Effectiveness of Digital Health Interventions on Stress, Anxiety, and Depression in University Students: A Systematic Review and Meta-Analysis

The mental health of young people has attracted increasing attention and is now recognized as a global public health challenge¹. University education aims to reinforce students' intellectual abilities and prepare them for productive and successful lives as adults². However, during this particular phase in life, students frequently encounter multiple stressors, such as moving away from home, becoming more independent, taking on new responsibilities, and managing academic workloads³. Such stressors can adversely affect both physical and emotional well-being, leading to diminished academic performance, lower life satisfaction, reduced self-confidence, increased dropout rates, and, in severe cases, suicidal thoughts⁴. Indeed, a considerable proportion of university students report elevated levels of perceived stress, which is defined as the appraisal of environmental demands as exceeding one's coping abilities⁵. This population is particularly vulnerable to stress, anxiety, and depression, with late adolescence through early adulthood representing the peak period for the onset of mental health disorders⁶. Studies indicate that compared with non-students of the same age, one-third of university students have experienced or are currently experiencing severe mental health problems and higher levels of depression, anxiety, and distress³. A recent review indicated that the global prevalence of depression and anxiety symptoms among college students is high, at 33.6% and 39.0%, respectively⁷. Notably, approximately 75% of adults diagnosed with mental disorders experience initial symptoms before the age of 25⁸. This early onset is particularly alarming given that suicide is a leading cause of death among young people aged 15-29 worldwide⁹. These findings collectively underscore the urgent need for effective interventions targeting stress, anxiety, and depression in university student populations.

Addressing the high prevalence of stress, anxiety, and depression among university students requires multifaceted approaches, including psychological counseling, pharmacotherapy, social support, physical exercise, and cognitive-behavioral interventions^10,11. With rapid advances in information technology, evidence-based interventions are being increasingly delivered through scalable digital platforms (e.g., smartphone apps, web portals, and teletherapy systems), complementing traditional face-to-face services¹². The term "digital health interventions" refers to responsive, technology-supported approaches that include personalized health communications, biometric or behavioral tracking, and just-in-time informational support¹³. These can be broadly categorized into two modalities: web-based platforms offering structured psychoeducational programs and mobile health (mHealth) applications delivering targeted interventions via smartphone-optimized interfaces.

Adolescents and young adults represent the most digitally engaged demographics worldwide, with a 70% internet penetration rate (versus 48% in the general population)¹⁴. Previous systematic reviews and meta-analyses have investigated the effects of online mindfulness-based interventions, physical activity interventions, computer-delivered and web-based interventions, and other psychotherapy interventions¹⁵ on reducing stress, anxiety, and depression among university students. For instance, a recent systematic review of nine RCTs involving 1,100 participants explored the effectiveness of online mindfulness interventions in improving mental health outcomes in this population¹⁶. However, the existing findings remain inconsistent. One study found that the mHealth application "Destressify" did not significantly improve stress, anxiety, or psychosocial functioning¹⁷. Similarly, Kvillemo et al. reported no statistically significant advantage of a mindfulness intervention over an active control condition¹⁸. While digital health interventions hold promise for promoting emotional well-being, enhancing treatment engagement, and offering a cost-effective alternative to conventional methods, a comprehensive synthesis of their overall efficacy is necessary to guide investment and implementation. To date, no systematic review or meta-analysis has integrated evidence across different types of digital health interventions targeting stress, anxiety, and depression specifically in university students.

To address this gap, we propose the following research questions: (1) Are digital health interventions effective for improving university students' stress, anxiety, and depression compared to active and passive control conditions? (2) What is the magnitude of the effect of digital health interventions on mental health outcomes in this population? (3) Which types of digital health technologies are most effective at alleviating depression, anxiety, and stress?

Therefore, the present study has three primary objectives. First, the evidence was evaluated regarding the effectiveness of digital health interventions in treating stress, anxiety, and depression in university students. Second, the efficacy of these interventions across reported outcomes was statistically summarized. Third, the quality of the available evidence was assessed. Additionally, given the diversity of the student population, we explored how participant characteristics may influence intervention outcomes.

To this end, we conducted a systematic review and three meta-analyses of randomized controlled trials (RCTs) measuring stress, anxiety, and depression outcomes. Through this work, we seek to support researchers in evaluating emerging evidence, identifying future research directions, and contributing to the development of effective digital treatment solutions for university students.

