We used a result-based convergent approach to provide a comprehensive synthesis for different research types on the impact of retention and adherence. Simple meta and narrative synthesis was used for quantitative findings and thematic synthesis for qualitative findings. This review is reported in accord with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) and PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses literature search extension) reporting guideline [20,21], and its protocol has been registered on PROSPERO before the process of data extraction (identification CRD420251068217).

We searched 5 electronic databases (PubMed, Web of Science, Embase, Cochrane, and CINAHL) in the last 10 years. The complete search strings for all databases, including Boolean operators, wildcards, MeSH, and applied filters, are provided in section S1 in Multimedia Appendix 1. To maximize the capture of relevant literature, we used a highly sensitive, 2-concept search strategy. This strategy combined (1) terms for metabolic dysfunction-associated steatotic liver disease (MASLD, metabolic dysfunction-associated fatty liver disease, or non-alcoholic fatty liver disease), and (2) terms for digital technologies (eg, eHealth, mHealth, telemedicine, mobile apps, wearables, and web-based interventions). Notably, we intentionally excluded “self-management” or “lifestyle” as a search concept. The rationale for this omission is that the terminology for such interventions (eg, “behavioral modification,” “patient education,” “coaching,” or “digital lifestyle support”) is highly heterogeneous and inconsistently indexed across bibliographic databases. Adding them to the electronic search string would risk prematurely filtering out eligible studies. The final search strategy was developed by the review team, and reference-list screening of included studies was also performed to identify additional records.

Studies were screened for inclusion based on the Population, Intervention, Comparison, Outcomes, and Study Design framework. The detailed inclusion and exclusion criteria are summarized in Table 1.

All retrieved records were imported into EndNote 20.5 (Clarivate), and duplicates were removed. Titles and abstracts of the remaining records were then independently screened by the same 2 reviewers (RP and YX) against the eligibility criteria. Subsequently, full texts were retrieved for all potentially eligible records and assessed independently by the same 2 reviewers. Any discrepancies were resolved through discussion, and a third reviewer (JW) was consulted when consensus could not be reached. The entire process was documented to ensure transparency and reproducibility, and a PRISMA flow diagram was used to illustrate study selection.

Data were extracted independently by 2 reviewers (RP and JW) using a prepiloted, standardized extraction form. Discrepancies were resolved by discussion and, where necessary, by re-examination of the original source text. Missing or ambiguous data were handled by contacting corresponding authors. The following variables were extracted: (1) bibliographic details: first author, year, country, and journal; (2) study design; (3) participant characteristics: enrolled and analyzed sample size, mean age, sex distribution, and mean BMI; (4) intervention characteristics: digital delivery modality, duration (weeks), and interactivity level; (5) retention outcome: numbers enrolled and completing the intervention at the primary end point, and the derived proportion; and (6) adherence outcomes: operational definition (or notation that none was provided), metric type, and reported rate.

Each intervention was classified a priori as high-interactivity or low-interactivity by both reviewers independently. High-interactivity was defined as the presence of bidirectional, personalized communication between the participant and a human professional or peer. Interventions in which all participant-facing communication was unidirectional were classified as low-interactivity. Discrepancies were resolved by consensus.

When retention data reported at multiple time points, the retention rate at the end of the active intervention period was extracted as the primary data point, to maximize comparability across studies with heterogeneous postintervention follow-up designs. Additionally, several included studies were associated with more than 1 publication reporting data from the same intervention cohort. In such cases, all companion publications were retained and cross-referenced, with data extraction anchored to the publication providing the most complete reporting of variables relevant to this review. Data unavailable in the primary source were supplemented from companion publications, with the originating source noted in the extraction form. Studies sharing a participant cohort are identified in sections S2-S4 in Multimedia Appendix 1 with a shared superscript indicator; participant counts were deduplicated before meta-analysis to ensure the independence of observations.

