Patterns of Uptake, Engagement, and Attrition in Randomized Controlled Trials of Digital Interventions for Eating Disorders: A Systematic Review and Meta‐Analysis

ABSTRACT

Objective

This review aimed to quantify rates of uptake (treatment initiation), adherence (program completion), and attrition (study dropout) in randomized trials of digital eating disorder (ED) interventions, and to synthesize engagement reporting practices, their consistency, and associations with clinical outcomes.

Methods

Randomized trials of digital interventions (web, app, computerized, chatbots, etc.) delivered to people with diagnostic, subthreshold, or self‐reported EDs were included. Random‐effects meta‐analyses were conducted to compute absolute rates of uptake, adherence, and attrition, while a narrative synthesis summarized engagement patterns and reporting.

Results

Forty‐eight trials were included. The weighted mean uptake rate from 44 intervention conditions was 89.7% (95% PI = 67.0–97.0) and the weighted mean adherence rate from 14 intervention conditions was 41.8% (95% PI = 0.9–83.3). These estimates remained similar in a series of sensitivity analyses that adjusted for biases, outliers, and when limiting to specific clinical population groups. The weighted trial attrition rate was 23.3% (95% PI = 0.8–54.4); for intervention arms specificially it was 29.7% (95% PI = 9.7–62.3) and for waitlist arms it was 18.7% (95% PI = 4.6–52.3). Attrition was lower in trials that had human‐participant interaction, offered therapeutic guidance, provided monetary reimbursement, tested a web/computer program (compared to a smartphone application), and had a longer follow‐up (> 6 weeks). Reporting of engagement was inconsistent and heterogeneous, with nearly 90 different metrics recorded across trials. There was some evidence linking sustained user engagement to greater clinical benefit.

Conclusion

Findings offer practical benchmarks to inform future trial planning and highlight design elements that could be leveraged to enhance user engagement and retention.

Key Points

• Nearly 9 in 10 participants initiated digital eating disorder interventions, yet fewer than half completed the full program, highlighting a gap between uptake and sustained engagement.

• Trial attrition averaged 20%–30% and was lower in studies incorporating human support, therapeutic guidance, monetary reimbursement, web‐based delivery, and longer follow‐up periods, pointing to actionable strategies for improving retention.

• Engagement was measured inconsistently across trials using nearly 90 different metrics, highlighting the importance of improved reporting standards.

1. Introduction

Enthusiasm is mounting for the potential of digital technologies to expand access to evidence‐based eating disorder (ED) treatment. This is driven by a rapidly growing evidence base demonstrating the efficacy of interventions delivered via the web, smartphone applications (apps), and to a lesser extent, chatbots (Hua et al. 2025; Torous et al. 2025). Yet, concerns persist regarding the quality of this evidence base, as randomized controlled trials (RCTs) of digital ED interventions repeatedly report problems with attrition and engagement (Linardon et al. 2020).

Attrition and engagement are related but distinct constructs. Attrition reflects the extent to which participants do not complete a research trial protocol, typically by failing to complete outcome assessments (Eysenbach 2005). It is a proxy of treatment engagement, given that disengaged participants are more likely to discontinue participation and fail to provide follow‐up data (Torous et al. 2020). Engagement, on the other hand, refers to the extent to which participants actively interact with intervention content, such as initiating, using, and completing program components (Ng et al. 2019). High attrition and low engagement pose threats to the validity of trial findings, because differential loss to follow‐up and limited exposure to intervention content can bias effect estimates and obscure the true therapeutic potential of an intervention (Linardon et al. 2025).

Although engagement and attrition are widely acknowledged challenges in digital ED interventions, remarkably little work has systematically quantified their extent across research trials. Our earlier meta‐analysis (Linardon et al. 2020) of digital ED interventions reported pooled attrition rates of 25.3% (95% CI = 18.4, 33.9) for treatment programs (8 conditions) and 21.0% (95% CI = 14.9, 28.7) for prevention programs (32 conditions). We also extracted engagement patterns but highlighted considerable variability in reporting that precluded any meaningful synthesis. Other, more narrowly focused reviews have found attrition to be significantly higher in digital ED treatment programs relative to non‐active control conditions (Zhong et al. 2024), but comparable to standard psychotherapy conditions (Samara et al. 2024).

