This SR and NMA follow the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA-NMA) guidelines [38]. The completed PRISMA-NMA checklist is provided in Checklist 1. The protocol for this study has been registered with PROSPERO (CRD42024527317).

The search strategy was developed and reported in accordance with the PRISMA-S guideline [46]. Searches for RCTs were conducted in PubMed (including MEDLINE), EMBASE, Web of Science, and the Cochrane Library. Searches were performed through the native interfaces of each database (PubMed via NCBI, EMBASE via Elsevier, Web of Science via Clarivate, and Cochrane Library via Wiley). In addition, the reference lists of relevant SRs were checked to ensure that no eligible trials were missed.

The search terms were formulated according to the PICOs framework, including participants or populations, interventions, outcomes, and types of research design. Both Medical Subject Headings and free-text terms were included as appropriate. Boolean operators (“AND,” “OR”) were used to combine search terms, and database-specific search techniques such as truncation, phrase marks, and wildcards were applied. The complete search terms and algorithm were provided in Multimedia Appendix 1, and search strategies for the other databases were adapted accordingly. The search was designed and executed by 2 reviewers (YZ and WP), who were trained in SR methodology, and the strategy was cross-checked for completeness and accuracy.

The literature search was initially conducted on June 13, 2024 and was last updated on November 15, 2025. All retrieved records were imported into EndNote X9 for citation management, and duplicates were removed using both automated and manual deduplication. Additional search methods included checking the reference lists of the included studies or relevant SRs. Gray literature was also searched via Google Scholar, OpenGrey, and ProQuest Dissertations. The search was limited to studies published in English due to resource constraints for translation.

Two reviewers (YZ and DH) independently screened the titles and abstracts against predefined protocol criteria. Full texts were retrieved for all potentially eligible studies. When multiple articles were identified from the same randomized controlled trial, the most recent or most comprehensive publication was retained for data extraction. Earlier reports were used to supplement missing information on study design, intervention details, or outcomes when necessary. Any discrepancies between the 2 reviewers were resolved by discussion. If disagreements persisted, a third reviewer (WP) was invited for adjudication. At the title and abstract screening stage, we excluded 14,788 records that clearly did not meet the eligibility criteria, most commonly because of wrong study design (eg, cross-sectional surveys, qualitative studies, reviews, protocols), wrong population (non-youth samples), ineligible intervention or comparator (ie, the difference between study arms did not lie in the use of a DHI), or an unrelated topic.

Data were extracted using a standardized and piloted form. Extracted variables included: first author, publication year, recruitment region, participant characteristics (mean age, SD, gender distribution), type of intervention and comparator, sample size per arm, intervention duration, and reported endpoints. Detailed characteristics of the DHIs were also extracted to facilitate subcategorization (Table 1). Further, outcomes and corresponding measurement methods were recorded, such as self-reported condom use and validated self-efficacy scales.

When outcome data were incomplete or unclear, study authors were contacted by email for clarification; trials with essential missing data were excluded from the quantitative synthesis and documented in Multimedia Appendix 2. All eligible studies were included in the SR, and only studies with usable and connected outcome data were included in the NMA.

The methodological quality of the included studies was independently assessed by 2 reviewers (YZ and WP), with disagreements resolved by a third reviewer (CZ). Risk of bias was evaluated using version 2 of the Cochrane risk of bias 2 tool (RoB 2) for randomized trials [47]. For each domain, studies were rated as having “low risk,” “some concerns,” or “high risk” of bias according to the Cochrane Handbook (version 6.5) [48]. Domain-level risk of bias judgments for each trial are summarized in Figure 1, with extended graphs and contribution matrices provided in Multimedia Appendix 3. The reference list of included studies is provided in Multimedia Appendix 4 [26,49-71].

Because blinding was generally not feasible for these nonpharmacological interventions, many trials were judged at high risk of bias in the domain of deviations from intended interventions [72,73]. As this limitation was expected and unlikely to influence objectively measured outcomes, we did not consider this domain when grading the certainty of evidence. The overall certainty of evidence was determined using the CINeMA (Confidence in NMA) web application, which is based on the GRADE framework [74,75]. In addition, we constructed GRADE “Summary of Findings” tables using the official template provided by the GRADE Working Group to summarize the key relative and absolute effects and certainty ratings for the primary outcomes (Multimedia Appendix 3).

