This systematic review was conducted following the guidelines of the “Preferred Reporting Items for Systematic Reviews and Meta-Analyses” (PRISMA) (Page et al., 2021). The review protocol was registered and described in detail in the international prospective register of systematic reviews (PROSPERO: CRD42020151336).

We conducted systematic searches in the following databases: MEDLINE (via PubMed), CENTRAL (via The Cochrane Library), Scopus, Web of Science, PsycINFO, PsycARTICLES, PsycEXTRA, ProQuest Dissertations and Theses Open, Open Access Theses and Dissertations, and Open Grey. The initial search was conducted between April 6, 2019, and May 3, 2019, with a subsequent search performed between June 1, 2021, and June 4, 2021, to capture studies published after the initial search. The search strategy was designed according to the Population, Intervention, Comparison, Outcomes and Study (PICOS) framework and adapted for each database using a combination of subject headings and free text terms related to aftercare, internet- and mobile-based interventions, depressive/anxiety disorders, and randomized controlled trials (see Supplementary material S1 for the complete search strategy). To ensure the comprehensiveness of our review, we performed an additional check 10–15 December 2023. This check specifically focused on completed trials that were identified in our previous searches but had not been published at that time. We re-examined these previously identified completed trials to determine if they had been published since our last search. For each of these trials, we conducted targeted searches in scientific databases and through general web searches to locate any resulting publications. Any newly published studies from these previously identified completed trials were assessed for eligibility and, if relevant, included in our data extraction process.

Retrieved records were managed using Mendeley. Two independent researchers [LP, MG] screened titles and abstracts to identify eligible studies. The researchers coded the records as eligible, ineligible, or unclear and resolved disagreements through discussion. Inter-rater reliability was calculated using the kappa statistic. Full texts of potentially eligible studies were then assessed for final inclusion.

Data were extracted using COVIDENCE. The following data were extracted from each study: (1) study identification, (2) study design characteristics, (3) intervention details, (4) primary diagnosis, (5) previous treatment type, (6) technical setting, (7) degree of professional support, (8) outcomes, (9) instruments, (10) results. Two independent researchers [LP, PP] extracted the data, and disagreements were resolved through discussion. Data from multiple reports of the same study were extracted into a single form.

The risk of bias was assessed using the Cochrane Collaboration’s tool (Higgins et al., 2011), covering the following domains: sequence generation, allocation concealment, blinding of participants, blinding of personnel, blinding of outcome assessment, incomplete outcome reporting, selective outcome reporting, other sources of bias. Each domain was rated as low, high, or unclear risk of bias. Disagreements were resolved through discussion.

We performed a qualitative synthesis of the included studies. The synthesis focused on summarizing the characteristics and findings of the included studies and discussing the implications of the results. Heterogeneity among studies was assessed based on study design, participant characteristics, intervention details, and outcomes measured. Potential sources of heterogeneity were explored qualitatively.

The systematic search of databases and registers yielded 5,278 studies, with an additional 2 references identified from other sources (Kok et al., 2014; Kraft et al., 2017). After removing 2,063 duplicates (2,061 identified by Mendeley and 2 by Covidence), 3,217 studies were retained for screening. Of these, 3,180 were excluded based on title and abstract review. The full texts of the remaining 37 studies were sought for retrieval, all of which were successfully obtained and assessed for eligibility. Following full-text review, 18 studies were excluded for the following reasons: not published (n = 4), other language (n = 1), wrong outcomes (n = 1), wrong intervention (n = 9), wrong study design (n = 1), and wrong patient population (n = 2). Ultimately, 19 studies met all inclusion criteria and were included in the review for qualitative synthesis. This selection process is illustrated in the PRISMA flow diagram (Figure 1).

The systematic search yielded 19 randomized controlled trials (RCTs), published between 2011 and 2023, comprising a total of 3,206 participants. Sample sizes ranged from 41 (Kraft et al., 2017) to 460 participants (Segal et al., 2020), with a mean of 160.4 participants per study (SD = 123.7). The mean age of participants across studies ranged from 37.2 years (Browning et al., 2012) to 51.0 years (Vicent-Gil et al., 2019). Most studies included predominantly female samples, with the percentage of female participants ranging from 51.1% (O’Reilly et al., 2019) to 80% (Hunkeler et al., 2012).