The methodological quality of all included studies was appraised using the Mixed Methods Appraisal Tool (version 2018) [22]. This validated instrument enables consistent evaluation across qualitative, quantitative (randomized controlled trial [RCT] and non-RCT), and mixed methods designs through 5 structured criteria specific to each study type. Each criterion is rated as “Yes,” “No,” or “Can’t tell,” with the overall quality score expressed as a percentage, where each unmet or unclear criterion results in a 20% reduction. For mixed methods studies, the final quality rating is determined by the weaker component between the qualitative and quantitative parts, while in multimethod studies, the mean value of both components is calculated. Studies were categorized as high (80%‐100%), moderate (40%‐79%), or low (0%‐39%) quality. The primary screening was undertaken by RP, and 20% (4/21) of the included studies were independently reassessed by 2 additional reviewers (YX and XY) to ensure consistency. Any discrepancies in scoring were discussed until consensus was achieved. No studies were excluded solely on the basis of methodological quality. A summary of the risk-of-bias assessment is presented in Table 2, and the full item-level Mixed Methods Appraisal Tool ratings are provided in section S5 in Multimedia Appendix 1.

All quantitative analyses were conducted in R (version 4.5.1; R Core Team) using the metafor package (version 4.6). Retention proportions were transformed to the logit scale before pooling to stabilize variance and normalize the sampling distribution. A random-effects model was fitted using the Hartung-Knapp-Sidik-Jonkman method [23] for variance estimation. This method allows the model to estimate the weighted pooled average with greater precision, also reducing the rate of false positives, particularly in meta-analyses with a small number of studies or substantial heterogeneity. The restricted maximum likelihood estimator was used to estimate the between-study variance component (τ²). Pooled estimates and 95% CIs were back-transformed to the proportion scale for reporting. Statistical heterogeneity was quantified using the Cochran Q test, the I² statistic, and the τ² estimate. The I² statistic describes the proportion of total variance attributable to between-study heterogeneity, but does not indicate the absolute magnitude of variation in true effects across populations [24]. Therefore, we calculated a 95% prediction interval for interpreting the weighted pooled average estimate of meta-analysis in the real-world.

Each prespecified variable (mean age, proportion of male participants, mean BMI, and intervention duration) was first tested in univariable models, followed by a multivariable model including conceptually relevant factors. The significance of moderators was assessed using the QM statistic, and the percentage of heterogeneity explained was quantified by R². Three preplanned subgroup analyses were conducted: (1) by digital delivery modality, (2) by study design, and (3) by intervention duration (section S7 in Multimedia Appendix 1). The robustness of the findings was examined using leave-one-out sensitivity analyses, and potential small-study effects were evaluated visually by funnel plot inspection and statistically using the Egger regression test. All statistical tests were 2-tailed, with a significance level of P<.05.

In total, 2 prespecified sensitivity analyses were performed to assess the robustness of the primary pooled retention estimate. First, to evaluate the influence of study design, the meta-analysis was repeated in the subset of randomized controlled trials only. Second, to evaluate the influence of the refined eligibility criteria applied during revision, the 4 previously meta-analyzed studies excluded under these refined criteria were reintroduced and the meta-analysis was refitted on the expanded datasets. Both analyses used the same inverse-variance random-effects model on the logit scale with restricted maximum likelihood estimation of τ², Hartung-Knapp adjustment, back-transformation to the proportion scale, and 95% prediction intervals. Robustness was judged by the consistency of the point estimate, the overlap of confidence and prediction intervals, and the stability of between-study heterogeneity (I² and τ²).

We followed a convergent integrated mixed methods design (per JBI guidance), with quantitative and qualitative strands analyzed separately and then integratedly discussed in discussion.

The certainty of evidence for the primary outcome was assessed using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework by 2 reviewers (HL and NC). Evidence was rated across 5 domains: risk of bias, inconsistency, indirectness, imprecision, and publication bias. The overall certainty was rated as high, moderate, low, or very low, and is reported in the Results section.

