Anxiety disorders are the most prevalent mental disorders in children, adolescents, and young adults [1-3]. Especially in adolescence, social anxiety disorder (SAD) is particularly common, with prevalence rates ranging from 12% to 36% [4,5]. Notably, 88% of individuals with SAD experience onset before the age of 25 years and 50% before the age of 14 years [6]. The first symptoms (eg, shyness and fear of embarrassment) often appear in childhood [7,8]; and in adolescence, up to 50% of individuals report subclinical levels of social anxiety symptoms [9,10].

SAD is characterized by an intense and persistent fear of embarrassing oneself in front of others or being judged negatively by others [11]. To endure feared situations, those affected by SAD often engage in safety behaviors to minimize negative evaluations or avoid social situations altogether [11]. These behaviors maintain social anxiety symptoms [12-14]; contribute to substantial impairment in their psychosocial functioning (eg, fewer friends than their healthy peers) and the development of comorbid disorders (eg, depression and substance use); and lead to negative long-term effects, such as poorer academic and job opportunities and a lack of social support due to a small or nonexistent social network [15-20]. These detrimental consequences demand prevention efforts and effective treatments to intervene early.

Established treatment protocols (eg, those based on the cognitive model developed by Clark and Wells [12]) in adults have been adapted to suit the specific needs of children and adolescents [21]. Meta-analyses support their efficacy in individual and group settings compared to waitlist or psychological placebo control groups (ie, attention control) [22-24]. Accordingly, individual or group cognitive behavioral therapy (CBT) is recommended as a first-line treatment for children and adolescents with SAD [25,26].

Although effective treatments exist, treatment delivery and uptake are impeded by poor mental health literacy (eg, poor symptom recognition); public and self-stigma, particularly in individuals with SAD; and a gap in the provision of care [27-30]. The care gap is further exacerbated in children and adolescents, individuals from ethnic minority groups, those living in rural areas or lower-income countries, and those with lower socioeconomic status [31-33]. Various efforts have addressed some of these barriers by integrating care into primary physician settings, implementing psychoeducational interventions in educational settings, or using task-sharing approaches (eg, interventions delivered by nonprofessionals under supervision) [34-36]. Digital interventions are a highly scalable and far-reaching solution that may address barriers not addressed by previous efforts (ie, limitations related to time and location).

Digital mental health interventions (DMHIs) aim to manage, alleviate, or treat mental health problems through a digital medium (ie, internet, mobile phone app, wearables, or SMS text messaging) [37,38]. They can include some form of human support (ie, guided self-help), for instance, delivered through asynchronous messages, synchronous real-time chat, or telephone calls with a professional; or they can be stand-alone interventions (ie, unguided self-help) [39]. Several DMHIs have been developed for adults with SAD, with the evidence base for SAD being among the most extensive, including interventions based on CBT, psychodynamic therapy, interpersonal therapy, and acceptance-based CBT [39,40]. For this age group, meta-analytic evidence supports the efficacy of DMHIs: internet-delivered CBT (iCBT) and virtual reality exposure significantly outperformed passive or waitlist control conditions. Compared to other active control conditions, iCBT yielded small effects and was found to be as efficacious as face-to-face psychotherapy [41-43].

In children and adolescents, meta-analyses on the efficacy of DMHIs have focused predominantly on depressive and anxiety symptoms but have not reported effect sizes for either the prevention or the treatment of SAD specifically [37,38]. Although SAD outcomes were presented separately in 1 meta-analysis, no pooled effect size was calculated [44]. For DMHIs for anxiety disorders in general, meta-analyses reported significant effects for interventions based on CBT principles compared to a waitlist control condition (ages 6-16 years: standardized mean difference [SMD] 0.68; ages 7-18 years: Hedges g=1.41; and ages 6-25 years: Hedges g=0.68) and no significant difference when compared to other active conditions (eg, face-to-face therapy: Hedges g=0.30; SMD −0.04) [45-47]. For DMHIs for the prevention of anxiety disorders, meta-analyses did not yield significant overall effects, but some randomized controlled trials (RCTs) found small effects in favor of DMHIs [48,49]. Similarly, the meta-analyses of prevention programs for anxiety disorders in school-based settings reported small to medium effect sizes compared to passive control conditions after the intervention [50,51].