Most studies (n = 16) primarily focused on major depressive disorder (MDD), while others included participants with anxiety disorders (n = 2), alcohol use disorder co-occurring with depression (n = 2), or persistent depressive disorder (n = 1). Eight studies specifically targeted participants in remission or with a history of depressive episodes. Study designs were predominantly parallel group randomized controlled trials (RCTs), with the number of arms ranging from two to five. Intervention durations varied widely from 2 weeks (Browning et al., 2012) to up to 12 months (Kordy et al., 2016), with follow-up periods extending to 24 months in some cases (Holländare et al., 2013). Studies were conducted in various settings, including outpatient, community, and post-inpatient contexts.

Most interventions (n = 16) were internet- or mobile-based, while five studies utilized primarily telephone-based interventions (e.g., van den Berg et al., 2015; Hunkeler et al., 2012). Professional support ranged from unguided self-help (e.g., Agyapong et al., 2012) to regular therapist contact (e.g., Vicent-Gil et al., 2022), with most studies (n = 12) providing some form of minimal guidance.

Interventions were diverse, including cognitive-behavioral therapy (CBT, e.g., Kok et al., 2015), mindfulness-based approaches (e.g., Kraft et al., 2017; Segal et al., 2020), attentional bias modification (Browning et al., 2012), and supportive text messaging (e.g., Agyapong et al., 2012; O’Reilly et al., 2019). Control conditions were predominantly treatment as usual (n = 12; e.g., Klein et al., 2018), with the remainder using wait-list controls (n = 4; e.g., Schlicker et al., 2017) or active comparators (n = 3; e.g., Browning et al., 2012).

Notably, the studies varied in their approach measuring relapse or recurrence. Some studies used structured clinical interviews (e.g., SCID; Klein et al., 2018), while others relied on self-report measures or predefined cut-off scores on depression scales (e.g., Holländare et al., 2013).

Table 1 presents a comprehensive summary of the characteristics of the studies included in this analysis.

Primary outcomes measured depressive symptoms, most assessed using the Beck Depression Inventory (BDI), Patient Health Questionnaire (PHQ-9) or Hamilton Depression Rating Scale (HDRS). Secondary outcomes often included anxiety symptoms, quality of life measures, and relapse rates (see Table 2). Participant attrition was reported in 17 of the 19 included studies, with varying levels of detail provided.

Table 2 presents a comprehensive summary of the results of the studies included in this analysis.

We conducted a comprehensive risk of bias assessment for all 19 included studies using the Cochrane Risk of Bias Tool (Higgins et al., 2011). This tool evaluates seven domains: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other biases (see Table 3). Random sequence generation was adequately reported in 78.9% (15/19) of the studies, employing methods such as computer-generated randomization, thus demonstrating a low risk of bias. However, 21.1% (4/19) of studies provided insufficient information, resulting in an unclear risk assessment. Allocation concealment procedures were less consistently reported. Only 31.6% (6/19) of studies clearly described appropriate methods, while 68.4% (13/19) lacked sufficient detail, leading to an unclear risk of bias assessment in this domain. Blinding presented a significant challenge, as is often the case with technology-based interventions. Double-blinding was reported in only one study, while 57.9% (11/19) implemented single-blind designs, typically involving blinded outcome assessors. The remaining 36.8% (7/19) did not specify blinding procedures, resulting in an unclear risk assessment. Outcome assessment blinding was adequately reported in 52.6% (10/19) of studies, indicating a low risk of bias. The other 47.4% (9/19) provided insufficient information, raising concerns about potential bias in this domain. Regarding incomplete outcome data, 84.2% (16/19) of studies demonstrated appropriate handling of attrition and missing data, indicating a low risk of bias. However, 15.8% (3/19) were assessed as having a high risk of bias in this area. All studies (19/19) showed a low risk of bias for selective reporting, suggesting consistent reporting of pre-specified outcomes across the research. Other potential sources of bias were unclear in 68.4% (13/19) of studies, primarily due to insufficient information about factors such as baseline imbalances or potential contamination between groups. The remaining 31.6% (6/19) were judged to have a low risk of other biases.

