The review was guided by the following Population-Exposure-Outcome (PEO) framework: Population: adolescents aged 10‐19 years; Exposure: digital health interventions (web-based, app-based, or text message–based) targeting one or more of 5 lifestyle health domains (physical activity, dietary habits, obesity management, alcohol consumption, and tobacco use); Outcome: adherence to the digital intervention, operationalized as the extent of use relative to intended use parameters specified by intervention developers, including completion rates and component-level adherence metrics.

A comprehensive literature search was conducted across 4 electronic databases: PubMed, PsycINFO, CINAHL, and Embase in March 2024. No restrictions were applied regarding publication date, language, or geographic region. The search strategy was developed iteratively by one researcher (NB) and used combinations of controlled vocabulary terms (MeSH, Emtree, CINAHL Subject Headings, and PsycINFO Thesaurus terms) and free-text keywords across 3 key concepts: adherence or engagement, digital health interventions, and specific health domains (physical activity, diet, obesity, alcohol, and tobacco). While our inclusion criteria encompassed internet-, computer-, and smartphone-based interventions, the term “digital interventions” is used throughout this review for the sake of brevity. The initial screening process involved reviewing titles and abstracts, followed by a detailed assessment of full-text articles for studies that potentially met our inclusion criteria. Two researchers from our team independently performed the search and screening processes, resolving any discrepancies through discussion until consensus was reached (OS and PdR). Discrepancies at each stage were resolved through discussion between the two reviewers. In cases where consensus could not be reached, the senior members of our team (NB and SH) were consulted. The complete search strategy, including detailed search strings, can be found in Multimedia Appendix 1. The search was updated on June 22, 2026, to capture studies published since the original search. Using the identical search strings across the same 4 databases (PubMed, PsycINFO, CINAHL, and Embase), we retrieved records covering the period from March 2024 to June 2026. The deduplicated updated records were screened against the original eligibility criteria; title and abstract screening, full-text screening, and data extraction of the newly included studies were performed by 2 researchers (IDY and NB), with discrepancies resolved through discussion. This systematic review was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 guidelines [19].

Studies examining digital interventions targeting adolescents aged 10‐19 years (as defined by the World Health Organization [WHO]) were required to focus on at least one of the following 5 specified health domains: physical activity, diet, obesity management, alcohol use, and tobacco use. Studies including some participants older than 19 years were retained when the intervention primarily targeted adolescents or youth and the sample mean age was within or very close to the 10‐ to 19-year adolescent age range. These 5 health domains were selected based on their established role as modifiable lifestyle factors associated with the prevention of cancer. The selection was informed by the SUNRISE project (Sustainable Interventions and Healthy Behaviours for Adolescent Primary Prevention of Cancer with Digital Tools; Horizon Europe, Grant Agreement Number 101136829), within which this review was conducted.

These domains were selected based on their established links with chronic disease (eg, cancer, cardiovascular disease, diabetes, and so on) prevention and represent modifiable lifestyle factors that have been shown to reduce cancer risk throughout life. Eligible studies were required to provide qualitative or quantitative data on the intended use parameters or actual adherence patterns with digital intervention components as defined by the developers.

No restrictions were applied regarding publication date or language. Interventions had to implement digital health promotion strategies as the primary therapeutic component; studies in which digital strategies merely supplemented in-person interventions were deemed ineligible. Intended use was operationally defined as the minimum level of digital platform interaction specified by the intervention developers as necessary for achieving therapeutic adherence. Actual use, on the other hand, encompassed the adherence patterns that were empirically measured within the study population. Studies combining multiple intervention modalities, such as automated messaging platforms integrated with counseling calls, were included, provided adherence measures could be isolated and attributed specifically to digital components. Supplementary elements were categorized as “personal support.”