Study registration
This systematic review and meta-analysis protocol was registered on PROSPERO with a registration number of CRD 42024610457. This study was designed and conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines¹⁹.

Search strategy
A comprehensive literature review was conducted by two authors (XXZ and JZ), who searched the Cochrane Library, Ovid Embase, Ovid MEDLINE, PubMed, Scopus, and Web of Science Core Collection databases. Their search encompassed all relevant records from the start of each database's coverage up to June 30, 2024.

The search strategy employed a combination of controlled vocabulary (e.g., MeSH) and free-text terms pertaining to digital health interventions, depression, anxiety, and stress. To maximize the retrieval of pertinent studies, Boolean operators (AND, OR) were utilized. The search was restricted to publications in the English language. The complete search strategy is detailed in eTable 1 in Supplement 1. Citation Chaser was used to search the reference lists of the included studies and to retrieve articles that had cited the included studies to find additional relevant studies²⁰.

Study selection and inclusion criteria
The inclusion and exclusion criteria for this systematic review were defined according to the Population, Intervention, Comparison, Outcomes, Study Design (PICOS)¹⁹ framework: Population (P): Studies involving university students, regardless of age, gender, or field of study. Intervention (I): Any digital health intervention primarily delivered via the internet or a mobile device and aimed at alleviating symptoms of stress, anxiety, or depression. This included, but was not limited to, web-based programs, mobile apps, and chatbot-delivered therapies. Comparison (C): Control groups including wait list controls, active controls (e.g., receiving nondigital psychoeducation), treatment as usual, or other digital health interventions. Outcomes (O): Studies had to report outcomes on at least one of the following, measured using a validated scale: stress (e.g., Perceived Stress Scale, PSS), anxiety (e.g., GAD-7), or depression (e.g., PHQ). Study Design (S): Only randomized controlled trials (RCTs) were included to ensure the highest quality of evidence.

Upon completion of the database search and duplicate removal, the titles and abstracts of the identified records were screened independently by two authors (XXZ and JZ). Any discrepancies encountered were adjudicated by a third reviewer (DZ). Subsequently, the full texts of the remaining articles were assessed for eligibility by two other independent reviewers (DZ and YFC), with a third reviewer (JZ) serving as an arbiter to reach a consensus on final inclusion.

The inclusion criteria encompassed all digital health intervention types (e.g., web-based programs, mobile applications) to facilitate a comprehensive analysis and direct comparison across technological modalities. Studies were required to report outcomes related to stress, anxiety, or depression using validated measurement scales. Studies were excluded if they possessed the following characteristics: lack of assessment of the relevant variables, population outside the intended scope, review articles, abstracts, editorials or letters, animal studies, or case reports. Studies that reported insufficient data for calculating effect sizes, even after attempts to obtain it from the authors, were excluded from the meta-analysis. Conference abstracts were excluded due to the lack of detailed information and in-depth methodology required for a robust assessment of quality and data extraction. Citations from all the databases were imported into an EndNote X21 library (Clarivate Analytics).

Appraisal of methodical quality
The quality of the included RCTs was independently assessed by two authors (DZ and YFC) according to the guidelines of the Cochrane reviews²¹. The evaluation contents include (1) random sequence generation, (2) allocation concealment, (3) blinding of participants and personnel, (4) blinding of outcome assessment, (5) incomplete outcome data, (6) selective reporting, and (7) other types of bias. Each included study was judged to have a "low", "high", or "unclear" risk of bias for each domain. If the researchers scored the four criteria as "low" and if no serious flaws were detected, then the study was scored as having a low risk of bias.

Extraction of the data
The following data from eligible studies were independently extracted by two authors (JZ and DZ): authors, year of publication, country, participant characteristics (including age, sample size, and distribution of groups), assessment tools employed for depression, anxiety, and stress, digital health interventions (delivery mode, treatment length, control group, intention to treat, attrition rate), and the related statistical data. The procedure for handling missing data (e.g., standard deviations) involved attempting to contact the corresponding authors twice within a four-week period. If no response is received, then SDs will be calculated from standard errors, confidence intervals, or P values provided in the studies, following the methods outlined in the Cochrane Handbook²¹. Any discrepancies or inconsistencies in the extracted data were resolved through discussion between the two reviewers (JZ and DZ). If a consensus could not be reached, then a third reviewer (XXZ) was consulted to make the final decision.