Notably, some studies suggest that young people with SAD benefit less from transdiagnostic interventions that target several anxiety disorders than young people with other anxiety disorders [52-54]. However, other studies found no difference between SAD-specific and transdiagnostic anxiety interventions for children and adolescents with SAD [55,56]. Thus, it would be relevant to compare the effects of DMHIs that specifically target SAD and those that target anxiety disorders in general. To the best of our knowledge, no systematic review and meta-analysis has investigated the effect of DMHIs on social anxiety in children and adolescents specifically.

DMHIs are not equally effective for all children, adolescents, and young adults. In the subsequent section, we discuss the potential moderators of treatment efficacy, which include control conditions, the psychological principles of the interventions, age, adherence, and the extent of support. Generally, the effect sizes are larger when DMHIs are compared with inactive control conditions (eg, waitlist) than with active control conditions [57,58]. Most DMHIs are based on CBT principles [37], although when compared to DMHIs based on other psychological interventions (eg, cognitive bias modification [CBM] and mindfulness), they do not result in better effects [59]. Evidence concerning age shows that older adolescents and young adults generally benefit more from DMHIs than children [46,60,61]. Similarly, adherence is supposed to be linked to symptom improvement, reflecting compliance with predefined and recommended use and engagement with the intervention [62-64]. However, the results regarding the association between adherence and outcome are more complex. Some studies point to a direct link between adherence and symptom improvement [65] and some to an inverse relationship [66], while others identify only single aspects of adherence (eg, module completion) to be relevant for symptom improvement [67]. Furthermore, studies indicate that adherence itself is, in turn, influenced by symptom severity, which may thereby affect symptom reduction [62]. Overall, adherence plays an important role in the implementation of DMHIs, and human support is a known factor in improving adherence [68,69]. This support can be provided in the form of guidance by professionals (ie, psychologists and therapists) or through the social environment of the young people (ie, parents and teachers).

DMHIs are especially promising in young people because they are digital natives and generally seek help later and to a lesser extent than adults [38,70,71]. Thus, the low-threshold and confidential nature of DMHIs may appeal especially to young people with SAD [30,72]. Accordingly, this systematic review and meta-analysis aims to evaluate the efficacy of DMHIs in reducing social anxiety symptoms in children, adolescents, and young adults.

Our primary research question is as follows:

The secondary research question includes moderator analyses:

In addition to evaluating efficacy, we will systematically synthesize and assess adherence to the DMHIs, focusing on how adherence is defined, measured, and reported across studies.

Two authors (NW and AF) independently extracted the data into a preformatted Excel file based on the Cochrane data collection form from the study by Higgins et al [80]. No automation tools were used. Information extracted from primary studies included the following: study characteristics (ie, country, ethics approval, trial preregistration, funding, study aims, study design, recruitment method, randomization procedure, justification of sample size, dropouts, and handling of missing values), sample characteristics (ie, informed consent procedure, eligibility criteria, sample size, sociodemographic information, and clinical status), intervention characteristics (description of the intervention; underlying psychological principles; duration, number of modules, and delivery setting of the intervention; extent of guidance; guidance provider and their training; and adherence to and satisfaction with the intervention), control condition characteristics (ie, waitlist; care as usual; treatment as usual; or another active intervention, including description and underlying psychological principles), and outcome data for all available measurement time points (ie, assessment tool, sample size, unstandardized means, SDs, and SEs if SDs were not available). If transdiagnostic anxiety measures were used, and no outcome data for SAD were reported, the corresponding authors were contacted.

Data analyses followed a practical guide for meta-analyses in mental health research [81] and were analyzed with Comprehensive Meta-Analysis software (version 4.0; Biostat, Inc) [82]. For each study, a standardized effect size was computed using Hedges g and 95% CIs derived from means, SDs, and sample sizes of the intervention and control groups. To account for potential heterogeneity, all analyses were computed with random-effects models [81]. In 3-armed RCTs, we prioritized comparisons with the waitlist control group over another active control group because there were more studies with passive comparison conditions, leading to a more homogeneous comparison group [83-86].