To focus on primary empirical evidence, we excluded systematic reviews and meta-analyses. Single-session interventions were excluded as they preclude meaningful longitudinal adherence assessments, resulting in universal participation rates that provide limited insight into sustained adherence patterns. Studies that used digital platforms solely for communication (eg, email-based coaching or video consultation sessions) were excluded as they represent interpersonal rather than digital behavioral intervention modalities. Additionally, studies that primarily targeted families or other populations rather than adolescents directly were deemed to be outside the scope of this review.

Data from studies included in the initial systematic literature search were extracted independently by 2 researchers (OS and PdR), while data for studies identified through the updated systematic search were extracted independently by 2 researchers (IDY and NB). Any discrepancies were resolved through discussion. The following information was extracted for each included study: general study characteristics, study population characteristics, intervention characteristics, and information on intervention use.

General study characteristics included the country, study design, and health domain. Study population characteristics comprised mean age, gender distribution (in %), target group (high school students, university students, and others such as adolescents not enrolled in formal education, those in vocational training programs, community-based youth groups, or clinical populations), intervention setting (online, school-based, home-based, community-based, or clinic-based), presence of clinical diagnosis, number of participants, and inclusion and exclusion criteria. Intervention characteristics included the intervention name, delivery type (web-based, through an app, and through text messages), presence and type of personal support, and provision of participant incentives. Information on intervention use encompassed study duration, number of modules, intended use, actual participant adherence, and attrition rate. Additionally, we extracted data on intervention effectiveness and all reported factors influencing adherence. These factors were classified as either intervention-related, that is, modifiable elements of the intervention design, or user-related, that is, inherent characteristics of the participants. The synthesis of studies identified through the initial systematic search was conducted by OS and reviewed by PdR, whereas the synthesis of studies identified through the updated systematic search was conducted by IDY and reviewed by NB. Factor categorizations and evidence classifications were cross-checked against the original study manuscripts. Any disagreements were resolved through discussion with the senior team members (NB and SH).

To guide our synthesis, we adopted a framework informed by the conceptualization of engagement with digital behavior change interventions proposed by Perski et al [7], which distinguishes between behavioral engagement (extent of usage) and experiential engagement (subjective experience). This review focuses on behavioral engagement, operationalized as adherence: the degree to which participants used the intervention as intended by developers. Adherence was captured through two primary metrics where reported: (1) completion rate, defined as the proportion of participants who completed the entire intervention program (binary measure), and (2) component adherence, defined as the mean proportion of individual intervention components completed by participants (continuous measure). The factors influencing adherence were categorized as intervention-related (modifiable elements of intervention design and delivery) or user-related (participant characteristics), and further classified by evidence type: statistical (inferential testing, P<.05), descriptive (quantitative patterns without significance testing), or qualitative (interviews, focus groups, and open-ended responses).

As a first step, the identified studies were categorized according to their respective health domains. In the second step, factors influencing adherence reported in the reviewed articles were qualitatively evaluated and summarized in individual tables for each health domain (Tables S1-S6 in Multimedia Appendix 1 [20-110]). These factors were then categorized into intervention-related factors that acted as either facilitators or barriers to adherence, as well as user-related factors that either facilitated or hindered adolescents’ adherence to digital interventions. In the third step, the factors were analyzed based on 3 types of evidence: statistical significance (findings derived from inferential statistical analyses with P<.05), descriptive evidence (quantitative patterns identified through descriptive statistics without formal significance testing), and qualitative evidence (findings from interviews, focus groups, and open-ended survey responses). The compiled tables present an integrated view of these 3 evidence types to provide a comprehensive understanding of factors influencing adherence.