Data synthesis
For each outcome, the quantitative data were synthesized and are presented as the weighted mean difference (WMD) with a corresponding 95% confidence interval (CI). The WMDs were calculated by pooling the pre-to-post change scores (mean and SD) extracted from each eligible study. Heterogeneity across studies in direct comparisons was assessed using the I² statistic, with values of 25%, 50%, and 75% conventionally denoting low, moderate, and high heterogeneity, respectively. A fixed effects model was applied when heterogeneity was not significant (I²< 50% and P > 0.1); otherwise, a random-effects model was employed.

The results were visualized using forest plots, which display the first author's name, publication year, sample size, effect estimate with its 95% CI, and associated P value for each study. Subgroup analyses were performed separately for each primary outcome (stress, anxiety, and depression). For each outcome, the analyses were stratified by the specific measurement instrument used (e.g., for stress: Connor-Davidson Resilience Scale (CD-RISC)²², Brief Resilience Scale (BRS)²³, Perceived Stress Scale (PSS)²⁴, Depression Anxiety Stress Scale (DASS)-Stress subscale^25,26,27; for anxiety: Generalized Anxiety Disorder Scale (GAD)²⁸, State-Trait Anxiety Inventory (STAI), DASS-Anxiety subscale^27,29; for depression: Patient Health Questionnaire (PHQ)³⁰, Beck Depression Inventory (BDI)³¹, DASS-Depression subscale^26,27). Other subgroup variables included the following: (1) intervention technique: web- or app-based cognitive-behavioral therapy (CBT), mindfulness-based intervention (MBI), physical-activity intervention (PAI), and other psychological intervention (OPI) excluding CBT, MBI, and PAI; (2) guidance format: reminder only, feedback only, mixed (reminder + feedback), or none; (3) delivery mode: smartphone app, web-based platform/program, or other; (4) treatment duration: ≤ 4 weeks, > 4 to < 8 weeks, or ≥ 8 weeks; (5) recruitment pathway: online, mixed, on-campus, or unspecified; and (6) control group type: active or passive. Passive controls receive no intervention as they serving to control for the natural history of the condition and placebo effects. In contrast, active controls receive an alternative, standard, or placebo intervention to control for nonspecific effects of the intervention process. The assessment of publication bias involved examining funnel plot asymmetry and performing Egger's regression test. Bias was considered to be present if visual asymmetry was accompanied by a significant Egger's test result (P < 0.05). Egger's test was conducted solely for outcomes with 10 or more studies to ensure the test's power³². The primary data analysis, including the generation of forest and funnel plots, was performed using Review Manager (RevMan) version 5.3 software. Additionally, Stata version 18 was used to conduct the statistical assessment for publication bias (Egger's test).

A total of 20,975 potentially relevant studies were initially identified from the six databases. After 7,374 duplicate records were removed, 13,601 studies underwent screening. Of these, 13,548 were excluded based on their titles and abstracts. The remaining 53 studies underwent full-text review for eligibility. Following this assessment, 31 studies were excluded for the following reasons: (1) use of non-validated or inappropriate measurement scales (n = 7), (2) intervention not meeting the inclusion criteria (n = 5), (3) incorrect publication type (n = 1), (4) outcomes not relevant to this analysis (n = 3), (5) population inconsistent with the review scope (n = 1), (6) non-randomized study design (n = 5), and (7) insufficient data for effect size calculation or meta-analysis (n = 9). Consequently, 22 RCTs were included in the final systematic review. The study selection process is detailed in Figure 1.

Figure 1: Flow chart for the selection of eligible studies. Please click here to view a larger version of this figure.

The basic characteristics of the included studies
The 22 included RCTs were conducted across nine countries: the United States (n = 6)33,34,35,36,37,38, Canada (n = 5)17,39,40,41,42, Germany (n = 2)43,44, Iran (n = 2)45,46, Italy (n = 2)47,48, South Korea (n = 2)49,50, the United Kingdom (n = 1)51, Australia (n = 1)52, and Sweden (n = 1)53. With respect to intervention types, seven studies employed cognitive behavioral therapy (CBT)-based interventions34,35,37,39,44,45,51, seven studies utilized mindfulness-based interventions (MBIs)34,39,40,43,44,45,49, seven were based on stress management and cognitive learning theories32,36,37,38,42,46,48and one study involved a physical activity intervention (PAI)19. Based on delivery modes, the interventions were categorized as follows: mobile phone applications (n = 11)17,20,21,33,34,37,40,45,46,48,49, web-based platforms or programs (n = 7)19,35,36,41,42,44,47, and other/nonspecific modes (n = 4)32,38,39,43. The basic characteristics of the included studies are summarized in Table 1.