The growing number of RCTs testing digital interventions for EDs in recent years offers an opportunity to build upon and extend the findings of prior reviews in several critical ways. First, there is now the potential to quantify the extent of treatment uptake, as this is an engagement metric increasingly reported across digital health trials, reflecting the proportion of participants who log in to and access their assigned digital program (Liu et al. 2026). Low rates of treatment uptake create a mismatch between randomization and actual intervention exposure, which can complicate outcome interpretations and obscure whether null or modest effects reflect intervention ineffectiveness or failure to initiate use. Second, it is now possible to quantify rates of adherence, which is typically operationalized as the proportion of participants completing all required modules, lessons, or sessions. Determining absolute rates of adherence is important for informing judgments about dose delivered versus dose received, which is central to interpreting efficacy estimates in digital intervention trials (Ng et al. 2019).

Furthermore, the growing number of trials now permits more precise estimates of trial attrition, as well as examination of factors that may influence attrition rates. Drawing from other fields, trials that incorporate some degree of human contact, including during onboarding, eligibility screening, or through therapeutic guidance during the intervention, tend to report lower attrition, likely due to increased accountability and sustained participant motivation (Linardon and Fuller‐Tyszkiewicz 2020; Liu et al. 2026). Delivery of repeated prompt reminders has also been linked with reduced attrition (Saleem et al. 2021), possibly by maintaining participants' awareness of trial demands and encouraging continued involvement. Paying participants for completing outcome assessments can also reduce attrition (Griffith Fillipo et al. 2022; Linardon 2023), as financial incentives can offset participant burden and encourage continued completion of follow‐up requirements. The expanded literature on digital ED interventions now provides an opportunity to examine the moderating role of such trial‐related procedures, which may offer insights into strategies that may improve participant retention in future studies.

The present review sought to examine patterns of uptake, engagement, and attrition in trials of digital ED interventions. We aimed to (1) quantify the extent of uptake, adherence, and attrition, (2) uncover factors associated with attrition rates, and (3) synthesize patterns of engagement with respect to how it has been operationalized and reported, the consistency of metrics captured, and their associations with clinical outcomes.

2. Methods

2.1. Search Strategy and Study Selection

This review was pre‐registered before data extraction (https://osf.io/ejzvr/overview) and was conducted in accordance with the PRISMA guidelines (Page et al. 2021). The MEDLINE, PsycINFO, Web of Science, and Scopus databases were systematically searched in October 2025 using combinations of keywords related to EDs, digital health, and RCTs, with full search terms detailed in the Supporting Information.

We included RCTs that examined the effects of a digital intervention for EDs and reported patterns of uptake, engagement, or attrition. Samples included participants preselected for elevated ED symptoms, as established through either a formal diagnostic interview confirming a threshold or subthreshold diagnosis, scores above a validated cut‐off on a standardized rating scale, or endorsement of ED behaviors at trial intake. Studies using unselected samples, or samples screened only for risk factors (e.g., weight concerns) in the absence of core ED symptoms, were excluded. Digital interventions were defined as those delivered via computers, mobile phones, or tablets, using platforms such as the Internet, downloadable software, CD‐ROMs, smartphone applications, or chatbots. Self‐guided, human‐guided, or blended (e.g., adjunct to traditional psychotherapy) digital interventions were eligible for inclusion. Any comparison group was permitted, and trials comparing multiple digital interventions were also eligible for inclusion. Published and unpublished studies were eligible.

2.2. Risk of Bias and Data Extraction

Four domains from the Cochrane Risk of Bias tool (Higgins et al. 2019) were used to evaluate study quality: random sequence generation, allocation concealment, blinding of outcome assessment, and incomplete outcome data. Each domain was rated as having low, high, or unclear bias. We also extracted characteristics related to the participants, intervention, control group, follow‐up lengths, trial procedures, uptake, attrition, and adherence proportions, and other engagement findings. Two researchers extracted information from eligible trials. Minor discrepancies were resolved through consensus.