To complement the narrative synthesis, we conducted a quantitative synthesis of attrition. For each intervention arm reporting sufficient data, we calculated the proportion of enrolled participants lost during the intervention and meta-analyzed the logit-transformed proportions using a random-effects model, with the between-study variance estimated using restricted maximum likelihood and CIs calculated using the Knapp-Hartung adjustment [111]. Summary estimates and CIs were backtransformed to the proportion scale. Heterogeneity was quantified using τ², I², and a 95% prediction interval. Robustness was examined in sensitivity analyses using a binomial-normal generalized linear mixed model with a logit link [112,113], and a 3-level random-effects model in which estimates from multiple intervention arms were nested within studies [114]. Subgroup analyses and random-effects meta-regression examined health domain, delivery modality, level of personal support, provision of incentives, intervention duration, and mean participant age as potential moderators. Potential small-study effects were explored using funnel plots and an Egger-type regression test [115] because conventional funnel-plot methods may be misleading in meta-analyses of proportions; these analyses were considered exploratory. Reported adherence determinants were not quantitatively pooled because of substantial heterogeneity in their measurement. Instead, we summarized the direction and consistency of each determinant across health domains and evidence types and reported this synthesis in accordance with the Synthesis Without Meta-analysis (SWiM) reporting guideline [116]. Analyses were conducted in R (version 4.5.3; R Core Team) [117] using the metafor package [118].

The original search of the electronic databases was performed on March 12, 2024, and yielded 6831 records. After removal of duplicates, 5385 records remained for title and abstract screening, and the full texts of 408 articles were subsequently examined, of which 85 records (82 studies) [20-76,119-146] were included. The search was updated on June 22, 2026, across the same 4 databases and identified 2298 records; after removal of 413 duplicates, 1885 records were screened by title and abstract, of which 101 were sought for retrieval and assessed as full texts. Of these, 63 were excluded (10 conference posters or abstracts, 21 for an ineligible population, 19 because the digital component was not the primary intervention, 6 for a health domain outside the review scope, 3 for reporting no adherence or engagement data, and 4 for an ineligible study design), leaving 38 newly included records. These corresponded to 34 additional study entries after publications reporting on the same participant cohort were combined, yielding a total of 116 included studies [20-69,71-109,119-144,147]. Figure 1 visualizes the selection process and reasons for exclusion for both the original and updated searches.

The 116 included studies comprised several designs: 77 (66.4%) were randomized controlled trials (RCTs) [24,26,28,31,34,37-41,43,46,48,49,51,53-55,57-60,63,66,68,69,71-77,79,81-83,85,87,92-94,96-98,100-107,109,119-121,123-129,131-140,143,144,147], 36 (31%) were quasi-experimental studies [20-23,25,27,29,30,32-34,36,42,44,45,47,50,52,61,62,64,65,67,78,80,84,88-91,99,108,122,130,141,142], and 3 (2.6%) were observational studies [56,86,95].

Geographically, the included studies spanned 26 countries across 5 continents, with persistent regional disparities. The United States remained the single largest contributor, accounting for 43.1% (50/116) studies, followed by Canada, Spain, and Australia. European studies again formed a substantial group, and the updated search broadened the geographic spread by adding studies from countries not previously represented, including Brazil, Uzbekistan, Vietnam, Taiwan, India, Turkey, and Portugal.

The analysis included 143 comparisons from the 116 included studies, as 23 studies [22,26,35,39,40,52,57,67,81,82,85,87,89,91,97,104,107,108,126,129,134,136,147] contained more than one intervention group relevant to our systematic review (21 contributed 2 intervention arms and 2 [85,108] contributed 4 intervention arms). Studies published on the same participant cohort were combined into single entries in the tables and handled as one study. The median intervention duration was 84 (IQR 42-135; range 2-730) days, with 27/143 (18.9%) comparisons lasting 1‐30 days, 54/143 (37.8%) lasting 1 to 3 months, 44/143 (30.8%) lasting 3 to 12 months, and 4/143 (2.8%) lasting 1 year or longer. Fourteen comparisons out of 143 comparisons (9.8%) did not report intervention duration. Of the 143 comparisons, 79 (55.2%) provided no incentives to participants, 44 (30.8%) offered monetary or material rewards such as cash, gift cards, vouchers, prize draws, or a retained device, and 20 (14%) offered nonmonetary incentives such as academic credit or gamified points. A detailed description of the study characteristics can be seen in Tables S7-S18 in Multimedia Appendix 1 [20-110,119-144,147].