In terms of study design, 14 RCTs employed a three-arm design (comprising one control group and two experimental groups)17,19,20,21,32,33,34,35,36,37,41,43,45,47. Six studies utilized a four-arm design (one control group and three experimental groups)38,39,42,46,48,49, and two studies featured a five-arm design (one control group and four experimental groups)40,44. Nineteen trials compared the experimental interventions against a no-intervention control (i.e., no treatment or waitlist)17,20,21,32,34,35,36,37,38,40,41,42,43,44,45,46,47,48,49, one trial compared the intervention to face-to-face mindfulness practice39, one compared an app-based CBT intervention to a non-app-based alternative33, and one compared online physical activity with psychological counseling19.

Table 1: General characteristics of the included studies. Please click here to download this Table.

Participant characteristics
The 22 studies included a total of 3,655 participants. Seven of these trials had sample sizes of 150 participants or fewer17,32,37,38,43,44,45. A definitive total of 3,041 participants were included in the final analyses, with individual study sample sizes ranging from 47 to 486. Twenty-one studies reported the age of the participants, with the overall age ranging from 17 to 54 years. The mean age of the participants was reported in 20 trials and ranged from 20.3 to 27.1 years; however, two studies did not report mean ages20,45. Twenty-one studies recruited both male and female participants, and 19 trials had female participants constituting more than 60% of the sample, with percentages ranging from 61.5%48 to 100%46. One study by Sookyung et al. enrolled exclusively female university students46, and two studies did not specify the gender distribution20,44. A notable gender imbalance was observed, as most trials included a significantly greater number of female participants.

Ten studies documented reasons for participant withdrawal, which included insufficient duration of core interventions21,45,47, voluntary dropout or withdrawal from the study36,40, personal or family reasons, as well as device incompatibility19,37,44, absence of baseline assessments or underage status39, and failure to wear devices or provide required feedback46. Eight studies performed intention-to-treat (ITT) analyses32,35,36,17,40,41,42,47. One study did not report withdrawal numbers43, whereas the remaining 21 studies provided data on participant attrition, with attrition rates ranging from 5% to 40.6%.

Intervention programs
Regarding intervention duration, seven studies implemented programs lasting ≤4 weeks32,33,34,40,45,46,49, eight studies had durations between 4 and 8 weeks17,35,36,37,41,44,47,48, and seven studies featured interventions lasting ≥8 weeks19,20,21,38,39,42,43. In terms of guidance and support, fourteen trials utilized semi-guided approaches such as telephone calls, text messages, or standardized emails to encourage intervention completion or reinforce learning principles in digital programs19,20,21,32,35,37,39,40,41,42,46,47,48,49. Three trials provided support or feedback from "program coaches"17,33 or researchers43 to facilitate intervention adherence and skill practice. One study employed audio-file reminders38, whereas four studies did not specify any explicit guidance methods34,36,44,45.

Risk of bias in included studies
The overall risk of bias across the included studies was assessed as moderate, primarily due to unclear reporting or insufficient methodological details (Figure 2). Although 21 out of the 22 included studies (95.5%) reported random allocation of participants, only thirteen studies (59.1%) explicitly described their randomization methods17,20,21,32,35,37,39,41,42,46,47,48,49, which included random number tables or lists21,37,42,47, computer-generated sequences17,32,35,39,41,49, randomized software programs46,48, and coin toss20. One study was rated at high risk for allocation concealment, as only the first participant was randomly assigned, with subsequent allocations occurring based on the characteristics of previously enrolled participants38. Blinding of outcome assessors was deemed feasible; however, participant blinding was often impractical due to the nature of the control conditions. Three studies reported that neither participants nor researchers were blinded17,21,41. One study blinded both participants and research assistants39, and another implemented single-blinding by researchers20. Eight studies explicitly applied ITT analysis, although a total of 502 participants were not included in these analyses32,35,36,40,41,42,45,47.