RoB 2 assessments were incorporated into the interpretation of NMA findings and into the GRADE/CINeMA evaluation of certainty, but they were not used to weight studies in the statistical synthesis.

From a total of 25,659 records initially retrieved, 24 RCTs published between 2004 and 2024 were included in the final analysis (Figure 2). These studies were conducted across 8 diverse countries, predominantly in the United States (n=14), with the remainder from China (n=2), the United Kingdom (n=2), Uganda (n=2), Singapore (n=1), the Netherlands (n=1), Australia (n=1), and Spain (n=1). The trials collectively enrolled 20,134 participants (range 50‐6248; mean 838.9, SD 1358.2), with a mean age of 19.5 years. Overall, 10,228 participants (53.4%) were male, although sex composition varied substantially—some studies enrolled only males [49-51], only females [52,53], or mixed populations. Of the 24 included studies, 21 studies were two-arm, and 3 studies were multi-arm. Across the included studies, 6 trials evaluated TCI, 8 assessed interactive online-based intervention (IOI), 6 examined MAI, and 8 investigated SWI. The total number of intervention approaches (n=28) exceeded the number of included studies (n=24) because several trials directly compared 2 or more active interventions (eg, IOI vs SWI) without including a conventional control group. Intervention durations ranged from brief sessions lasting 10‐20 minutes up to 12 months. Follow-up periods were heterogeneous, spanning from immediate postintervention assessments to 24 months. Most studies reported outcomes at 3‐6 months, while only a few provided longer-term follow-up beyond 12 months. Five studies performed analyses for more than one time point. To enhance consistency and reduce potential bias associated with short-term variability, we extracted outcomes at the longest follow-up time point, thereby facilitating a more comprehensive evaluation of the intervention’s long-term effectiveness [44,45].

For condom use at last sexual contact, 4 out of 7 studies reported ORs greater than 1, suggesting a possible beneficial effect of the interventions; however, only one trial [54] showed a clear statistical significance (OR 1.13, 95% CI 1.01‐1.25). Regarding consistent condom use, 6 of 11 studies showed ORs above 1, with substantial heterogeneity; one study reported a very large effect (OR 4.55, 95% CI 1.15‐17.95) [55]. For the proportion of condom use, 5 out of 6 trials reported ORs above 1, suggesting a tendency towards higher condom use in the intervention groups; however, CIs were wide and often included the null (overall OR range 0.48‐2.43), indicating that the evidence for this outcome is imprecise. Finally, for STIs incidence, including HIV, effects varied substantially across 7 trials, with ORs ranging from 0.53 to 2.10. Four of the 7 trials had point estimates below 1 [49,50,53,79], and 3 trials had 95% CI that excluded 1 [49,56,57], indicating heterogeneous and partly conflicting evidence for this outcome. These results summarized the observed effects across trials, highlighting that effect estimates varied considerably across outcomes and studies. The characteristics and effect estimates of the included studies were summarized in Table 2.

Risk-of-bias assessments using the RoB 2 tool are summarized in Figure 1. Overall, most trials were judged to be at low risk of bias for the randomization process, outcome measurement, and selection of the reported result. However, a substantial minority of studies had some concerns or high risk of bias in at least one domain, most frequently for deviations from the intended interventions and missing outcome data. Consequently, several trials were rated as having some concerns or a high overall risk of bias.

Although condom use self-efficacy and number of sexual partners were prespecified as secondary outcomes, too few studies reported these measures to allow meta-analysis. Only one trial evaluated condom use self-efficacy. In Rinehart et al [66], this construct was assessed using 3 items developed within the Health Belief Model (range 0‐12; Cronbach α=0.72). At the 3rd month, the intervention group reported significantly higher self-efficacy scores than the control group (7.38 vs 6.68; P=.04), but this difference was no longer significant at the 6th month (7.39 vs 6.99; P=.20). Two trials reported on the number of sexual partners. In Shafii et al [79], participants in the intervention arm reported a 29% reduction in the number of sexual partners at follow-up, although the effect did not reach statistical significance (IRR 0.71, 95% CI 0.50‐1.03, P=.07). Changes in the control group were not reported. By contrast, Free et al [54] examined the proportion of participants reporting 2 or more sexual partners over 12 months. At one year, this outcome was reported by 56.9% of intervention participants compared with 54.8% of controls (OR 1.11, 95% CI 1.00‐1.24, P=.06). Overall, the evidence on the impact of digital interventions on the number of sexual partners remains limited and inconsistent.