The included studies comprised a total of 46,029 participants summed across intervention and control arms, with a median of 52.5 (IQR 28-232; range 10-5260) participants per study. Of the 142 comparisons that reported exact participant numbers, 91 (64.1%) included fewer than 100 participants, 33 (23.2%) had between 100 and 500 participants, 9 (6.3%) included 500 to 1000 participants, and 9 (6.3%) had more than 1000 participants. The overall mean age was 14.48 (SD 2.07) years. Across the studies reporting participant sex, female participants comprised 59.1% of the combined sample (115 [20-69,71-87,90-109,119-144,147] of 116 studies reported sex), essentially unchanged from the original cohort’s 59.2.

The great majority of studies enrolled adolescents with no prior clinical diagnosis. Among studies recruiting clinically defined populations, obesity or overweight remained the most common, followed by childhood cancer, congenital or coronary heart disease, type 1 diabetes mellitus, and juvenile idiopathic arthritis. The update introduced clinical populations absent from the original synthesis, including single studies of chronic kidney disease, cystic fibrosis, elevated blood pressure or cardiovascular risk, and adolescents recruited through mental health services, together with a larger group of studies enrolling adolescents who used nicotine or electronic cigarettes.

Across the 143 reviewed comparisons, 71 (49.7%) were delivered through location-independent digital platforms accessible via smartphones or personal devices, 33 (23.1%) were school-based, 19 (13.3%) were home-based, 11 (7.7%) were clinic-based, 5 (3.5%) were community-based, and 4 (2.8%) combined settings such as summer camp plus online or school plus mobile delivery.

Regarding delivery mode, 51 (35.7%) comparisons were app-based and 50 (35%) were internet- or web-based, together accounting for roughly 7 in 10 interventions; 13 (9.1%) paired a wearable activity tracker or accelerometer with a smartphone or web platform, 11 (7.7%) relied exclusively on text messaging, 9 (6.3%) were computer- or software-based, 8 (5.6%) used a mixed multimodal approach, and 1 (0.7%) delivered exercise through immersive virtual reality.

Personal support beyond initial recruitment was provided in 72 comparisons (50.3%), most commonly through research-team contact, health-professional input, or coaching, while the remaining 71 (49.7%) offered no personal support or relied solely on automated prompts and reminders.

Effectiveness was determined based on whether studies demonstrated statistically significant improvements in their prespecified primary outcomes, distinct from adherence metrics; 52.8% (47/89) of the interventions evaluated for effectiveness were reported as effective. The mean intervention completion rate across the 68 comparisons reporting it was 66.9% (SD 26%), and component adherence averaged 55.2% (SD 25.5%) across the 66 comparisons that reported it, with the remaining comparisons not reporting full completion or quantitative adherence data.

As displayed in Tables S7-S18 in Multimedia Appendix 1 [20-110,119-144,147], the included studies were categorized into 5 distinct health domains based on their intervention focus (physical activity, diet, obesity, alcohol use, and tobacco use), with an additional category for multicomponent lifestyle interventions that targeted multiple health domains. Of the 116 included studies [20-69,71-109,119-144,147], 31 (26.7%) [61-69,71-76,98-109,141-143,147] were multicomponent lifestyle interventions, 32 (27.6%) [20-33,77-86,119-125,144] focused on physical activity, 19 (16.4%) [34-48,127-130] targeted obesity, 14 (12.1%) [55-57,90-94,131-136] targeted tobacco use, 10 (8.6%) [49-54,87-89,126] focused on diet, and 10 (8.6%) [58-60,95-97,137-140] addressed alcohol use.