Figure 2: Risk of bias in included studies. (A) Summary of risk of bias across domains. (B) Risk of bias by domain for each study. Please click here to view a larger version of this figure.

Effects of digital health interventions on stress, anxiety, and depression
The pooled effect estimates for each outcome are summarized in Table 2. Forest plots illustrating the outcome analyses are presented in Figure 3 to Figure 5, and the results of the publication bias assessments using funnel plots are shown in Figure 6 to Figure 8.

Table 2: Pooled effects of digital health interventions on target outcomes compared with control groups. Please click here to download this Table.

Effects of digital health interventions on stress
Fifteen studies involving 1,076 participants in the digital health intervention groups and 958 in the control groups assessed stress (Figure 3). A high degree of heterogeneity was observed (I2= 81%, P<0.001), and a random-effects model was applied. The meta-analysis revealed that digital health interventions significantly reduced stress compared to the control group (WMD = -1.79; 95% CI: -2.51, -1.07; P<0.001).

Figure 3: Forest plot of the effects of digital health interventions on stress. Please click here to view a larger version of this figure.

Effects of digital health interventions on anxiety
Fourteen studies, including 1,041 participants in the intervention groups and 1,129 in the control groups, evaluated anxiety outcomes (Figure 4). Heterogeneity was moderate (I2= 51%, P = 0.005), and a random-effects model was used. Digital health interventions were associated with a statistically significant reduction in anxiety (WMD = -1.73; 95% CI: -2.20, -1.25; P < 0.001).

Figure 4: Forest plot of the effects of digital health interventions on anxiety. Please click here to view a larger version of this figure.

Effects of digital health interventions on depression
Fifteen studies, comprising 973 subjects in the intervention groups and 946 in the control groups, reported depression outcomes (Figure 5). Due to substantial heterogeneity (I2= 64%, P < 0.001), a random-effects model was employed. The meta-analysis indicated a significant beneficial effect of digital health interventions on depression (WMD = -2.05; 95% CI: -2.91, -1.19; P < 0.001).

Figure 5: Forest plot of the effects of digital health interventions on depression. Please click here to view a larger version of this figure.

Subgroup analysis
The results of the subgroup analyses for stress, anxiety, and depression are presented in Table 2. Multiple significant between-subgroup differences were identified. The choice of assessment scale significantly moderated the effects on stress, anxiety, and depression (all P < 0.001). Intervention technique also served as a significant effect moderator across all three outcomes (all P < 0.001). With respect to stress and depression, compared with the control conditions, both the CBT-based and the MBI-based conditions significantly reduced stress. With respect to stress, CBT-based interventions had the greatest effect size (WMD = -3.08, 95% CI:-5.44, -0.71), which was greater than that of MBI (WMD = -2.70, 95% CI: -3.91, -1.50). Similarly, for depression, CBT-based interventions also yielded the most substantial benefit (WMD = -3.48, 95% CI: -5.33, -1.63). For anxiety, OPI-based interventions demonstrated the strongest effects (WMD = -2.73, 95% CI: -4.19, -1.26).

Guidance modality significantly moderated outcomes for all three measures (all P < 0.001). With respect to stress, the mixed feedback and reminder approach yielded the greatest effects (mixed) (WMD = -2.67, 95% CI: -3.78, -1.56). With respect to depression, reminder-based guidance had the strongest effect (WMD = -2.1, 95% CI: -3.70, -0.49).

Delivery mode was a significant moderator for all outcomes (all P < 0.001). Compared with mobile applications, web-based programs demonstrated superior effects across all the measures: stress (WMD = -3.07, 95% CI: -5.23, -0.92), anxiety (WMD = -3.2, 95% CI: -4.79, -1.62), and depression (WMD = -2.24, 95% CI: -3.67, -0.81).

Treatment duration significantly moderated effects across all outcomes (all P < 0.001). Interventions lasting 4-8 weeks demonstrated the greatest effect sizes compared with shorter or longer durations: stress (WMD = -3.7, 95% CI: -5.02, -2.39), anxiety (WMD = -2.77, 95% CI: -3.46, -2.08), and depression (WMD = -4.1, 95% CI: -5.31, -2.89).