2.3. Outcomes

Meta‐analyses were conducted on three outcomes.

1. Intervention uptake: The proportion of participants randomized to the digital intervention who logged in, used, or activated an account at least once.

2. Intervention adherence: The proportion of participants allocated to the digital intervention group who completed all relevant sessions, modules, or skills.

3. Study attrition: The proportion of participants who were randomized to either the digital intervention or comparison group but did not complete the posttest assessment.

2.4. Analytical Approach

We conducted both qualitative and quantitative syntheses of the included studies. The qualitative synthesis examined engagement patterns, focusing on the number and type of engagement metrics reported, patterns observed, and dose–response findings.

We then conducted meta‐analyses estimating absolute rates of uptake, attrition, and adherence. Event rates were calculated for each study as the proportion of participants who accessed the digital intervention (uptake), completed all required content (adherence), or failed to complete posttest assessments (attrition). To permit valid variance estimation and inclusion of studies with zero or complete event rates, proportions were continuity corrected and logit transformed prior to pooling (Borenstein et al. 2009). Meta‐analyses were conducted on the logit scale using inverse‐variance weighting under random‐effects models. Pooled estimates and 95% confidence intervals (CIs) were back‐transformed to the proportion scale and expressed as a percentage for ease of interpretation (Borenstein et al. 2009). We also present the 95% prediction interval (PI), which indicates the range within which the true effect of a future study is expected to fall. Heterogeneity was assessed through the I ² statistic.

We also conducted various sensitivity analyses. First, we recalculated pooled event rates for each outcome when limiting to trials that received a lower risk of bias rating (meeting three or four criteria). Second, we recalculated pooled event rates by applying the trim‐and‐fill method to account for small‐study bias (Duval and Tweedie 2000). Third, we pooled effects while excluding outliers using the nonoverlapping CI approach (Harrer et al. 2021). Fourth, we calculated pooled effects when limiting to trials that sampled specific population groups (e.g., samples with bulimia nervosa, binge‐ED).

We also conducted subgroup analyses to examine potential moderators. These analyses were specified a priori and restricted to attrition outcomes in the digital intervention arms, as attrition was the most frequently reported outcome, and to minimize the number of inferential tests conducted. Subgroup analyses were performed under a mixed effects model (Borenstein et al. 2009). The operationalization of study subgroups can be seen in the Supporting Information. Although fewer studies reported uptake and adherence outcomes, we also present pooled estimates for these outcomes across selected subgroups to facilitate a descriptive, visual comparison of these engagement rates under different participant, intervention, and trial characteristics.

Relative rates of attrition between app and control conditions were also calculated using random effects models. Odds ratios (ORs) were calculated separately for each study, weighted by their inverse variance and pooled to create a total effect (Lipsey and Wilson 2001). Analyses were performed separately for waitlist controls, other controls (i.e., treatment as usual, information resources), and active therapeutic comparisons (e.g., face‐to‐face treatment). To avoid double‐counting, when trials included more than one eligible intervention or control group, effect sizes were averaged within studies before pooling (Borenstein et al. 2009). A OR < 1.0 indicates that attrition rates were lower in the intervention group. For studies with zero events in one or both conditions, a continuity correction of 0.5 was applied to all cells (Sweeting et al. 2004). All analyses were performed using Comprehensive Meta‐Analysis Version 4.

3. Results

3.1. Study Characteristics

A flowchart of the literature search is presented in Figure 1. Forty‐eight trials met full inclusion criteria. Table S1 presents the characteristics of included trials. Twenty‐five trials sampled participants with diagnostic‐level EDs confirmed by structured interviews, including transdiagnostic presentations (k = 8), bulimia nervosa (k = 7), and binge‐ED (k = 10). The other trials either recruited subthreshold EDs (confirmed via semi‐structured interviews) or symptomatic individuals, established by scoring above a threshold on a validated questionnaire or by endorsing core ED behaviors at baseline. There were 59 digital intervention conditions; delivery formats ranged from the web (n = 39), CD‐ROMs (n = 4), apps (n = 13), chatbots (n = 3), or a mix of these (n = 1). Most programs delivered were guided self‐help in nature (n = 29), with fewer unguided (n = 24) or adjunctive (n = 6) interventions delivered. Most comparison conditions were waitlists (n = 33), and the length of follow‐up ranged from two to 24 weeks. Only 10 trials paid participants for completing assessments, and 20 trials reported delivering recurring reminders to engage. Thirty‐three trials received a lower risk of bias rating, defined as meeting three or four of the risk of bias criteria (see Table S1).