Within the 31 multicomponent lifestyle interventions, the most common combination was diet and physical activity (12 studies, 38.7% [61,62,64-66,68,74,76,102,104,107,142]), followed by alcohol and tobacco (5, 16.1% [63,69,71,72,143]), diet and obesity (4, 12.9% [100,101,105,147]), all 4 of alcohol, diet, physical activity, and tobacco simultaneously (4, 12.9% [73,75,103,109]), diet, physical activity and obesity (3, 9.7% [67,98,99]), physical activity and obesity (2, 6.5% [106,108]), and diet, physical activity, and tobacco (1, 3.2% [141]).

Across the health domains, dietary interventions showed the highest mean intervention-completion rate (83.9%, SD 21%) and alcohol-focused interventions the lowest (48.4%, SD 35.4%); for component adherence, dietary interventions again showed the highest mean rate (63.7%, SD 23%, essentially level with physical activity interventions at 62.8%), while obesity interventions showed the lowest (45.9%, SD 22.5%). These domain-level figures rest on small and unevenly distributed numbers of reporting comparisons (n=4‐24 per domain) and on widely overlapping ranges, and are therefore indicative rather than definitive. In the meta-analysis of attrition, pooled dropout was lowest for physical activity and highest for alcohol interventions, as detailed in the “Quantitative Synthesis of Attrition” section.

The findings presented below are organized according to 3 distinct levels of evidence synthesis: statistical evidence, descriptive evidence, and qualitative evidence. Where individual studies did not generate all 3 types of evidence, only the categories of evidence that were empirically identified are reported and analyzed in the subsequent sections. For a detailed description of all the reported quantitative adherence values, please refer to Tables S8, S10, S12, S14, S16, and S18 in Multimedia Appendix 1 [20-110,119-144,147].

Across the included studies, 108 comparisons (88 studies [20,23-30,32-34,37-42,46,49-51,53-60,62,64,65,67,69,72-83,85,87-90,92-95,97-102,104,106-110,119-125,128-130,132-134,136-138,140,142,143,147]; 30,956 participants) reported extractable attrition data. The pooled attrition proportion was 16.9% (95% CI 13.6-20.9). Heterogeneity was very high (I²=98.3%; τ²=1.51), and the 95% prediction interval was correspondingly wide (1.7-70.3), indicating that attrition varied substantially across study contexts. Estimates were robust to model choice (exact binomial-normal model: 14.3%, 95% CI 11.1-18.2; 3-level model accounting for the nesting of multiple intervention arms within studies: 17.3%, 95% CI 13.6-21.7).

Pooled attrition was lowest for physical activity (14.1%, 95% CI 9.9-19.7; k=32), diet (14.9%, 95% CI 6.2- 31.5; k=9), and obesity (15.1%, 95% CI 7.3-28.4; k=14) interventions, intermediate for multicomponent interventions (17.9%, 95% CI 11.1-27.8; k=31), and highest for alcohol (25.3%, 95% CI 7.0-60.5; k=8) and tobacco (21.4%, 95% CI 11.9-35.5; k=14) interventions, although the CIs overlapped substantially (Figure 2). None of the examined moderators reached significance in the omnibus Wald-type tests of the meta-regression models (QM): health domain (QM₅=0.79, P=.56), delivery modality (QM₃=1.22, P=.30), personal support (QM₁=0.57, P=.45), incentives (QM₁=1.03, P=.31), intervention duration (QM₁=2.60, P=.11), or mean participant age (QM₁=0.62, P=.43), with each moderator explaining essentially none of the between-study variance. An Egger-type regression test indicated possible small-study effects (z=−2.75, P=.007), with smaller studies tending to report lower attrition, although this test should be interpreted cautiously at this degree of heterogeneity (Figure 3).

Notably, pooled attrition (approximately 17%) was considerably lower than the degree of underuse implied by the completion and component-adherence metrics, under which roughly one-third of participants did not complete the full intervention, and approximately 45% of intervention components went unused. This pattern indicates that many adolescents remained nominally enrolled while substantially under-using the intervention, so that retention statistics alone overstate true adherence. Overall, attrition in adolescent digital lifestyle interventions is substantial but highly heterogeneous and is not explained by standard study-level characteristics.