Deviations from the registered protocol, including updated disease terminology (nonalcoholic fatty liver disease to MASLD), separation of retention and adherence as distinct outcomes, expansion of eligible digital platforms beyond smartphone apps, and change of statistical method from risk ratio to single-proportion meta-analysis, were determined before data analysis and are reported in full in section S8 in Multimedia Appendix 1.

The first search was performed in June 2025 (22,864 citations) and was followed by an update on April 17, 2026 (3949 citations). Due to the broad search strategy prioritizing sensitivity, the initial screening yielded 17,131 records. At the title and abstract screening stage, the majority of records were excluded based on the following categories of reasons: (1) studies using digital tools solely as measurement or sequencing instruments (eg, digital calipers and next-generation sequencing platforms) without any intervention or management component, (2) studies not involving human participants with MASLD or nonalcoholic fatty liver disease, (3) nonprimary research (editorials, conference abstracts, and letters without data), and (4) studies unrelated to MASLD management. Given the volume of exclusions at this stage, individual record-level exclusion reasons were not tracked, consistent with established practice for large-scale systematic reviews. The remaining 146 studies were screened in full and excluded if applicable, leaving a final set of 21 studies [10,25-44]. The detailed process of study selection is presented in the PRISMA flow diagram (Figure 1).

Our synthesis included 21 studies [10,25-44] published between 2016 and 2025. The geographic distribution of the evidence base was broad, yet heavily concentrated in high-income Western countries, with 3 studies [31,34,38] originating from the United States. Additional contributions came from Europe, including Germany (n=3) [29,43], Italy (n=2) [28,33], the United Kingdom (n=2) [25,40], and Iceland (n=2) [26,35]. Studies from Asia and Australia were also represented (Australia: n=2 [27,42], China: n=2 [36,41], Republic of Korea: n=2 [10,39], Thailand: n=1 [44], India: n=1 [30], and Singapore: n=1 [37]). Methodological approaches were notably heterogeneous. While RCTs (n=8) [10,31,36,37,39,41,43,44] represented a substantial portion of the included literature, the evidence base also comprised a large number of prospective single-arm, feasibility, or observational designs (n=9) [26,27,32-35,38,40,42]. A smaller subset of studies (n=4) [25,28-30] used mixed methods approaches to explore usability, acceptability, or specialist perspectives. Intervention sample sizes varied dramatically, from 22 to 314 participants. Participants were typically characterized by overweight or obesity, with mean ages ranging from 36.7 to 63.5 years across the included studies.

The interventions leveraged various digital modalities to deliver the therapeutic content. Mobile apps (n=11) [10,25,26,30,35-41] were the most common platform, followed by telehealth-based delivery (n=4) [27,34,42,43], web-based programs (n=4) [28,29,32,33], and SMS text messaging interventions (n=2) [31,44]. Moreover, 1 entry was a qualitative study [25].

The primary therapeutic focus was most frequently on comprehensive lifestyle modification (n=11) [10,25-27,30,35,38-42,44] or weight loss (n=7) [28,31,33,34,36,37]. A smaller cohort of studies (n=3) [29,32,43] focused specifically on physical exercise. Intervention duration was highly variable, ranging from short-term 4-week protocols to long-term 24-month follow-up, with 12-week and 6-month being the most common. A key characteristic of most interventions was the integration of human support to facilitate a 2-way feedback loop, contrasting with 1 study [44] that used a purely 1-way information broadcast. This support was delivered by a range of professionals, including lifestyle coaches, dietitians, exercise physiologists, or advanced practice providers. Communication channels were diverse, encompassing asynchronous methods (eg, in-app chat, email, and bidirectional SMS text messaging) and synchronous modalities (eg, scheduled phone calls, group telehealth sessions, and real-time audio-visual supervision).