FIGURE 1.

Flowchart of literature search.

3.2. Qualitative Synthesis

3.2.1. Engagement Metrics Reported

Table S2 presents the engagement metrics captured across trials. Of the 48 trials, 9 reported no metrics of engagement. Of the 39 trials that did, four relied on participants' self‐reported engagement, while 35 captured objective metrics. The mean number of engagement metrics reported was 1.83 (min = 0, max = 6, mode = 1). There was substantial variability in how engagement was operationalized, with approximately 88 unique definitions of engagement used. The most common metrics of engagement reported were the number of symptom monitoring/diary completions (k = 16 trials), the number of sessions/modules completed (k = 19 trials), the number of activities/skills/posts completed (k = 9 trials), the number of program logins or views (k = 5 trials), and the mean or total number of days of intervention use (k = 4 trials). Other examples of engagement metrics captured included time spent on the program, time to treatment commencement, percentage of program or pages accessed, duration of use, and time between sessions (see Table S2).

3.2.2. Dose–Response Associations

Nine studies reported exploring associations between markers of engagement and clinical outcomes. Five studies reported no significant association between various engagement metrics (e.g., module completions, journal entries, page opens) and the degree of symptom change (de Zwaan et al. 2017; Flatt et al. 2022; Hartmann et al. 2024; Jones et al. 2008; Saekow et al. 2015). Four studies, however, did identify an association: Fitzsimmons‐Craft et al. (2020) and Pruessner et al. (2024) both found that greater program completion rates predicted larger reductions in symptom scores in binge‐spectrum presentations. Hildebrandt et al. (2017) found that greater reductions in binge eating among participants receiving CBT plus the Noom app were partially mediated by the percentage of days on which participants recorded consuming three meals per day. Schmidt et al. (2008) reported that participants with bulimia nervosa who attended four or more CD‐ROM intervention sessions were more likely to experience remission (62%) than those who attended fewer than four sessions (33%).

3.3. Quantitative Synthesis

3.3.1. Intervention Uptake

The weighted pooled uptake rate from 44 digital intervention conditions was 89.7% (95% CI = 87.0, 92.0), with high heterogeneity (I ² = 78%) and a broad 95% PI (67.0, 97.0). Weighted uptake rates were similar when applying the trim‐and‐fill procedure (86.3%), when removing higher risk of bias trials (88.8%), and when removing outliers (91.7%). Uptake rates were also comparable when restricting the analyses to specific diagnostic subtypes, delivery formats, and treatment modalities, ranging from 87.1% to 94.0%. See Table 1 for patterns of uptake across these sensitivity analyses.

TABLE 1.

Meta‐analyses on rates of uptake, adherence, and attrition.