Recruitment method significantly moderated effects for all outcomes (all P < 0.001). In terms of stress and depression, online recruitment strategies yielded stronger effects than other methods (stress: WMD = -4.97, 95% CI: -8.42, -1.52; depression: WMD = -4.66, 95% CI: -7.88, -1.44). For anxiety, mixed online and campus recruitment had the strongest effects (WMD = -1.69, 95% CI: -2.53, -0.85).

Across all the outcomes, compared with the passive control group, the interventions demonstrated significantly greater effect sizes (stress: WMD = -2.46, 95% CI: -3.56, -1.36; anxiety: WMD = -2.32, 95% CI: -2.89, -1.75; depression: WMD = -2.61 95% CI: -3.92, 1.29). Compared with the active control group, the intervention effect sizes were smaller (stress: WMD = -1.34, 95% CI: -3.05, 0.37; anxiety: WMD = -0.38, 95% CI: -1.24, 0.48; depression: WMD = -1.29, 95% CI: -2.22, -0.35).

Publication bias
Publication bias was assessed by visual inspection of funnel plots and Egger's linear regression. For stress (15 trials), the funnel plot was broadly symmetrical, with most studies clustering in the upper-mid section (Figure 6). Egger's test indicated no significant asymmetry (P = 0.259). With respect to anxiety, all 14 included studies were clustered in the upper-middle section of the funnel plot (Figure 7), and the results of Egger's test were not significant (P = 0.424). For depression (15 trials), studies were dispersed throughout the middle and upper portions of the funnel (Figure 8), with no evidence of asymmetry (P = 0.793).

Figure 6: Funnel plot of stress. Please click here to view a larger version of this figure.

Figure 7: Funnel plot of anxiety. Please click here to view a larger version of this figure.

Figure 8: Funnel plot of depression. Please click here to view a larger version of this figure.

Data Availability
This study is a systematic review and meta-analysis. All the data underlying the findings presented in this manuscript have been extracted from the previously published studies cited in the reference list. The complete dataset generated during the extraction process is provided as a Supplementary File with this article.

Supplement 1: eTable 1-The complete search strategy. Please click here to download this File.

Supplement 2: The extracted data from all included studies. Please click here to download this File.

Supplement 3: PRISMA 2020 Main Checklist. Please click here to download this File.

The present meta-analysis identified 22 RCTs with 3,655 participants to examine the effectiveness of digital health interventions in reducing stress, anxiety, and depression among university students. As a scalable, low-cost, and readily accessible form of psychological support, digital health interventions have demonstrated strong implementation potential in the college population. Our systematic review and meta-analysis demonstrate that digital health interventions lead to significant reductions in students' stress, anxiety, and depression. Web-based programs were found to be more effective overall than smartphone applications, and a medium-term intervention duration of 4-8 weeks yielded the greatest benefits for all three mental health outcomes. Collectively, digital health interventions furnish a flexible, confidential channel for psychological support and constitute a valuable adjunct to traditional campus mental health services.

Our study revealed that the intervention technique served as a significant effect moderator across all three outcomes. Compared with the control conditions, both online CBT and MBIs significantly reduced stress and depression, which was consistent with the findings of previous studies^{17,20,21,33,35,54}. Digital health interventions that incorporate CBT and psychological intervention content can help students effectively identify mental health disorders and play a significant role in promoting the mental well-being of university students^55,56,57. Additional meta-analytic evidence further confirms that online MBIs are particularly efficacious for improving university students' mental health¹⁶. This may be attributed to participants' preference for the flexibility and accessibility of online CBT and MBIs for self-monitoring combined with relaxation techniques. These approaches may help users identify and track evidence countering worrisome predictions in real time, enabling them to challenge anxious expectations and alleviate stress⁵⁸. Compared with single intervention measures, online CBT and MBIs demonstrate multiple benefits, particularly in terms of significant improvement in patients' social functioning and prolonged remission of depressive symptoms⁵⁹. These findings on digital health interventions contribute empirical evidence for constructing a multidimensional mental health intervention system, of which they are an integral component.