This systematic review synthesized evidence from 116 studies to identify the factors influencing adherence to digital lifestyle interventions for adolescents. Our analysis revealed that, although just over half (47/89, 52.8%) of the interventions successfully achieved their primary behavioral outcomes, this effectiveness is likely to be undermined by significant adherence challenges. On average, participants completed 66.9% (SD 26%) of the interventions (the mean completion rate across the 68 comparisons that reported it). Mean adherence to individual intervention components was lower still: 55.2% (SD 25.5%) (across the 66 comparisons that reported it). This disparity highlights a critical issue, namely that an intervention is unlikely to be effective if users do not engage with it as intended [148]. Therefore, understanding the drivers of adherence is crucial.

Several intervention-related elements were consistently effective in promoting adherence across multiple health domains. Personalization capabilities, intuitive user interfaces, strategically timed reminder systems, goal-setting functionalities, and gamification elements were found to be particularly influential design features. These findings align with and extend research on adult populations. For instance, the importance of personalization and user-friendly design echoes findings from systematic reviews on digital mental health interventions for adults [10]. However, our review indicates that for adolescents, these elements may need to be combined with gamification and engaging activities to be successful. This is consistent with self-determination theory, which posits that supporting autonomy (personalization), competence (usability and goal-setting), and relatedness (human support) fosters the intrinsic motivation crucial for sustained engagement [5,149]. The effectiveness of gamification elements, such as music integration, interactive quizzes, and competitive features, suggests that interventions that leverage intrinsic motivation through enjoyable activities may achieve better adherence outcomes than purely informational approaches [150]. Importantly, while gamification and “engaging activities” are related, they are operationally distinct in the included studies. Gamification refers to the application of game-design elements (eg, points, badges, leaderboards, and competition) to nongame contexts [151], whereas engaging activities encompass a broader category of design features intended to sustain interest, including interactive quizzes, music integration, narrative storylines, and varied content formats. Both contributed to adherence, but through different mechanisms: gamification primarily leverages extrinsic motivational pathways through reward and competition, while engaging activities may support intrinsic engagement through enjoyment and curiosity.

The prominent role of human support elements, which is particularly evident in obesity interventions through health coach communication and parental involvement, highlights the ongoing importance of interpersonal connections in digital health contexts [5,152]. This finding supports the emphasis placed on social support as a critical determinant of behavior change maintenance in social cognitive theory [153,154]. However, the mixed evidence regarding the effectiveness of social intervention components, such as message boards and expert forums, suggests that not all forms of social interaction enhance adherence. The effectiveness of human support may depend on the quality, timing, and perceived relevance of the interaction, rather than on the mere availability of social features.

Implementing comprehensive human support presents significant scalability challenges, particularly for universal prevention programs, where providing individual attention becomes logistically and economically unfeasible. Constraints on resources in health care and education systems often prevent the provision of sustained professional support for large groups of participants. While hybrid models combining digital tools with personal support can optimize adherence in clinical or targeted prevention contexts, alternative strategies deserve consideration for population-level implementations. These could include peer mentoring systems, AI-driven personalized feedback mechanisms, or tiered support models where human interaction is strategically deployed for participants showing signs of early disengagement or elevated risk [155].

Technical implementation barriers consistently emerged as substantial obstacles to sustained adherence in all health domains, regardless of the focus of the specific intervention. Synchronization difficulties, authentication problems, and inappropriately timed notifications emphasize the importance of robust technological infrastructure and comprehensive, user-centered design principles. These technical challenges were particularly evident in interventions targeting physical activity and alcohol use, potentially due to the real-time tracking requirements and environmental context sensitivity inherent to these health domains. The consistent finding of declining adherence over the duration of interventions [156], particularly in dietary and multicomponent domains, suggests the necessity of adaptive engagement strategies that evolve dynamically to maintain user interest and combat habituation effects.