Outcome	Analysis	N	Estimate (95% CI)	I 2	95% PI
Uptake	Total effect	44	89.7% (87.0, 92.0)	78%	67.0, 97.0
	Trim‐and‐fill	12 trimmed	86.3% (82.8, 89.2)	—	—
	Higher risk of bias removed	29	88.8% (85.2, 91.3)	79%	65.6, 97.0
	Outliers removed	33	91.7% (89.5, 93.4)	63%	78.2, 97.1
	Clinically diagnosed samples only	17	90.7% (84.9, 94.4)	79%	55.5, 98.7
	Bulimia nervosa	8	91.2% (81.4, 96.1)	75%	44.1, 99.3
	Binge‐eating disorder	7	91.1% (79.6, 96.4)	87%	28.1, 99.6
	Delivery format a
	Web/computer	32	89.9% (86.3, 92.6)	77%	62.0, 98.0
	Smartphone app	9	90.8% (83.9, 94.9)	83%	56.8, 98.7
	CBT digital intervention	34	90.2% (87.4, 92.5)	73%	71.3, 97.2
	Modality
	Guided self‐help	22	90.1% (86.0, 93.1)	74%	63.8, 97.9
	Unguided self‐help	17	87.1% (82.2, 90.9)	80%	62.9, 96.4
	Adjunctive care	5	94.0% (90.4, 96.4)	0%	87.0, 97.4
Adherence	Total effect	14	41.8% (31.0, 53.5)	93%	0.9, 83.3
	Trim‐and‐fill	0 trimmed	41.8% (31.0, 53.5)
	Higher risk of bias removed	12	39.9% (28.5, 52.6)	93%	0.8, 52.6
	Outliers removed	9	41.5% (34.0, 49.6)	79%	19.6, 67.4
	Clinically diagnosed samples	9	52.3% (40.3, 64.0)	88%	16.4, 86.0
	CBT digital intervention	10	46.9% (32.1, 62.2)	93%	0.7, 90.3
	Modality b
	Guided self‐help	9	45.0% (31.9, 58.8)	92%	0.9, 86.2
	Unguided self‐help	4	28.6% (15.6, 46.6)	93%	0.1, 94.2
Attrition	Total effect (all trial arms combined)	48	23.3% (21.9, 29.1)	92%	0.8, 54.4
	Trim‐and‐fill	11 trimmed	31.4% (27.3, 35.7)
	Higher risk of bias removed	31	26.6% (22.4, 31.4)	93%	0.9, 56.4
	Outliers removed	26	24.8% (21.9, 27.9)	79%	13.5, 41.0
	Bulimia nervosa samples only	8	18.6% (11.8, 28.0)	86%	3.7, 57.5
	Binge‐eating disorder samples only	9	23.2% (13.4, 37.1)	94%	2.5, 78.0
	Total effect (digital intervention arms)	58	29.7% (25.7, 34.0)	87%	9.7, 62.3
	Trim‐and‐fill	10 trimmed	34.0% (29.7, 38.06)
	Higher risk of bias removed	38	31.3% (26.4, 36.7)	89%	10.2, 64.7
	Outliers removed	22	27.1% (23.2, 31.5)	73%	13.4, 47.3
	Bulimia nervosa samples only	10	21.8% (14.6, 31.3)	72%	5.6, 56.8
	Binge‐eating disorder samples only	9	31.0% (19.5, 45.5)	88%	4.7, 80.4
	Total effect (waitlist control arms)	31	18.7% (14.5, 23.7)	87%	4.6, 52.3
	Total effect (other control arms)	7	22.3% (14.5, 32.6)	80%	5.3, 59.5
	Total effect (active therapy arms)	8	30.6% (20.7, 42.6)	77%	7.6, 70.2

^a Chatbots (two conditions) and mixed delivery formats (one condition) were also featured but not included here due to the low number of conditions.

^b Adjunct treatment only featured in one condition so this was not presented.

3.3.2. Intervention Adherence

The pooled adherence rate from 14 intervention conditions was 41.8% (95% CI = 31.0, 53.5), with high heterogeneity (I ² = 93%) and a PI of 0.9–83.3. This estimate remained comparable when applying the trim‐and‐fill estimate, and when removing higher risk of bias trials and outliers. A pooled adherence rate of 52.3% was observed for trials with clinically diagnosed samples. Guided self‐help interventions produced a pooled adherence rate of 45.0% compared to 28.6% found for unguided interventions (Table 1).

3.3.3. Trial Attrition

The total weighted attrition rate from 48 trials was 23.3% (95% CI = 21.9, 29.1), with high heterogeneity (I ² = 92%) and a 95% PI of 0.9, 56.4. Comparable estimates were observed in the series of sensitivity analyses, except that a pooled attrition rate of 31.4% was observed when applying the trim‐and‐fill correction.