The overall results across included studies showed that the experimental groups had superior scores on psychometric scales for stress, anxiety, and depression compared to the control groups. While these findings were statistically significant, their clinical relevance requires careful consideration. For instance, the WMD for the stress scale was 1.79. However, when considering the Perceived Stress Scale (PSS), existing literature suggests that the Minimal Clinically Important Difference (MCID) for the PSS is generally 2 points or higher^6,61,62. Although the observed reduction in stress scores fell below the conventional MCID of 2 points, it still represents a positive shift in the desired direction. Population-based research suggests that even a 1- to 2-point decrease in PSS scores can meaningfully shift the distribution of stress at a population level, thereby conferring public health relevance⁶³. Therefore, these findings are best interpreted as indicating a modest yet meaningful reduction in stress at the population leve⁶⁴, which may be particularly valuable as preventive or low-intensity support for students with subclinical symptoms. This principle similarly applies to the anxiety outcomes. The observed WMD of -1.73, while below the conventional clinical benchmark of 3.5-4 points for the GAD-7 scale^65,66, exceeds the MCID of 1.5 points identified in a meta-analysis for individuals with moderate baseline anxiety (GAD-7 score≥8)⁶⁷, a population that is well-represented in digital intervention trials. Collectively, these findings support the position that digital interventions can yield clinically relevant anxiety reduction in students with mild-to-moderate symptoms. Collectively, these findings support the position that digital interventions can yield clinically relevant anxiety reduction in students with mild-to-moderate symptoms. The depression outcome can similarly be interpreted as a modest but meaningful population-level improvement, given that the observed WMD of -2.05 represents a substantial proportion of the PHQ score indicative of mild depression (3.20 points, with a clinical cutoff of 5-15)⁶⁸.

Compared with mobile applications, web-based programs demonstrated superior effects across all the measures (stress, anxiety and depression). These findings align with earlier research underscoring the superior adherence and user-engagement rates associated with web platforms⁶⁹. Network- or platform-based interventions consistently outperformed mobile apps, a finding that aligns with the methodological recommendation from the Cochrane Handbook v6.4 that high-engagement interfaces for digital health interventions can enhance effects⁷⁰. Additional meta-analytic evidence further confirms that web-based MBIs are particularly efficacious for improving university students' mental health, notably in alleviating symptoms of anxiety and depression¹⁶. These results support prioritizing web-CBT or MBI pathway programs on campuses that face counsellor shortages. Although web-based platforms currently demonstrate superior effectiveness compared to mobile applications, future studies should explore hybrid models that integrate both web and app interfaces. Incorporating artificial intelligence (AI) and machine learning (ML) technologies could enable personalized content delivery based on user profiles, behavioral data, and real-time emotional feedback, thereby enhancing intervention precision and engagement¹².

Our studies indicated that a medium-term intervention duration of 4-8 weeks had the greatest effect on stress, anxiety, and depression. A medium-term intervention may be positively associated with improved intervention efficacy, and existing evidence further supports a positive correlation between intervention duration and therapeutic outcomes⁷¹. Additionally, an online intervention study among Irish university students⁴² and another five-week personalized online psychological intervention app targeting Australian university students⁴⁸both demonstrated positive effects on reducing stress, which was consistent with the findings of this study. Our findings suggested that an intervention duration of 4 to 8 weeks may be an "optimal dose" and that prolonged interventions do not necessarily yield greater gains and may risk nonadherence. An intervention duration of 4 to 8 weeks reflected a period during which students demonstrated relatively high compliance. To maximize effectiveness, future digital health interventions for university students are advised to implement a duration of 4 to 8 weeks.

Compelling evidence highlights a significant, unmet need for accessible therapeutic interventions among university students^69,72, with key barriers to implementation including time constraints, personal stigma, and self-discipline⁵⁴. The analytical framework developed in this work provides a replicable model for other digital therapeutics (e.g., targeting sleep or addiction), supplying researchers with the a priori estimates needed to lock in effective components, optimize intervention packages, and calculate sample sizes. As early adopters of digital technology, students represent a population for whom digital health interventions hold particular promise⁷³. This work, therefore, advances the field of digital health by furnishing a novel methodological vantage point with substantial academic value and broad translational potential.

However, despite the use of a random-effects model, the very high level of statistical heterogeneity (I² = 81%) observed for the stress outcome indicates substantial variation in effect sizes across the included studies. While our prespecified subgroup analyses explored several important moderators, such as treatment length and delivery mode, a significant portion of this heterogeneity remains unexplained. This suggests the influence of other clinical and methodological factors that we were unable to quantitatively assess. For instance, there was considerable variation across studies in the intensity of the interventions (e.g., recommended daily versus weekly use), the level of human support (ranging from fully automated to therapist-guided programs), and the baseline severity of participants' stress symptoms. Furthermore, differences in intervention duration, specific technological features, and study settings may have contributed. Therefore, while our meta-analysis confirms the overall efficacy of digital interventions for stress, the high heterogeneity underscores that the effect is not uniform. The findings highlight the need for future primary studies to consistently report on these implementation and contextual factors, enabling a more granular understanding of what works, for whom, and under what conditions.