All of these enrolled studies were RCTs, and the quality of evidence was evaluated by the Cochrane Handbook and graded each potential source of bias as low, high, or some concerns; the details were displayed in Figure 1. We assessed confidence in the results of the NMA using the CINeMA framework. Of the 4 outcomes analyzed, only “consistent condom use” met the criteria for CINeMA assessment. The remaining outcomes were excluded due to insufficient numbers of studies or disconnected network structures. For consistent condom use, certainty of evidence was mainly downgraded for within-study bias and imprecision, resulting in overall ratings of “low” to “very low” confidence. Among the 10 comparisons, 1 (10%) was rated as “very low” and 9 (90%) as “low” certainty. For condom use at last sex, the proportion of condom-protected acts, and STI incidence, the certainty of evidence is less well characterized, but given the sparse data, risk of bias, and wide intervals, these estimates should similarly be interpreted as low certainty. A detailed summary of risk-of-bias judgements, CINeMA assessments, and the corresponding GRADE “Summary of Findings” information for each primary outcome is provided in Figure 1 and Table 4, with full GRADE “Summary of Findings” tables formatted according to the GRADE Working Group template available in Multimedia Appendix 3 to aid interpretation of the magnitude and certainty of the main comparisons.

To the best of our knowledge, this is the first NMA to systematically evaluate the comparative effectiveness of different modalities of DHIs on promoting safer sexual behaviors among youth. By simultaneously examining 4 distinct digital modalities and 3 behavioral outcomes, and by incorporating STI incidence (including HIV) as a biological endpoint, this review expands the current evidence base and clarifies which intervention types are better suited for immediate versus sustained behavior change, and highlights the gap between improvements in self-reported safer behaviors and reductions in biological HIV/STI infection. Drawing upon data from 24 RCTs across diverse contexts, our findings offer comprehensive insights to inform future development, optimization, and implementation of DHIs in HIV/STIs prevention, underscore the need for designing multimodal, context-aware digital interventions that integrate behavioral support with access to testing and care services, and provide practical considerations for policymakers, program designers, and digital platform developers who seek to tailor DHIs to youth populations.

Across outcomes, between-study heterogeneity and statistical inconsistency were generally low to moderate, and the network satisfied the assumptions of transitivity and global consistency. However, most comparisons were informed by a small number of trials, many of which had some concerns or a high risk of bias in at least one RoB 2 domain. The complementary frequentist NMAs showed that 95% PIs were typically wide and often crossed the null, even when average effects appeared beneficial. Consistent condom use was the only outcome that met CINeMA requirements, and all network comparisons for this outcome were rated as having low or very low certainty. Together, these features suggest that our estimates reflect uncertain average effects rather than precise predictions for specific programs or settings. These patterns of risk of bias, particularly deviations from intended interventions and missing outcome data, may have led to overestimation of some intervention effects or increased uncertainty in the network estimates and contributed to downgrading the certainty of evidence in our GRADE/CINeMA assessment.

In assessing condom use at the last sexual encounter, TCI emerged as the only approach showing statistically significant improvement compared with NDI. This finding aligns with prior studies reporting absolute increases in condom use among participants receiving SMS or phone-based reminders [80]. The relatively stronger performance of TCI may be attributable to its simplicity and immediacy, directly prompting protective behaviors without requiring advanced digital literacy or prolonged engagement [81]. Previous evidence has also highlighted TCI as one of the more acceptable and widely used forms of digital intervention among young people [30]. Thus, TCIs may offer an immediate behavioral benefit, especially for outcomes tied to the most recent sexual event. Interestingly, while MAI ranked highest in SUCRA probability, its effect did not reach significance, suggesting inconsistency between ranking and statistical evidence. This discrepancy could stem from limited trial numbers, variability in app engagement, or short intervention duration, which may have constrained the power to detect statistical changes. However, this apparent benefit of TCI is based on a few trials with some concerns or high risk of bias, and the wide PIs suggest that similar effects may not be consistently achievable in all implementation settings.