Heterogeneity was assessed through visual inspection of the forest plot, the Cochran Q test, and the calculation of the I² statistic. The Q test assesses whether the observed effect sizes significantly differ from each other beyond chance. A significant Q statistic indicates heterogeneity, but in small meta-analyses, there might be insufficient power to detect it [81]. I² values explain the proportion of heterogeneity, that is, the variability in treatment effect estimates due to real study differences and not due to chance, with thresholds of 25%, 50%, and 75% indicating low, moderate, and high heterogeneity, respectively [87]. In addition, we calculated tau and tau squared. Tau measures the SD of the true effects across studies, while tau squared is its square representing the estimated between-study variance. A significant tau-squared value suggests heterogeneity, supporting the use of random-effects models [88].

As high heterogeneity was expected, we explored possible sources of heterogeneity in moderator analyses. Moderators were explored in subgroup analyses if the moderator was categorical (ie, control condition, risk of bias, psychological principles, SAD-specific intervention, guidance, parental involvement, or age group) or in a meta-regression analysis if the moderator was continuous (ie, age or symptom severity). Subgroup analyses included the computation of average effect sizes and the Q statistic as an indicator of a common effect size. A significant Q statistic suggests heterogeneity in effect sizes within the subgroup.

Sensitivity analyses were performed to investigate the robustness of our findings. We conducted analyses excluding studies with the highest and lowest effect sizes to assess the potential influence of outliers on the meta-analytic results. This allowed us to evaluate the stability of the overall effect size when extreme values were omitted. In another sensitivity analysis, we only included studies in which all participants met diagnostic criteria for SAD to control for heterogeneous diagnostic samples. Moreover, in a subgroup analysis, studies were stratified based on their risk-of-bias ratings to investigate the influence of the methodological quality. In addition to these preregistered analyses, we conducted a sensitivity analysis by excluding studies with participants aged >25 years to test whether older participants altered the effects, and we carried out a moderator analysis with age groups such as children (<11 years), adolescents (11-17 years), and young adults (18-25 years) as moderators.

In addition, numbers needed to treat were computed for all effect sizes using the dmetar package in R with the default Kraemer and Kupfer method [89]. Numbers needed to treat indicate the number of participants to be treated to achieve 1 case with reliable symptom change.

To address a possible publication bias, we visually inspected the funnel plot for symmetry and conducted a Begg and Mazumdar rank correlation test deter and an Egger linear regression test to quantify asymmetry. Using the trim-and-fill approach proposed by Duval and Tweedie [90], we assessed missing studies in the funnel plot to estimate the effect size after imputation [81,91,92].

Risk of bias was assessed by 2 authors (NW and AF) independently following the revised Cochrane Risk of Bias Tool for Randomized Trials (RoB 2) procedure and using the RoB 2 Excel template [93]. After the individual rating, discrepancies between ratings were discussed until consensus was reached. The RoB 2 tool includes five domains for randomized trials: (1) bias arising from the randomization process, (2) bias due to deviations from intended interventions, (3) bias due to missing outcome data, (4) bias in measurement of the outcome, and (5) bias in selection of the reported result. Risk of bias was rated separately for each domain based on whether it was likely to have influenced study outcomes. Studies categorized as low overall risk of bias were rated as low risk of bias across all 5 domains. A rating of “some concerns” in at least 1 domain led to an overall judgment of “some concerns” for the study. Studies were rated as high risk of bias if at least 1 domain was rated as high risk [81].

The systematic literature search yielded 2149 records across all electronic databases, from which 474 (22.06%) duplicates were removed before screening. Of the remaining 1675 records, due to nonfulfillment of eligibility criteria, 1357 (81.01%) were excluded after title screening, and 205 (12.23%) were removed after abstract screening. Therefore, 113 full texts were assessed for eligibility, of which 22 (19.5%) were ultimately included in the systematic review and 21 (18.6%) in the meta-analysis. The study by Vigerland et al [94] was not included in the meta-analytic calculations because there were insufficient data related to social anxiety at the postintervention assessment. A complete overview of the study selection process is presented in Figure 1.

The included papers were published between 2011 and 2024. The 22 trials were conducted in Australia (n=7, 32%), the United States (n=4, 18%), Sweden (n=3, 14%), the United Kingdom (n=2, 9%), China (n=2, 9%), Spain (n=1, 5%), Iran (n=1, 5%), Canada (n=1, 5%), and the Netherlands (n=1, 5%). A detailed overview of study characteristics is presented in Tables 1 and 2.

All non–US currency values were converted to US dollars using exchange rates corresponding to the date of first manuscript submission for each study. Converted amounts are shown in parentheses.