This mixed methods systematic review provides a comprehensive picture of attrition dynamics in digital MASLD interventions. Quantitatively, pooled retention across meta-analysis was 80%, with substantial between-study heterogeneity that no single participant-level moderator—mean age, BMI, sex distribution—was able to meaningfully explain. Neither intervention duration nor digital delivery modality significantly differentiated retention outcomes, though mobile app–based platforms demonstrated the highest point estimate and lowest within-group heterogeneity (section S7 in Multimedia Appendix 1). Qualitatively, facilitators and barriers operated across 4 intersecting domains: platform and design, human support and professional integration, motivational and behavioral strategies, and patient-level characteristics. Taken together, multifaceted determinants of attrition in digital MASLD self-management cannot be adequately captured by single variation, requiring the integration of quantitative data with patient-centered narrative evidence.

The point estimate observed in this synthesis (80%, 95% CI 72%‐87%) nominally exceeds the retention rates reported across several digital chronic-disease contexts. However, a definitive comparison cannot be established due to the wide PI (95% PI 47%‐95%) and substantial heterogeneity (I²=73.6%). In the field-defining meta-analysis by Meyerowitz-Katz et al [15], the pooled dropout rate across 17 app-based chronic-disease studies was 43% (95% CI 29%‐57%), corresponding to retention of roughly 57%. In smartphone-delivered mental health trials, Linardon and Fuller-Tyszkiewicz [19] reported a short-term study-attrition rate of 24.1%, aligning more closely with the present estimates. Several features of the included MASLD studies may partly contribute to the relatively favorable point estimate: a predominance of structured RCT designs with active follow-up (8/15 studies) [10,31,36,37,39,41,43,44], intervention durations frequently aligned with research-protocol windows rather than open-ended use, and the concentration of dropout reasons in situational rather than motivational categories. Consequently, such high pooled retention is likely a context-specific average rather than a guaranteed outcome for subsequent MASLD digital interventions. Ultimately, attrition is not an anomaly, but an intrinsic, structurally heterogeneous feature of eHealth trials [14].

A cross-cohort analysis of more than 100,000 participants in remote digital studies presented attrition curves with displayed steep early decay [49], and panel analysis of mental-health app usage by Baumel et al [50] found median open-day windows under 1 month. A similar trajectory was also observed in our study, suggesting that the drivers of long-term adherence differ qualitatively from those of early dropout in digital MASLD interventions. Limited by a small sample of 15 studies and substantial intrastratum heterogeneity, our analysis cannot definitively isolate true duration effects from confounding factors. Further studies should focus on exploring the underlying mechanisms that drive these 2 distinct phases, rather than treating attrition as a linear decay.

A critical methodological gap in the current literature is the systematic conflation of retention (study completion) with adherence (intervention fidelity). As previously noted, adherence in eHealth research remains conceptually underdeveloped, frequently misoperationalized, and rarely anchored in a prespecified definition of “intended use” [12]. Compounding this issue, objective platform-derived attrition metrics and patient-reported data are rarely integrated within the same analytical framework. Despite the longstanding recommendation of the WHO-endorsed mERA reporting checklist [51], few studies report fidelity-to-intervention and exposure-to-content in parallel. However, a promising trend toward behavior-specific adherence measurement, particularly for dietary and exercise components, is emerging. This granularity is of high clinical significance, as component-level reporting empowers interventionists to pinpoint specific behavioral failures and tailor support accordingly. Notably, Zeng et al [52] introduced the Exercise and Diet Adherence Scale to quantitatively assess adherence to lifestyle interventions among patients with MASLD. This scale represents a promising structured approach for evaluating compliance. But none of the studies included in this review applied it, which may be due to the digital format of interventions and the time limitation.

While the inductively derived 4-domain structure aligns with broader frameworks for digital chronic-disease interventions [16,53], our synthesis emphasizes that general digital tools must be specifically calibrated to the MASLD population.