This study has several notable strengths. It provides a unique contribution by conducting a meta-analysis that distinctly compares the efficacy of two predominant modes of digital health delivery: web-based programs versus mobile applications, specifically for managing stress, anxiety, and depression in university students. The review adhered to rigorous systematic review standards, including a preregistered protocol and duplicate, independent processes for study selection and data extraction. Furthermore, our extensive subgroup analyses across intervention types, therapeutic modalities, and durations provide valuable insights into the potential moderators of effectiveness, moving beyond a monolithic view of digital health.

Our meta-analysis has several limitations. First, the broad categorization of "digital health interventions" combines a wide range of technologies (e.g., chatbots, virtual reality, serious games, interactive web programs, passive symptom tracking apps) and therapeutic modalities (e.g., CBT, mindfulness, acceptance and commitment therapy)⁷⁴. This approach, while useful for generating an overall effect estimate, likely obscures important differences in how these interventions function and their respective effectiveness. Although we performed subgroup analyses by "intervention technique" (CBT, MBI, etc.), we could not fully account for heterogeneity stemming from the specific technology used or the level of human support provided. Consequently, we advise against over-generalizing our findings to all types of digital health tools. Future research requires a more refined framework that concurrently considers the therapeutic component and technological delivery format. A critical next step involves head-to-head comparisons to determine, for instance, whether chatbot-delivered CBT matches the efficacy of human-guided online CBT programs. These direct comparisons are essential for identifying the optimal digital interventions for different user groups.

Second, the long-term efficacy of digital health interventions remains uncertain, as most studies assessed outcomes only immediately after the intervention period. While some interventions lasted eight weeks or longer, few evaluated whether benefits were sustained beyond that point. Our subgroup analysis indicated that interventions of four to eight weeks yielded the strongest effects, suggesting that longer durations do not necessarily improve outcomes and that an optimal therapeutic dose may exist⁷⁵. Future research should prioritize longitudinal designs with follow-ups extending to six or twelve months to determine not only the initial impact but also the durability of benefits and their potential for preventing symptom relapse¹⁷.

Third, the risk of bias present in the included studies introduces uncertainty into the findings. Although the overall risk was rated as moderate, the lack of participant blinding and associated expectancy effects may have led to an overestimation of the intervention benefits⁷⁰. Future research in this area should prioritize refining study designs to mitigate such biases where possible. For example, incorporating objective outcome measures (e.g., physiological markers of stress, behavioral tracking data) alongside self-reported outcomes could help reduce reliance on participant perceptions, thereby minimizing the impact of expectancy effects⁷⁶.

Lastly, the generalizability of our findings may be limited by the geographical and cultural concentration of the included studies, which were predominantly conducted in Western countries. This raises concerns regarding the cross-cultural applicability of digital health interventions, particularly in non-Western settings such as Asia, where differing cultural norms, values, and digital engagement patterns may influence intervention acceptability and effectiveness. Future research should expand geographic coverage to include more non-Western populations. It should also integrate qualitative methods, including interviews, focus groups, and user feedback, to guide cultural adaptation of digital mental health tools⁵⁶.

In conclusion, the findings of this systematic review and meta-analysis indicate that digital health interventions are effective at reducing stress, anxiety, and depression in university students. The delivery mode was identified as a significant moderator of intervention effects, with web-based platforms or programs demonstrating superior outcomes compared to mobile applications. These results underscore the value of integrating digital health interventions as accessible and scalable support for student mental health within higher education settings. To build upon these findings, future research should prioritize several key areas. There is a pressing need for head-to-head comparisons of different delivery technologies to isolate their specific effects. Furthermore, longitudinal studies with follow-up periods of six to twelve months are essential to determine the long-term sustainability of benefits. Finally, greater attention should be paid to the cultural adaptation of interventions and the integration of personalized approaches, such as those leveraging artificial intelligence, to enhance their relevance and effectiveness.