For digital intervention arms only, the weighted pooled event rate from 58 conditions was 29.7% (95% CI = 25.7, 34.0), with a PI of 9.7, 62.3. Attrition rates ranged from 21.8% (samples with bulimia nervosa) to 34.0% (trim‐and‐fill adjustment) in the sensitivity analyses.

For comparison conditions, the pooled event rate was 18.7% (95% CI = 14.5, 23.7) for waitlists, 22.3% (95% CI = 14.5, 32.6) for “other” controls, and 30.6% (95% CI = 20.7, 42.6) for active therapy comparisons.

Attrition rates were significantly higher in digital intervention versus waitlist control conditions from 32 comparisons (OR = 2.18, 95% CI = 1.74, 2.74; I ² = 62%; PI = 0.83, 5.75). However, no significant differences in attrition were observed when comparing digital intervention conditions to other control conditions (eight comparisons; OR = 1.14, 95% CI = 0.61, 2.11; I ² = 74%; 95% PI = 0.19, 6.84) and to active psychological therapy comparisons (eight comparisons; OR = 0.92, 95% CI = 0.48, 1.97; I ² = 78%; 95% PI = 0.10, 9.45).

3.3.3.1. Subgroup Analyses

Results from the subgroup analyses of attrition rates for intervention arms are presented in Table 2. Attrition rates were significantly lower when computerized/web programs were delivered (compared to apps), when a guided intervention was used (compared to unguided interventions), when human‐participant interaction was described (compared to when the study was entirely automated), when monetary reimbursement was provided for completing assessments (compared to when participants were not reimbursed), and when follow‐up assessments were 6 weeks or longer (compared to < 6 weeks). Sample type, reminders, and CBT interventions were not significantly associated with attrition rates.

TABLE 2.

Results from the subgroup analyses on attrition rates for digital intervention arms.

Subgroup analysis	N	Estimate (95% CI)	I 2	p
Clinically diagnosed sample				0.143
No	31	32.5% (27.1, 38.4)	89%
Yes	27	26.2% (20.5, 32.7)	84%
Reminder prompts				0.521
No	27	31.1% (25.5, 37.3)	87%
Yes	31	28.3% (22.8, 34.6)	88%
CBT intervention				0.108
No	11	38.2% (26.8, 51.0)	93%
Yes	47	28.0% (23.9, 32.5)	85%
Delivery format a				0.009
Smartphone app	13	39.2% (32.1, 46.8)	82%
Web/computer‐based	42	27.4% (22.6, 32.9)	87%
Guidance offered				0.004
No	20	38.1% (31.6, 45.2)	87%
Yes	38	26.1% (21.8, 30.9)	83%
Human‐participant interaction				0.009
No	15	38.5% (31.1, 46.5)	88%
Yes	43	26.8% (22.3, 31.8)	86%
Participant payment				0.030
No	48	31.7% (27.3, 36.5)	86%
Yes	10	19.7% (12.5, 29.5)	89%
Follow‐up length				< 0.001
≤ 6 weeks	11	47.0% (38.0, 56.3)	88%
> 6 weeks	47	26.0% (21.9, 30.5)	85%

^a Two conditions were chatbots, and one was a mix of web and app. Given the low number, we excluded these and only performed a comparison of web/computer‐based versus smartphone app.

4. Discussion

We synthesized 48 randomized trials of digital ED interventions and examined patterns and reporting of uptake, engagement, and attrition. From 44 treatment conditions, we identified a pooled uptake rate of 89.7% (95% CI = 87.0, 92.0), with a relatively broad PI of 67.0–97.0. This estimate remained stable across different population groups, treatment settings, and intervention modalities. This shows that 9 in 10 participants enrolled in digital ED trials initiate treatment, a finding that closely mirrors uptake rates in digital health trials for depression and anxiety (Liu et al. 2026). This high uptake rate is not surprising, given that trial participants knowingly and voluntarily consent to participate in research, which may select for individuals who are more motivated or enthusiastic about engaging with digital interventions than the broader population. An important next step is to better identify the subset of participants who fail to initiate treatment so that targeted strategies aimed at improving early buy‐in can be developed. This is especially important as digital interventions are implemented beyond clinical trial settings, where opportunities to promote awareness, motivation, and onboarding support are likely to be more constrained.