Participants were recruited through various strategies, primarily using self-selected sampling procedures such as posters and flyers, word of mouth, social media and web-based advertisements, postings on platforms or forums used by parents and health care providers, and media and newsletter articles (13/22, 59%) [56,83,86,94,96,98,103-105,107-109,111]. The remaining RCTs (9/22, 41%) recruited participants via convenience sampling (ie, student research pool [99,106], school-based screening procedures [84,95,97,100,110], or an associated mental health clinic [102]). Only a few of the studies (2/22, 9%) mentioned efforts to ensure socioeconomic diversity, either by choosing schools representing a range of socioeconomic backgrounds [100] or by aligning the study cohort’s demographic profile with that of the general population [56].

In total, 4196 participants were randomly assigned across all RCTs. They ranged in age from 3 to 48 years, with mean ages spanning from 4.8 to 24.9 years. In all studies, more than half of the participants were female. A comprehensive compilation of participant-related characteristics, including ethnic diversity and socioeconomic information for the final samples, is presented in Tables 1 and 2.

The risk-of-bias assessment indicated that most of the studies (16/22, 73%) raised some concerns (Figure 2). A main concern was the fact that RCTs investigating psychological interventions did not facilitate adequate masking of participants to the treatment condition. Therefore, only studies that included another active intervention mimicking the initial intervention and in which participants were blinded to the condition [103] were rated as low risk in the bias in measurement of the outcome domain. Of the 22 studies, 3 (13.6%) were classified as high risk due to insufficient information on randomization; a lack of analysis estimating the effect of assignment to the intervention group; and high missingness, with absence of measures to address it [86,105,109]. Insufficient information on the randomization process (ie, lack of details on allocation sequence concealment or whether baseline differences between groups were assessed) was reported in 41% (9/22) of the studies. In 32% (7/22) of the studies, there was no information on possible trial deviation or estimation of an effect of assignment to the intervention. Most of the studies (19/22, 86%) followed the intention-to-treat approach and investigated their missing data. Of the 22 studies, 17 (77.3%) reported their plan of analysis, and the reporting of results was transparent and complete. The risk-of-bias assessments for all studies are presented in Multimedia Appendix 3 [56,83-86,94-106,108-111].

The random-effects model yielded an overall effect size in favor of DMHIs compared to any control condition at the postintervention assessment (Hedges g=0.508, 95% CI 0.308-0.707; Figure 3 [83-86,94-111]). Sensitivity analyses were performed (1) excluding studies with the highest [100] and lowest [84] effect sizes, (2) excluding studies with participants aged >25 years [83,86,105,106], and (3) including studies in which all participants had a diagnosis of SAD [86,95,99,100,103,107,111]. Two additional analyses were performed to examine between-group effect sizes at follow-up: one using data collected 3 to 6 months after the postintervention assessment and another using data collected 12 months after the postintervention assessment. A detailed overview is presented in Table 4.

In subgroup analyses, we investigated the moderating effect of the control condition, risk of bias, age group, underlying psychological principles, the focus of the intervention on SAD or transdiagnostic anxiety symptoms, guidance, and parental involvement. Except for the subgroup analysis comparing the effect of control conditions, both active and inactive conditions were included in all analyses. The effect size was significantly greater when the DMHI was compared to a waitlist (Hedges g=0.576, 95% CI 0.343-0.809) than to another active intervention (Hedges g=0.101, 95% CI 0.021-0.261). Furthermore, DMHIs based on CBT (Hedges g=0.610, 95% CI 0.361-0.859) were significantly more effective than DMHIs based on other therapeutic interventions (Hedges g=0.176, 95% CI −0.052 to 0.405; ie, CBM). DMHIs including therapeutic elements specifically addressing SAD (Hedges g=0.878, 95% CI 0.469-1.278) were significantly more effective than DMHIs addressing transdiagnostic anxiety symptoms (Hedges g=0.195, 95% CI 0.040-0.277). Guided self-help (Hedges g=0.825, 95% CI 0.425-1.224) was more efficacious than unguided self-help (Hedges g=0.271, 95% CI 0.083-0.458). Effect sizes differed significantly across age groups (children: Hedges g=0.185, 95% CI −0.008 to 0.379; adolescents: Hedges g=0.575, 95% CI 0.270-0.880; and young adults: Hedges g=0.578, 95% CI 0.276-0.879), with the smallest and nonsignificant effect observed in children. No significant difference was found for DMHIs with (Hedges g=0.281, 95% CI 0.117-0.445) or without (Hedges g=0.602, 95% CI 0.242-0.962) parental involvement. The computed effect sizes for different levels of risk of bias were not significantly different from each other (low risk: Hedges g=0.201, 95% CI −0.029 to 0.431; some concerns: Hedges g=0.622, 95% CI 0.359-0.885; and high risk: Hedges g=0.399, 95% CI −0.213 to 1.011); however, while the 95% prediction interval for the low-risk and some-concerns subgroup analyses were small, the one in the high-risk bias subgroup analysis spanned from –6.972 to 7.770. A complete overview is presented in Table 4, and all forest plots are included in Multimedia Appendix 4 [83-86,95-111].