Table 3 presents detailed ratings of the risk of bias domain for each of the studies included in this analysis.

This systematic review synthesized evidence from 19 randomized controlled trials examining the efficacy of digital interventions for depression and anxiety aftercare, revealing several key theoretical pathways through which these interventions may influence mental health outcomes. Of the included studies, 14 reported significant improvements in depressive symptoms (effect sizes ranging from small to large, Cohen’s d = 0.20–0.80), while only 6 studies assessed anxiety outcomes, highlighting a theoretical gap in understanding transdiagnostic effects. The variation in adherence rates (reported in 11 studies) and engagement patterns supports theoretical frameworks emphasizing the importance of human support in digital interventions (Torous et al., 2020a,b). Notably, guided interventions consistently showed higher completion rates compared to fully automated ones, with dropout rates varying from 7.7 to 50% across studies. These patterns, combined with the finding that only 4 studies assessed quality of life outcomes, suggest the need for more sophisticated theoretical models of digital intervention implementation that address both clinical efficacy and user engagement (Torous et al., 2020a,b).

The efficacy of digital interventions for depression appeared to operate through multiple theoretically-grounded mechanisms. Internet-based cognitive therapy (Kok et al., 2015) and mobile cognitive therapy (Klein et al., 2018) both demonstrated promise in reducing depressive symptoms through cognitive modification pathways, supporting cognitive theories of depression that emphasize the role of information processing in symptom maintenance (Bockting et al., 2015). Our findings align with the established efficacy of cognitive therapy in traditional face-to-face settings and corroborates findings from previous meta-analyses on internet-based cognitive behavioral therapy (iCBT) for depression (Andersson et al., 2014; Karyotaki et al., 2017; Karyotaki et al., 2021).

Mindfulness-based interventions, particularly the Mindful Mood Balance program (Segal et al., 2020), showed efficacy consistent with metacognitive models of depression (Spijkerman et al., 2016). These findings extend our understanding of how mindfulness mechanisms (specifically decentering and metacognitive awareness) can be effectively translated to digital formats, building on established theoretical frameworks of mindfulness-based cognitive therapy (Kuyken et al., 2016).

In the context of anxiety symptoms, only 6 out of the 19 studies provided clear data on this domain. Digital interventions were shown to operate through complementary theoretical mechanisms. Notably, the effectiveness of the Mindful Mood Balance program in reducing anxiety symptoms (Segal et al., 2020) supports transdiagnostic theories of emotional regulation in anxiety and depression (Sauer-Zavala et al., 2020). Additionally, multi-component interventions combining telephone contact and text messages (van den Berg et al., 2015) align with theoretical models emphasizing the importance of multiple channels of support in anxiety management (Craske and Stein, 2016). However, given the limited number of anxiety assessments across studies, further conclusions on the theoretical pathways for the efficacy of digital interventions remain challenging. Regarding relapse prevention, this systematic review identified five studies that specifically addressed relapse or recurrence rates. The interventions that showed promise in reducing relapse rates, such as internet-based cognitive behavioral therapy (iCBT) (Holländare et al., 2013) and mobile cognitive therapy (M-CT) (Klein et al., 2018), share a foundation in cognitive-behavioral principles. These findings support the role of ongoing skill practice and monitoring in maintaining therapeutic gains (Bockting et al., 2018). However, the variability in effect sizes—from highly significant to statistically non-significant – raises questions about other factors that may moderate intervention efficacy.

For example, the effectiveness of the SUMMIT program, which led to an increase in “well weeks” over a 24-month period (Kordy et al., 2016) aligns with theoretical framework conceptualizing depression as a chronic, recurrent condition, which requires continuous support in maintaining recovery (Bockting et al., 2015). However, it also raises questions about the optimal duration and intensity of digital interventions, and how they might be integrated into stepped care approaches. The tendency for more positive results in studies with longer follow-up periods (e.g., 24 months) suggests that the benefits of these interventions may accumulate over time, possibly by helping patients internalize coping strategies or by providing a safety net during vulnerable periods.