Despite high rates of treatment initiation, only a minority complete the full digital intervention. The meta‐analytic adherence rate from 14 treatment conditions was 41.8% (95% CI = 31.0, 53.5), with a PI of 0.9, 83.3. This estimate remained stable in various sensitivity analyses and for different population groups, but reduced to 28% when limiting the analyses to unguided intervention conditions. Unguided interventions may have lower completion rates because the absence of human guidance reduces accountability and personalized support, increasing the likelihood of disengagement in the face of technical challenges, motivational barriers, or uncertainty about progress (Perret et al. 2023). These findings provide useful benchmarks for expected program completion rates in digital ED trials, which may inform the design, powering, and interpretation of future intervention studies.

However, these adherence findings should be interpreted with appropriate caution. Less than one‐third of included trials reported program completion rates, which raises the possibility of selective reporting bias. Indeed, our narrative synthesis revealed significant variability in the reporting of engagement, with nearly 90 different operationalizations featured across studies. This variability, together with heterogeneous samples, interventions, and trial procedures, limited the ability to meaningfully synthesize engagement and adherence patterns or to draw firm conclusions about expected usage patterns or their influence on clinical symptoms across studies. These findings emphasize the need for greater standardization and consensus in how engagement is defined, measured, and reported in trials of digital ED interventions, in order to improve interpretability, comparability, and cumulative knowledge in this field. Achieving such consensus is unlikely to rest with any single research group, but rather will require collective effort and dialog among various digital health researchers in the ED field, drawing on the diverse methodological and theoretical perspectives to establish shared priorities for conceptualizing and reporting engagement processes in digital treatment trials (Beintner et al. 2019; Cipriani et al. 2025).

Moving forward, greater consideration is needed for understanding the temporal patterning of engagement. Few trials reported information on the duration between first and last use, spacing of session completions, and whether engagement was sustained across the trial period versus concentrated early and then discontinued. This distinction is necessary, as similar adherence rates may reflect qualitatively distinct usage trajectories. More granular reporting of engagement trajectories would allow for more nuanced interpretation of these adherence estimates and better inform intervention design and optimization.

On the other hand, attrition is a commonly reported and standardized trial metric that is central to evaluating the validity of study findings and is considered a close proxy for treatment engagement. The mean trial attrition rate from 48 studies was 23.3% (95% CI = 21.9, 29.1), with a broad PI (0.8–54.4), and comparable estimates were found for specific populations. Attrition increased to nearly 30% (95% CI = 25.7, 34.0) in the 58 intervention arms, while in the comparison arms it was 18.7% (95% CI = 14.5, 23.7) for waitlists, 22.3% (95% CI = 14.5, 23.7) for other controls, and 30.6% (95% CI = 20.7, 42.6) for active therapies. Finding that around one in four to five participants fail to complete follow‐up assessments is broadly comparable to the pooled attrition rates reported in digital health trials applied to other mental disorders (Liu et al. 2026), suggesting that retention challenges are not unique to digital ED interventions.

Subgroup analyses may shed light on possible factors that could minimize attrition. Trials that included some form of human‐participant contact, either during onboarding, eligibility screening, or through regular therapeutic support during treatment, demonstrated lower attrition rates compared with trials that were entirely automated and delivered without human involvement. This is a well‐replicated finding (Linardon and Fuller‐Tyszkiewicz 2020) that may be explained by the therapeutic alliance and accountability created through human interaction, which helps participants navigate obstacles, maintain motivation, and feel personally invested in completing the trial. We also show that attrition was approximately 38% lower in trials that paid participants for completing trial assessments. Payment may enhance retention by increasing extrinsic motivation to complete assessments and signaling the value placed on participants' time and contributions (Griffith Fillipo et al. 2022). While payment appears to be an effective retention strategy, it is important to acknowledge that compensating participants may not be feasible for all research groups, particularly those with limited funding conducting large‐scale remote trials that recruit hundreds of participants.