Age and baseline symptom severity of social anxiety were introduced as moderators in the main analysis. The meta-regression indicated that age had no significant effect on the efficacy of DMHIs on social anxiety (𝛽=.038, 95% CI –0.003 to 0.079; P=.07). Higher baseline symptom severity increased the effect size significantly (β=.014, 95% CI 0.005-0.023; P=.002).

A visual inspection of the funnel plot suggested potential publication bias: the SEs were not symmetrically distributed around the overall mean (Figure 4). Both the Begg and Mazumdar rank correlation test and the Egger linear regression test of the intercept were significant (P<.001 for both), indicating an underlying publication bias. When adjusting for missing studies using the trim-and-fill approach proposed by Duval and Tweedie [90] in the random-effects model, the resulting effect size was Hedges g=0.506 (95% CI 0.308-0.707).

The meta-analysis evaluated the efficacy of DMHIs on SAD in young people. Across 21 RCTs and 4196 participants, we found a small to medium pooled effect size in favor of DMHIs compared to both inactive and active control conditions assessed after the intervention. The effect size remained after taking the publication bias into account. Notably, the treatment gains diminished by the 3- to 6-month follow-up and were no longer significant at the 12-month follow-up. Subgroup analyses revealed that DMHIs based on CBT, including SAD-specific elements, and supported by human guidance implemented in adolescents and young adults resulted in higher pooled effect sizes compared to other psychological interventions, transdiagnostic interventions for several anxiety disorders, and unguided self-help programs in children. Mean age and parental involvement had no significant effect on the outcome.

Compared to other meta-analyses of DMHIs addressing anxiety disorders, the overall effect size found in this meta-analysis was smaller. In a similarly aged sample but one with mixed anxiety disorders, a meta-analysis found a large pooled effect size (Hedges g=0.68) [45]. Even in children and adolescents at risk for an anxiety disorder (ie, subthreshold of diagnosis), where smaller effect sizes would be expected, a large pooled effect size (SMD 0.77) was found [44]. Possible reasons for the smaller effect compared to previous meta-analyses could be the large heterogeneity between the included studies regarding the interventions, the samples, and the study designs. Another reason could be the fact that despite the focus of this meta-analysis on samples with SAD, the primary studies included a broad range of mental health problems or disorders (eg, other anxiety disorders and depression). Consequently, study participants did not necessarily have social anxiety. In other meta-analyses, either a measure assessing various anxiety disorders was defined as the primary outcome, or the interventions and samples were primarily geared toward SAD. The sensitivity analysis in this meta-analysis showed that when only studies with participants with a SAD diagnosis were included, the calculated effect size was large (Hedges g=1.15). Meta-analytic evidence from face-to-face settings for children and adolescents diagnosed with SAD only also yielded a larger effect (Hedges g=0.71), which increased at follow-up [22]. Similar effect sizes [130] were found in iCBT for adults with SAD (Hedges g=0.55 [130]; Hedges g=0.76 [131]). In line with these findings, the moderator analyses showed that there are, in some cases, large differences between the effect sizes of the subgroups. These differences are discussed in the following subsections.