Our systematic review revealed that only 4 out of 19 studies specifically assessed quality of life outcomes. Quality of life improvements across psychodynamic, disease management, and mindfulness-based interventions (Zwerenz et al., 2017; Hunkeler et al., 2012; Segal et al., 2020) align with the broader literature suggesting that internet-based interventions can improve quality of life in individuals with depression (Karyotaki et al., 2017; Karyotaki et al., 2021). These results support theoretical frameworks looking at the interconnection between symptom improvement and functional recovery, and more than that, they emphasize the importance of functional recovery as a primary goal in treating major depressive disorder, advocating for clinical assessments that go beyond symptom alleviation to address work, social, and overall life functionality (Lam et al., 2015) Adherence rates (ranging from 50 to 92.3%) are consistent with previous evidence, which found dropout rates in smartphone app trials for depression ranging from 0 to 62.7%, with a pooled dropout rate of 31.3% (Torous et al., 2020a,b). These results highlight the importance of human support in digital interventions (Mohr et al., 2019), with guided interventions showing higher completion rates.

Our synthesis suggests that digital interventions operate through multiple theoretical pathways, including cognitive modification, behavioral activation, emotional regulation, and social support mechanisms. Effectiveness appears to be moderated by factors such as intervention design, level of human support, and individual patient characteristics, suggesting the need for more sophisticated theoretical models of digital intervention implementation and personalization (Torous et al., 2021).

The integration of digital interventions into existing care pathways reveals both design opportunities and implementation complexities. Anaysis of intervention design features showed that shorter programs generally achieved higher adherence rates. For example, Aggestrup et al.’s (2023) 4-week intervention achieved 81.6% completion, contrasting lower adherence rates in longer programs (e.g.: Vicent-Gil et al., 2022). This inverse relationship between duration and adherence aligns with findings from the broader digital health literature (Kelders et al., 2012). The level of human support emerged as crucial - guided interventions, such as the therapist-supported program in Holländare et al. (2013), demonstrated higher completion rates compared to full automated ones. This observation is consistent with meta-analytic findings suggesting that guided internet-based interventions yield larger effect sizes and better adherence than unguided ones (Baumeister et al., 2014). Furthermore, the mode of delivery significantly impacted engagement patterns, with text message-based interventions (Agyapong et al., 2012; O’Reilly et al., 2019), showing high engagement rates (possibly due to low burden and daily life integration), while complex web-based platforms showed variable adherence.

Integration strategies varied significantly across studies. While some interventions, like Kordy et al.'s (2016) SUMMIT program, functioned as stand-alone aftercare interventions, others, such as Agyapong et al.'s (2012) text message intervention, served as adjuncts to standard care. This diversity highlights the need for flexible integration approaches. Studies like Simon et al. (2011) and Hunkeler et al. (2012) demonstrated how digital tools can effectively complement usual care, though successful implementation required substantial provider support and training. The role of healthcare providers proved critical, i.e., studies where guided interventions were consistently incorporated showed better outcomes and high adherence rates, ranging from 76.2 to 96.1% (Hunkeler et al., 2012; Holländare et al., 2013).

Several implementation barriers emerged across studies. Data integration and interoperability presented significant challenges - none of the reviewed studies successfully addressed the integration of intervention data into existing electronic health records. The potential to exacerbate health inequalities through digital divides was evident but inadequately addressed, particularly regarding accessibility for populations with limited technological literacy or access. Regulatory and ethical considerations, especially regarding data privacy and security, were not prominently featured in the studies despite their importance for implementation. Cost-effectiveness data was largely absent, with only limited evidence from Kordy et al. (2016) suggesting potential economic benefits.

Personalization emerged as a promising strategy for enhancing engagement. This was demonstrated in the SUMMIT-PERSON intervention (Kordy et al., 2016), which showed improved adherence compared to equivalent non-personalized version. Engagement-promoting features such as reminders and progress tracking (Simon et al., 2011) showed potential benefits, though their relationship with clinical outcomes requires further investigation. The evidence suggests that successful implementation depends on matching intervention complexity and support levels to both healthcare context and patient characteristics, while ensuring adequate provider training and clear integration protocols (Hornstein et al., 2021).