The synthesis revealed complex and occasionally contradictory relationships between demographic characteristics and adherence patterns. This indicates that associations are context-dependent and cannot be easily generalized. While some studies identified higher adherence among female participants and younger adolescents, these relationships demonstrated considerable variability across health domains and study contexts. This suggests that demographic predictors may be moderated by intervention characteristics, the health domain in question, and cultural factors. This heterogeneity is consistent with ecological systems theory, which emphasizes the multifaceted influences on adolescent behavior, including individual, family, peer, and broader societal factors [157].

Socioeconomic indicators, particularly educational attainment and household income, exhibited more consistent associations with adherence; however, the nature of these relationships varied across health domains. Higher educational levels were associated with better adherence to dietary and alcohol interventions, which may reflect enhanced health literacy and digital competency. However, conflicting evidence regarding the effects of household income on obesity interventions was found, suggesting that socioeconomic influences may operate differently across behavioral targets. These findings emphasize the importance of considering the implications for health equity in the design of digital interventions and the potential need for differentiated implementation strategies across socioeconomic groups.

These socioeconomic patterns in adherence have important implications for health equity that extend beyond the adolescent-specific literature examined here. Systematic reviews examining digital health interventions in adult populations have identified similar socioeconomic gradients in engagement. Szinay et al [158] found that the uptake, engagement with, and effectiveness of mobile interventions for weight-related behaviors were not equally distributed across socioeconomic groups, raising concerns that digital health interventions may inadvertently widen health inequalities. Western et al [159] further argued that the “digital health divide” operates at multiple levels, including individual digital literacy, infrastructure access, and the socioeconomic context of health domains. Our findings in adolescent populations echo these adult-focused reviews and suggest that the equity implications of digital health interventions may emerge early in the lifespan. Interventions targeting adolescents from lower socioeconomic backgrounds may require differentiated design and implementation strategies, such as offline functionality, reduced data requirements, or integration with existing community infrastructure, to ensure equitable engagement.

Psychological factors, particularly intrinsic motivation and the perceived personal benefits of the intervention, emerged as powerful and consistent facilitators. This is particularly relevant for adolescents, who, as noted in the literature, are in a critical period of identity formation during which their cognitive decision-making and critical thinking functions are still maturing [11,12]. Interventions that do not align with their developing personal values or are not perceived as immediately relevant or beneficial are likely to be abandoned, which is consistent with motivation theories that emphasize the need for perceived efficacy and value alignment to encourage long-term participation [149,160]. However, the frequent reporting of loss of motivation and declining interest as barriers to adherence, particularly in the obesity and physical activity domains, highlights the difficulty of maintaining initial enthusiasm throughout extended intervention periods.

Adherence-related outcomes varied across health domains, illuminating the different challenges and motivators associated with specific behavioral targets. In the quantitative synthesis of attrition, dropout was lowest in physical activity interventions, potentially reflecting the immediate feedback and tangible outcomes associated with movement tracking, and highest in alcohol and tobacco interventions, possibly reflecting the sensitive nature of substance-use disclosure and the complex psychological factors surrounding these behaviors in adolescents. Completion and component-adherence rates also differed across domains, although the wide and overlapping ranges caution against overinterpreting any single domain ordering.

Multicomponent lifestyle interventions illustrate the tension between comprehensiveness and adherence. While targeting multiple risk factors simultaneously is theoretically appealing, the expanded evidence base indicated that such programs frequently encountered engagement decline, self-monitoring burden, and competing demands, suggesting that they may present unique adherence challenges that require specialized design considerations to maintain participant involvement across diverse behavioral targets.