The MASLD population is, on average, older, more multimorbid, and more digitally heterogeneous than the populations in which platform comparisons have most often been conducted (eg, mental health and smoking cessation) [1,54]. This profile amplifies the early access barriers, such as extra download steps or nonintegrated links. Consequently, the superior engagement achieved through preexisting messaging ecosystems (eg, LINE and WeChat [Tencent Holdings Ltd]) indicates that architectural embeddedness is a crucial, transferable design principle for this population. To further address this early attrition, stepped-care approaches demonstrate broader potential by calibrating to early engagement signals to offer a structural response [55]. Indeed, adaptive escalation triggered by early dropout markers has been shown to successfully optimize engagement trajectories in broader digital health implementations [56]. Moreover, digital platforms provide a structurally destigmatizing environment where patients can share their experiences without the social costs that drive avoidance behaviors. This affordance is particularly critical given the high prevalence of self-blame and shame documented in populations with MASLD [57,58].

Aligned with previous evidence that self-efficacy, relevant knowledge, and social support constitute the principal motivational substrates for MASLD self-management [59,60]. In MASLD, hepatic steatosis is asymptomatic in early and intermediate stages, biochemical and elastographic improvements occur over months [61], and patients lack the sensory cues that would otherwise maintain motivation in the interval [62]. Digital monitoring of distal proxies (weight and step count) therefore carries a heavier motivational load in MASLD than in conditions with intrinsic biological feedback. Consistently, foundational distinction between autonomous motivation (grounded in personally internalized goals) and controlled motivation (driven by external pressure) by Ryan and Deci [63] predicts that only the former reliably maintains long-term behavior change. This prediction is extensively confirmed in the meta-analytic synthesis by Ng et al [64] across 184 health-behavior studies. Further studies should investigate specific digital design strategies that cultivate this autonomous motivation in long-term MASLD management.

This review underscores the critical influence of both intervention design and patient-related characteristics on adherence. For lifestyle interventions, participant acceptance and adherence are significantly higher when the program resonates with their personal habits and cultural environment. Although digital delivery removes 1 dimension of structural inequity, it does not resolve broadband access, device compatibility, or the disproportionate domestic and caring obligations that, as several included studies documented, fell more heavily on women and minoritized populations [34]. Evidence suggests that participants with higher levels of education and socioeconomic resources generally demonstrate stronger adherence, highlighting the interplay between personal context and the effectiveness of self-management interventions [60]. However, many included studies in this review are pilot trials or feasibility studies, lacking robust sample sizes, long-term follow-up, and the statistical power to draw definitive conclusions. Therefore, future research should focus on implementing and evaluating personalized lifestyle programs within larger, real-world populations, exploring the socioeconomic factors to adherence and access.

While the meta-regression did not identify any study-level moderator that statistically explained between-study heterogeneity (section S7 in Multimedia Appendix 1), the integrated mixed methods analysis nevertheless yielded informative convergence between the 2 strands of evidence. Among delivery modalities, mobile app–based interventions showed the highest pooled retention of 83% (95% CI 73%‐90%; I²=66.9%; τ²=0.310) and, notably, the lowest within-group heterogeneity, suggesting a relatively consistent participant response to this format. This quantitative signal aligns with qualitative themes in which participants repeatedly emphasized the convenience, accessibility, and personalization afforded by smartphone-based platforms (Table 4). We interpret the absence of statistical significance in the subgroup and meta-regression analyses with caution rather than as evidence of no effect. Moreover, 3 factors plausibly account for this discrepancy. First, the included studies were unevenly distributed across delivery modalities, leaving most subgroup comparisons statistically underpowered. Second, design features such as interactivity, feedback, and personalization operate as continuous, multidimensional constructs that are poorly captured by coarse study-level moderator coding and are more faithfully represented in qualitative narrative. Third, participant- and context-level determinants were inconsistently or insufficiently reported at the study level, precluding their inclusion in the meta-regression.