For clinical practice, our findings provide several actionable recommendations. First, the evidence supports using digital interventions specifically for preventing depressive relapse, with studies like Holländare et al. (2013) and Klein et al. (2018) demonstrating significant reductions in relapse rates over extended follow-up periods. Second, intervention selection should be tailored to patient characteristics – e.g., the Circadian Reinforcement Therapy showed particular benefit for patients with circadian rhythm disturbances (Aggestrup et al., 2023), while cognitive control training benefited those with persistent rumination (Hoorelbeke et al., 2015). Third, to address adherence challenges, clinicians should prioritize interventions incorporating regular support or guidance, exemplified by the therapist-supported internet-based cognitive therapy (Kok et al., 2015). The high satisfaction rates reported in studies like Zwerenz et al. (2017) and Simon et al. (2011) indicate that patients find digital interventions acceptable when properly supported and integrated into their care plan. For successful implementation, clinicians should: assess patient suitability based on technological literacy and access; select interventions matching existing care pathways; ensure adequate monitoring systems; maintain flexibility in delivery based on patient response. This stepped approach to implementation, supported by regular outcome monitoring and adjustment, may optimize the potential benefits while minimizing dropout risks.

The review of digital interventions for depression management and relapse prevention revealed several methodological considerations and limitations that inform future research directions.

The heterogeneity of studies posed a significant challenge, with intervention types ranging from text message-based systems (Agyapong et al., 2012) to complex web-based platforms (Kordy et al., 2016). This variability in intervention design, combined with differences in duration, intensity of human support, and theoretical frameworks, further limited our ability to draw definitive conclusions about optimal implementation strategies.

Control conditions presented another methodological concern, with most studies employing treatment as usual or waitlist controls, rather than active comparators. The predominance of non-active control conditions limits our understanding of specific intervention effects versus non-specific engagement benefits. The lack of attention-placebo controls in digital intervention research particularly affects our ability to distinguish between therapeutic components and general benefits of technological engagement.

Outcome measurement heterogeneity further complicated cross-study comparisons. While depression measures were consistently reported, the diversity of instruments used (e.g., BDI-II, HDRS, PHQ-9) and inconsistent reporting of secondary outcomes, particularly anxiety symptoms and quality of life measures, hindered comprehensive effectiveness evaluation. Furthermore, the duration of follow-up varied widely, from immediate post-intervention assessments to 24-month follow-ups, which limits our interpretation of long-term effects.

Sample sizes and statistical power were concerns in several studies. While some larger trials provided robust evidence (e.g., Segal et al., 2020), many studies had relatively small sample sizes, limiting detection of meaningful effects in secondary outcomes or subgroup analyses.

In addition, the inconsistent reporting of adherence and engagement metrics further complicated interpretation, with many studies failing to provide detailed information on participants’ interaction with interventions.

With respect to generalizability, issues pertaining to selection bias were evident across studies, with recruitment conducted primarily through online platforms or specialized clinics, potentially excluding broader population segments. Individuals with severe depression, suicidal ideation, or comorbid conditions were frequently excluded, which further limits generalizability. Additionally, the rapid pace of technological advancement means that the study interventions might be outdated by publication time, a unique challenge in digital health research.

These methodological considerations point to three priority directions for future research. First, mechanism-focused studies are needed to elucidate how digital interventions affect change, particularly examining specific therapeutic components versus non-specific effects. This review identified several promising mechanisms, which may warrant further investigation, including cognitive processes (Browning et al., 2012; Zwerenz et al., 2017), emotion regulation (Hoorelbeke et al., 2015) and behavioural activation patterns. Second, implementation research should focus on optimizing intervention delivery through personalization algorithms and adaptation to individual patient characteristics. Finally, larger-scale trials with active control conditions and longer follow-up periods are needed to establish comparative efficacy and sustainability of treatment effects.