In this study, consistent condom use is improved more effectively by SWI and IOI than by TCI, with SWI ranking the highest. This finding aligns with prior evidence indicating that web-based and online interventions contributed to a substantial proportion of effective digital interventions [21,30]. First, that SWI outperformed IOI may seem counterintuitive, given the absence of interactive features. However, SWI could deliver standardized, theory-based content in a less distracting, user-driven format, allowing youth to process key prevention messages at their own pace. Moreover, while the IOIs included in this review are mostly delivered via computers or tablets, SWIs—though static—are often accessible on smartphones or distributed through popular platforms such as Facebook or WeChat with text, images, or videos [82]. This accessibility and portability may explain why SWI demonstrates stronger effects on consistent condom use. Second, TCI was less effective than both SWI and IOI for improving consistent condom use. Previous studies have suggested that TCI, especially SMS-based reminders, remain controversial in terms of their long-term effectiveness for HIV prevention behaviors [83]. Therefore, although TCI can effectively prompt immediate behaviors, its brief and repetitive messages may lack the depth and reinforcement required to sustain consistent condom use over time. Nevertheless, most trials contributing to this outcome had some concerns or high risk of bias, and CINeMA rated the certainty of these network estimates as low to very low, so the apparent superiority of SWI and IOI should be considered tentative.

When examining the proportion of condom use, IOI demonstrated a significant advantage over SWI, indicating the added value of interactive engagement. This aligns with prior evidence showing that increases in condom use were significantly associated with the use of tailored strategies [21], feedback provision, and guided navigation in digital interventions. Unlike static websites, IOIs integrated tailored feedback, quizzes, or real-time support, which represent core behavior change techniques such as personalized feedback, problem-solving, and self-regulation strategies [84]. These core behavioral change techniques are particularly effective in influencing situational decisions and negotiations during sexual encounters, thereby enhancing the overall proportion of condom use [85,86]. In other words, consistent condom use reflects the internalization of long-term protective norms, and it could be reinforced by standardized and less distracting formats like SWI, while the proportion of condom use is more sensitive to moment-to-moment decision-making. This divergence in findings—SWI being more effective for consistent use, while IOI excels in overall proportion—highlights that different behavioral outcomes may respond to distinct mechanisms of action. Besides, the evolution of technology-based intervention modes has gradually expanded from web-based formats to SMS and social media [87-89]. This trend further supports the notion that while static formats may effectively reinforce long-term norms, interactive platforms provide additional advantages for immediate behavioral decisions during sexual encounters. Yet the CrIs and PIs for this outcome were wide and frequently included the null, indicating considerable heterogeneity and imprecision and implying that any average benefit in the proportion of condom-protected acts may not translate into clear improvements in every context.

Unlike the behavioral outcomes, the effectiveness of DHIs on reducing STI infection reveals a different pattern: NDI and TCI performed more favorably, while IOI and SWI ranked lowest. This finding contrasts with prior studies showing that digital interventions can improve HIV/STI care engagement, such as testing uptake or service use [90-92], highlighting that improvements in care engagement do not necessarily translate into reductions in biological outcomes like STI incidence. Several factors may explain this inconsistency. First, STI incidence represents a distal biological endpoint that may require longer follow-up to capture meaningful reductions, and improvements in self-reported behaviors, such as condom use, may not directly translate into biological protection due to reporting bias or inconsistent application in high-risk contexts. Second, reductions in STI incidence depend not only on safer behaviors but also on timely testing, treatment, and linkage to care. However, evidence shows that stigma related to gender identity, socioeconomic status, race, and ethnicity can delay care-seeking and discourage individuals from accessing necessary services [93]. Nondigital approaches, such as community outreach or peer education programs, often combine behavioral education with direct access to services such as STI testing, treatment linkage, and ongoing support from trained staff or peers, which can directly impact biological outcomes. In contrast, many DHIs focus primarily on education and motivation, without providing structured access to testing or clinical care. Additionally, NDIs may facilitate stronger trust and engagement through in-person interactions, which can overcome barriers related to stigma, confidentiality concerns, or digital literacy limitations. Therefore, while DHIs can effectively change self-reported behaviors, the integrated, multi-component structure of NDIs may explain their relative advantage in reducing actual STI incidence (including HIV). Besides, in this study, TCI is the only digital intervention showing a relatively favorable effect on reducing STI incidence, second only to NDI. Prior studies have demonstrated that TCI can facilitate access to HIV prevention services for youth and achieve high patient and provider satisfaction [93,94]. By providing a comfortable, judgment-free platform, telemedicine may be particularly preferred by marginalized populations, especially transgender youth [95,96]. Given the small and heterogeneous evidence base, important risks of bias in several trials, and wide PIs, these findings on STI incidence should be regarded as exploratory and hypothesis-generating rather than definitive.

