head-mounted display–based systems versus non-immersive or semi-immersive systems); (3) level of interactivity, classified as video-based exposure (passive viewing) versus interactive VR exposure involving user engagement or task-based interaction; and (4) participant characteristics, including baseline symptom severity and key demographic or clinical features when reported.

Intervention duration will be treated as a continuous variable when sufficient data are available, and categorised into higher versus lower exposure groups for subgroup analyses if necessary. VR system type and level of interactivity will be analysed as categorical variables. Participant characteristics will be examined to assess whether baseline clinical differences contribute to variability in treatment effects. Where sufficient studies are available, exploratory random-effects meta-regression analyses will be conducted for key continuous variables (e.g., intervention duration) to further investigate potential sources of heterogeneity.

In this study, when significant heterogeneity is detected, leave-one-out sensitivity analysis will be used to evaluate the impact of individual studies on the overall results (46). Each study will be sequentially removed, and the pooled estimates and heterogeneity statistics will be recalculated. This procedure will help identify potential outliers and assess the robustness and stability of the meta-analytic findings.

In this meta-analysis, when a certain outcome measure involves 10 or more studies, potential publication bias will be evaluated through a funnel plot (47). The visual symmetry of the plots will be inspected to identify potential small-study effects or selective publication tendencies. To provide a quantitative assessment, Egger’s regression test will be performed at a significance level of 0.05 to examine funnel plot asymmetry, thereby enhancing the objectivity and statistical rigour of the analysis.

When fewer than 10 studies are included, formal statistical testing will not be conducted owing to insufficient power. Instead, a qualitative assessment will be performed based on study characteristics, sample sizes, effect directions, and the overall consistency of the findings. This comprehensive approach will facilitate a more reliable evaluation of potential publication bias and strengthen the robustness of the meta-analytic conclusions.

However, it should be noted that the ability to detect publication bias may be limited when the number of included studies is small or when relevant interventions are underreported. Under such circumstances, funnel plot–based methods and statistical tests may have reduced sensitivity, and the results of publication bias assessments will therefore be interpreted with caution.