Another noteworthy finding was that delivery format was associated with attrition, with trials delivering apps demonstrating 44% higher attrition than those delivering web or computer‐based programs. This difference may reflect the distinct intended use and design philosophies of these formats. Web and computer‐based programs typically feature structured, session‐based content with extensive psychoeducation delivered in discrete lessons that have clear start and end points, fostering a more formal commitment to the intervention (Andersson 2024). In contrast, apps are designed for brief, frequent, and on‐the‐go interactions that allow users to engage with therapeutic content in real‐time during moments of need (Mohr et al. 2019). While this portability and flexibility may enhance ecological validity and momentary support, the more fragmented and context‐sensitive nature of app engagement may inadvertently reduce users' sense of structured investment in the intervention and commitment to completing trial assessments, thereby contributing to higher attrition rates. However, it is also possible that this difference reflects temporal or cohort effects, as apps tend to be evaluated in more recent trials. In this context, higher attrition may partly reflect broader shifts in how individuals engage with digital technologies, including increased multitasking and reduced sustained attention, rather than format‐specific design features alone.

There are important limitations to the present review that should be considered. First, substantial statistical heterogeneity was observed across several meta‐analytic estimates, as reflected by high between‐study variability and relatively wide PIs. Although we conducted an extensive series of sensitivity and subgroup analyses that reduced heterogeneity in some cases, meaningful variability remained, indicating that pooled estimates should be interpreted as average effects that may not apply uniformly across all contexts. Second, the limited number of trials available for certain diagnostic groups precluded stratified or subgroup analyses by specific ED populations. Thus, pooled estimates largely reflect aggregated effects across heterogeneous samples, which may limit the generalizability of findings to particular diagnostic groups or clinical presentations. Third, considerable heterogeneity in the reporting of engagement metrics precluded quantitative synthesis of different engagement outcomes, such as frequency of use, number of completed modules, or completion of monitoring activities. Therefore, it was not possible to aggregate rates of specific engagement behaviors across trials. Explicitly documenting this heterogeneity is, however, informative, as it highlights a critical gap in current reporting practices and may help motivate greater standardization of engagement metrics that can be meaningfully synthesized in future meta‐analyses.

In conclusion, this review provides the most comprehensive synthesis to date of uptake, engagement, and attrition in randomized trials of digital ED interventions. Across trials, treatment uptake was high, with approximately nine in 10 participants initiating assigned interventions, whereas adherence to full program completion hovered around 42%. Attrition rates were moderate overall (20%–30%) and comparable to those observed in digital interventions for other mental health conditions. We highlighted substantial heterogeneity in the operationalization and reporting of engagement, which limited our ability to draw firm conclusions about engagement patterns and their links with clinical outcomes. Notwithstanding this, the pooled estimates reported here offer pragmatic benchmarks that can inform expectations around recruitment yield, anticipated completion rates, sample size planning, and power calculations in future trials. More broadly, these findings highlight the importance of moving beyond efficacy alone to better understand how digital ED interventions are used in real‐world settings. Advancing the field will require greater consensus around engagement reporting—for example, adhering to existing reporting standards that have been proposed in the broader digital health field (Beintner et al. 2019)—alongside trial designs and implementation strategies that balance scalability with sufficient human support to promote sustained use, thereby maximizing both scientific rigor and the real‐world impact of digital health tools.

Author Contributions

Claudia Liu: conceptualization, writing – original draft, writing – review and editing, data curation. Cleo Anderson: conceptualization, writing – editing, data curation. Mariel Messer: conceptualization, writing – review and editing. Zoe McClure: conceptualization, data curation, writing – review and editing. Jake Linardon: conceptualization, writing – original draft, methodology, supervision, writing – review and editing.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1: Supporting Information.

Liu, C. , Anderson C., Messer M., McClure Z., and Linardon J.. 2026. “Patterns of Uptake, Engagement, and Attrition in Randomized Controlled Trials of Digital Interventions for Eating Disorders: A Systematic Review and Meta‐Analysis.” International Journal of Eating Disorders 59, no. 5: 854–863. 10.1002/eat.70046.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.