The superiority of CBT-based treatments over other treatments, such as internet-delivered CBM or mindfulness exercises, is also reflected in other studies [58,132]. However, most treatments for young people with SAD are based on CBT, and the evidence for other treatments is still limited [133]. Interestingly, SAD-specific treatments, including CBT and CBM, outperformed treatments addressing anxiety in general. This finding contradicts results from RCTs directly comparing 2 treatments (transdiagnostic anxiety treatment vs a SAD-specific approach [55,56]) but aligns with other evidence suggesting that adolescents with SAD have poorer outcomes than those with other anxiety disorders [53,54,134]. This finding is quite striking because DMHIs addressing SAD in adults have an extensive evidence base, with RCTs demonstrating efficacious interventions based on CBT, psychodynamic therapy, and acceptance-based interventions [39] and meta-analytic evidence supporting iCBT to be as efficacious as face-to-face CBT [42]. In young people, the question remains as to which treatment components have a generic effect in anxiety disorders and which would improve the treatment effect in SAD and thus indicate a disorder-specific treatment. Treatment components specific to SAD are theoretically discussed in etiological and maintenance models of SAD [12-14]. Empirical evidence remains limited regarding whether these components are specifically important for young people with SAD [21]. It would be valuable to test both transdiagnostic anxiety and SAD-specific components individually in interventions. Compared to traditional RCTs, study designs such as factorial or leapfrog designs would be better suited to evaluate the efficacy of individual components and determine the most effective combination for intervention development [135-138].

Similar to previous evidence, moderator analyses supported the efficacy of guided DMHIs but not unguided DMHIs [59]. Notably, there was great variety in the extent of human guidance—from summaries of exercises sent to participants via chat or email to regular telephone contact. The amount of support was quantified only in a minority of the trials (6/22, 27%). Furthermore, while some guidance formats reflected strategies to improve engagement with the intervention, the guidance of other trials could be viewed as additional treatment, aligning more with blended therapy. To investigate how to optimize the amount, the content, or the delivery setting of guidance for young people, future studies should be more detailed in their report of the guidance format and its intended purpose and may even compare different support strategies for young people [139]. Possible results could streamline future implementation research and support the use of, and continuous engagement with, digital approaches.

Another aspect that varied considerably across the studies was parental involvement. In the study by Morgan et al [108], the intervention was only directed at parents, while other studies included specific parent information or modules that complemented the children’s modules [56,94,102,103] or only addressed the young people. Research to date shows an inconclusive picture of how parents can be best involved in treatments to increase their overall effectiveness compared to interventions meant only for children or adolescents [140]. Comparable to our results, other meta-analyses also found no additional effect of parental involvement [141-143]. Especially in adolescence and young adulthood, the social context relevant to anxiety-inducing situations lies primarily outside the parental home (eg, at school and with peers). Therefore, the involvement of parents might be less relevant for symptom reduction. Nevertheless, for young people, parents play an important role in accessing mental health services [144,145] and could therefore be a crucial supporting factor in their children’s adherence to the intervention [146].

Adherence was defined and reported, if at all, in great variety in the included studies. Most commonly, completed modules and time spent in the DMHI were reported; however, their effect on treatment outcome was rarely examined. Although these metrics offer some insight into DMHI use, they do not provide any information about reasons for nonadherence. Previous efforts have sought to explain nonuse of DMHIs or treatment. In adult samples, a proposed explanation is the “good enough” effect—participants may discontinue a DMHI after achieving their personal goals [147]. A qualitative examination of adolescents who stopped treatment for depression identified a similar group, along with 2 additional types: the dissatisfied, who did not find the intervention helpful; and the troubled, who reported a lack of stability in their lives that hindered engagement with an intervention [148]. Future research should report a minimum of adherence parameters (eg, completed modules and time spent in the program) and might even add reasons for nonadherence. Furthermore, it would be important to report the analysis of whether adherence influenced the outcome within the publication reporting the main results of the RCT. With this information, studies could be better compared, and the influence of adherence on the outcome could be assessed meta-analytically [149].

Regarding potential moderators, age did not emerge as a moderator of treatment efficacy in the meta-regression analysis, but the subgroup analysis reported significantly higher effect sizes for adolescents and young adults compared to children. Likewise, other meta-analyses have indicated that older adolescents profit more from DMHIs than children [37]. However, the small number of studies including child samples in this meta-analysis could also be a reason for this result. Furthermore, although the influence of age was taken into account in this meta-analysis—through meta-regression (mean age), subgroup analyses (age groups), and sensitivity analyses excluding studies with participants aged >25 years—many of the primary studies included wide age ranges and did not report outcomes separately by age group. This highlights the need to replicate these findings in future studies.