Our use of PIs further illustrates the extent to which the observed benefits of DHIs may vary across settings. For most comparisons, the 95% PIs were wide and often crossed the null value, even where the corresponding credible or CIs suggested modest advantages over NDI. This pattern indicates that, although certain DHI modalities tend to improve condom use on average, implementation in new populations or health systems may yield smaller effects or no clear benefit, underscoring the need for careful adaptation, monitoring, and evaluation when scaling up digital prevention programs.

These findings have several implications. First, the differential effectiveness of DHIs suggests tailoring intervention types to targeted outcomes. TCI showed immediate benefits for condom use at last sexual contact and a relative advantage for STI incidence, indicating that simple, low-burden interventions may prompt rapid behavior change and reach marginalized youth. In contrast, SWI and IOI were more effective for consistent and habitual condom use, highlighting the value of self-paced content and behavior change techniques that enhance motivation and self-regulation. Multi-modal approaches combining these strengths may maximize overall effectiveness. Second, DHIs alone may be insufficient to reduce STI incidence. Biological outcomes require longer follow-up, and self-reported behavior change does not always translate to infection reduction. Integrating DHIs with offline services, such as condom distribution, PrEP promotion, routine testing, and clinical linkage, is likely necessary to achieve meaningful improvements. Third, these results have implications for intervention design and digital health policy. When targeting youth populations, accessibility, acceptability, and engagement should be prioritized. For example, interventions delivered via mobile platforms or telecommunication may overcome barriers related to stigma or limited digital literacy, while interactive online content can leverage BCTs to support skill acquisition and habitual behavior change. Intervention planners should also consider the balance between immediacy and sustainability of effects: brief, repeated prompts may drive immediate behavior, whereas structured, self-paced content may reinforce long-term habits.

Several limitations of this study should be noted. First, substantial clinical and methodological heterogeneity across trials—including differences in intervention intensity and duration, digital platforms, follow-up periods, and outcome definitions—may have contributed to between-study heterogeneity and could challenge the transitivity assumption underpinning some indirect comparisons in the NMA. Second, we restricted inclusion to randomized, prevention-focused DHIs among HIV-negative or status-unknown youth and excluded nonrandomized studies and trials targeting HIV-positive adolescents; the findings may therefore not generalize to these subgroups or to broader digital programs. Third, all behavioral outcomes relied on self-report, and secondary outcomes such as self-efficacy and number of sexual partners were too sparsely and inconsistently reported to be synthesized, limiting our ability to evaluate the broader psychosocial impact of DHIs beyond condom use. Fourth, the number of trials assessing certain interventions, particularly MAI, was limited, which may reduce statistical power and the precision of effect estimates. Fifth, we were unable to formally assess small-study or publication bias, and selective nonpublication or outcome reporting cannot be ruled out. Finally, most network comparisons were rated as low or very low certainty because of within-study bias and imprecision, and PIs, estimated using standard random-effects methods in netmeta, were wide; the true comparative effects may therefore differ meaningfully from our estimates, underscoring the need for rigorous, adequately powered RCTs with standardized outcomes and longer follow-up.

In conclusion, this NMA provides the most comprehensive synthesis to date on the comparative effectiveness of DHIs in promoting safer sexual behaviors among youth. This NMA highlights that the effectiveness of DHIs for HIV prevention among youth depends on both intervention modality and targeted outcomes. While DHIs can enhance knowledge and protective behaviors, their impact on biological endpoints remains limited without integration with offline services and broader structural support. Tailoring interventions to behavioral targets, engagement strategies, and contextual factors is essential to maximize their potential in promoting youth sexual health. We hope that these results will inform the design of youth-centered digital prevention programs, guide clinicians and educators in selecting appropriate modalities, and support policymakers and guideline developers in integrating digital strategies into national HIV prevention frameworks. Future studies should focus on the specific characteristics of patients to provide personalized estimates of comparative effectiveness and individualized predictions regarding the probability of response to treatment and of side effects.