This review specifically interrogates attrition dynamics in digital self-management interventions for MASLD. Whereas previous systematic reviews have predominantly evaluated clinical outcomes [13,65] or analyzed retention and adherence in isolation [15,16], this review integrates these 2 dimensions using a convergent segregated design. Driven by the substantial heterogeneity of adherence data in current literature, this mixed methods approach functionally contextualizes objective pooled dropout rates with qualitative, patient-centered narratives of adherence. This approach yields a comprehensive picture of attrition dynamics, revealing that while retention compares favorably to adjacent fields, sustaining true adherence depends on resolving entry friction, ensuring accessible human support, and aligning feedback with the asymptomatic disease course [61,62]. Although the conclusions warrant cautious interpretation given the predominance of pilot and feasibility designs among included studies, these findings offer a preliminary framework for prioritizing engagement architecture in the design and evaluation of future digital MASLD interventions.

Primarily, few mixed methods systematic reviews have examined adherence and retention to digital self-management interventions for MASLD, and our study adds to this emerging body of evidence. The integration of both quantitative and qualitative evidence provided a more holistic understanding of attrition conditions in digital self-management interventions among patients with MASLD. These findings carry practical and methodological implications for researchers and clinicians alike. Long-term engagement in digital MASLD interventions depends on accessible professional support and contextual integration into routine clinical care, suggesting that implementation design deserves as much attention as technology selection. Future research should prioritize clear and comprehensive reporting of attrition across multiple dimensions, including reasons for dropout, timing of disengagement, and differential dropout rates across intervention components, alongside standardized adherence definitions and multifaceted engagement metrics capturing task completion and behavior-specific fidelity. Longitudinal designs incorporating mediating constructs, such as motivation and self-efficacy, would further clarify how study participation translates into meaningful behavioral change and clinical benefit.

Several limitations warrant consideration. First, the overall certainty of evidence was rated very low under the GRADE framework, reflecting design constraints intrinsic to the digital intervention literature. The transparent nature of digital tools precludes participant and provider blinding in most included studies, and a large number of observatory studies further limit causal attribution. Second, despite the breadth of prespecified moderators examined, none reached statistical significance in subgroup analyses or meta-regression. The constraints of uneven study distribution across delivery modalities, the multidimensional nature of design features, and context-level determinants limit the integration of quantitative and qualitative findings. Third, the operationalization of retention varied across included studies. Although all reported a proportion completing their designated end point, specific definitions, follow-up windows, and counting rules differed and were not always explicitly stated, introducing definitional heterogeneity (I²=73.6%) into the pooled estimate. Fourth, the wide prediction interval indicates that the pooled retention estimate cannot be reliably extrapolated to specific clinical implementations without consideration of local contextual factors. Fifth, adherence was reported in 18 studies among included studies [10,25-27,29-32,34-38,40-44], that the metrics varied widely (eg, logins, app sessions, days of use, and percentage of behavioral targets met), and that this prevented quantitative pooling for adherence. Finally, included studies were predominantly conducted in high-income or East Asian settings, limiting generalizability to low- and middle-income contexts and underrepresented cultural populations.

This mixed methods systematic review found that the pooled retention estimate in digital MASLD interventions nominally compared favorably with broader digital health benchmarks, yet substantial unexplained heterogeneity and a wide PI limit the generalizability of this finding. No participant-level or intervention-level moderator significantly differentiated retention in meta-regression. The qualitative synthesis highlighted facilitators and barriers spanning platform and design, human support and professional integration, motivational and behavioral strategies, and patient-level characteristics. Retention and adherence were empirically distinct across included studies, and pooled retention estimates alone are insufficient to evaluate intervention fidelity. Furthermore, the inconsistent operationalization of adherence constrains meaningful cross-study comparisons. Addressing this lack of measurement standardization must be a priority for methodological harmonization in future trials. Ultimately, identifying for whom and under what conditions these interventions are most effective will require continued innovation in both trial design and implementation strategies.