Quantitatively synthesizing attrition reinforced and sharpened the narrative findings. Although a substantial minority of adolescents (pooled 16.9%) dropped out of the included studies, the between-study heterogeneity was extreme (I²>98%) and was not explained by health domain, delivery modality, human support, incentives, intervention duration, or participant age. The level of disengagement is highly context-specific and is not predicted by the broad design characteristics that are typically reported, which supports our central recommendation that the field should adopt standardized adherence metrics. The contrast between modest attrition (approximately 17%) and lower intervention use, with mean completion of approximately two-thirds and mean component adherence of approximately one-half, is equally instructive. Adolescents tend to remain nominally enrolled while ceasing to engage with the intervention as intended, so that retention statistics alone substantially overstate true adherence.

These quantitative patterns are consistent with recent syntheses reporting low and declining real-world engagement with digital interventions among young people and a comparable facilitator and barrier structure [161-164].

This systematic review has several important limitations. First, the study protocol was not prospectively registered, and we did not conduct forward or backward citation searching, which may have resulted in the omission of relevant studies not indexed in the searched databases. Second, we did not conduct formal risk of bias assessments of included studies. While validated tools exist for assessing risk of bias across different study designs (eg, the revised Cochrane Risk of Bias tool for randomized trials [RoB 2] [165] and the Risk Of Bias In Non-randomized Studies of Interventions [ROBINS-I] tool [166]), our primary objective was to identify and categorize factors reported as influencing adherence rather than to evaluate intervention effectiveness. Because we do not aggregate effect estimates or draw causal conclusions about intervention efficacy, a formal risk-of-bias assessment of each study’s primary outcomes would not directly inform our narrative synthesis. We acknowledge, however, that this means findings from studies with varying methodological rigor receive equal weight in our synthesis. Third, this review focused on 5 lifestyle-related health domains selected based on their relevance to cancer prevention. The findings may not be generalizable to digital interventions targeting other health domains, such as mental health or sexual health, where the developmental and behavioral contexts differ substantially. Finally, although attrition could be meta-analyzed, the pooled estimates should be interpreted in light of very high between-study heterogeneity and the inconsistent operationalization of attrition and adherence across studies; the estimates are descriptive, are not weighted by study-level risk of bias, and the small-study-effect test should be read cautiously at this level of heterogeneity. Completion and component-level adherence were reported too heterogeneously to support a defensible pooled estimate and were therefore synthesized descriptively and through the SWiM approach rather than meta-analytically.

The findings of this review have several implications for future research and practice. First, the considerable heterogeneity in adherence measurement and reporting represents a fundamental barrier to advancing the field. We recommend that future studies adopt standardized adherence metrics, distinguishing at minimum between completion rates (proportion completing the intervention) and component-level adherence (engagement with specific intervention elements), and report both measures alongside their operationalization. Second, the consistent role of intervention design features across multiple health domains suggests that these elements should be considered core components in the design of digital health interventions for adolescents. Third, the mixed evidence regarding social features highlights the need for experimental research that systematically varies the type, timing, and intensity of social components to determine optimal configurations for different target groups. Fourth, the clear impact of technical barriers across all domains underscores the importance of rigorous usability testing and iterative design processes in intervention development. Finally, future research should explicitly examine the dose-response relationship between adherence and intervention effectiveness. Meta-analyses of individual participant data could provide enhanced statistical power for identifying optimal intervention parameters while accounting for participant-level moderators of adherence [167].

This systematic review provides a comprehensive overview of the factors influencing adherence to digital lifestyle interventions among adolescents. The findings show that intervention-related design features and user-related characteristics consistently affect adherence patterns, although these relationships are often context- and domain-specific. Complementing this narrative synthesis, a random-effects meta-analysis of attrition across 108 comparisons produced a pooled dropout of approximately 17%, but with extreme between-study heterogeneity that none of the examined study-level moderators explained, reinforcing that the degree of disengagement is highly context-specific rather than predictable from broad design features. The substantial variation in adherence rates and measurement approaches highlights the urgent need to establish standardized reporting frameworks that incorporate validated measures and clearly defined usage parameters. Such methodological advancements are essential for synthesizing robust evidence and translating research findings into effective, engaging digital health interventions for adolescent populations.