Baseline symptom severity moderated treatment outcome with a small effect. In other studies, larger treatment effects were found for adults with higher initial symptom severity [150]. The evidence in childhood and adolescence is mixed: some meta-analyses or studies found a significant effect, while others did not [151]. Therefore, this result is in line with the overall mixed picture.

This review included DMHIs that were developed primarily in high-income countries and provided in English, Swedish, Chinese, Persian, Spanish, or Dutch. Furthermore, only a few undertook efforts to match the socioeconomic composition of the study with that of the home country [56,100]. A recurring argument in favor of DMHIs is that they offer help at low cost with a low threshold and high accessibility to young people from nonurban and lower socioeconomic groups [38,152,153]. However, this presumption can only be realized if measures are taken to design DMHIs for diverse target groups; evaluate them in different demographic groups; and support their implementation, particularly in those groups that otherwise have little access to mental health care. To achieve this, young people need to be actively involved in designing interventions and validation studies, recruiting participants for trials, and communicating about and implementing digital self-help options [133,154].

This review and meta-analysis have some limitations, and the findings should be considered thoughtfully. First, the meta-analysis and included subgroup analyses were based on a small number of studies, and the results should be interpreted with caution. The limited number of studies precluded an examination of the efficacy of specific intervention components (eg, psychoeducation, exposure therapy, social skills training, CBM, or relaxation and mindfulness exercises) or the differentiation of effects between diagnostic categories (subclinical vs clinical). Most of the studies included participants with elevated scores in social anxiety screenings based on self-reports, and only a few of the identified prevention studies made diagnostic efforts to distinguish youths with subclinical social anxiety symptoms from those with a previous or current SAD diagnosis. Thus, prevention and treatment studies could not be categorized reliably to investigate potential differential treatment effects in subgroup analyses. Moreover, only 2 studies provided long-term follow-up data, resulting in limited information on the sustainability of treatment effects. Next, the meta-analysis encompassed a broad age range, spanning distinct developmental stages and diverse target groups; for instance, while adolescents are typically the primary end users of digital interventions, parents or legal guardians generally serve this role for preschool-aged children, who often lack the cognitive and emotional capacities to engage with such interventions independently. To explore potential age-related effects, subgroup analyses were conducted across 3 age categories (ie, children, adolescents, and young adults). However, these analyses were limited by the small number of primary studies within each subgroup and did not account for developmental differences within or between age groups. Given the substantial developmental variability that exists within these broad categories, future research should examine potential moderator effects using more granular age distinctions and developmental profiles, particularly in terms of cognitive, emotional, and social competencies.

Furthermore, factors influencing efficacy, such as adherence, were inadequately reported in primary studies, preventing their incorporation into the meta-analysis, for instance, in a meta-regression. Finally, the methodological quality of the studies varied in terms of transparency in randomization, masking procedures, and the consideration of appropriate statistical procedures.

Therefore, future intervention studies should be methodologically more rigorous; include thorough diagnostic procedures; provide long-term follow-up data; and report on potential moderators such as adherence, parental involvement, guidance, and participant characteristics in more detail. Moreover, more details on how access to the modules was managed would be interesting because it may have an impact on participant adherence as well as intervention efficacy. Furthermore, future studies should aim to target a socioeconomically diverse study population and report on demographic information that may influence the availability and efficacy of treatments (eg, gender, age, country, urbanicity, socioeconomic status, culture, and minoritized status [155]). This additional information could inform future meta-analyses or, preferably, meta-analyses of individual participant data that do not rely on summary statistics to investigate subgroup effects.

In conclusion, while further research is necessary to provide conclusive evidence on the efficacy of DMHIs for young people, our systematic review and meta-analysis show their potential to alleviate SAD symptoms. The preliminary findings underscore the importance of future research efforts aimed at clarifying the specific treatment components that render DMHIs effective, optimizing human guidance strategies and individualizing interventions to enhance therapeutic outcomes. Moreover, our study highlights the need for greater inclusivity in DMHIs and research design, emphasizing the need for future studies to encompass participants from all socioeconomic backgrounds. By addressing these considerations, DMHIs hold promise in bridging the care gap and democratizing access to psychotherapeutic interventions for all young people, thus fostering mental health equity in the future.