Mental health conditions remain among the most prevalent global health concerns, with current trends indicating that nearly half of the world’s population will have experienced at least one mental disorder by the age of 75 years [1]. Despite the high prevalence, fewer than half of those in need seek professional support [2]. For those who do, access is often limited by long waiting times, high costs, and a shortage of trained professionals [3-7]. Even when treatment is received, many individuals do not experience a meaningful improvement, despite evidence that many psychotherapeutic treatments are generally effective [8-11]. These limitations point to the need for interventions that are more accessible, personalized, efficient, and scalable. Digital mental health interventions (DMHIs) show great promise in overcoming common barriers to care and reducing the treatment gap [12]. For example, they can lower delivery costs, increase accessibility in remote regions, and enable personalization, a potential amplified by rapid advances in AI. DMHIs take on various forms; they can complement the clinicians’ work or function as fully autonomous, unguided formats, including web and smartphone-based modalities [13]. Some smartphone apps are now formally recognized as medical tools that can be prescribed and reimbursed by health care providers in several countries such as the Netherlands, Germany, and the United Kingdom [14,15]. Importantly, the potential of digital tools extends beyond treatment. As unguided formats are usually low-intensity, they are well suited for prevention and health promotion, for example, by offering interventions that improve coping skills and resilience [16]. In this context, prevention may involve targeting risk factors or strengthening protective factors and mental well-being. Despite their potential, DMHIs face challenges similar to face-to-face interventions. While they are generally effective, treatment response varies across individuals, and improving their overall (cost) effectiveness over time remains difficult. To enhance effectiveness, it is important to precisely understand how these interventions achieve their effects. Randomized controlled trials (RCTs), considered the gold standard for evaluating interventions, assess the global effectiveness of an intervention against a control condition, but they do not allow researchers to isolate which components are responsible for change, how these elements interact, or whether they operate through the hypothesized mechanisms [17,18]. This way, it is difficult not only to improve interventions over time but also to identify ineffective or even potentially harmful components, making them less effective but also less efficient and economical. Finally, the interventions that are tested within an RCT are often not scalable because they are too complex, resource-intensive, and insufficiently tailored to local contexts and policies [19,20]. Such interventions are often altered postevaluation. However, because the exact mechanisms through which interventions work are not certain, decisions about how to modify intervention packages to fit real-world demands rely on theoretical assumptions and practical considerations rather than empirical data. This risks perturbing active ingredients and the underlying working mechanisms. At the same time, retaining components that contribute to effectiveness may come at the cost of greater intervention length, participant burden, or resource use, potentially affecting adherence and scalability. These trade-offs are difficult to evaluate post hoc and should ideally be considered explicitly prior to evaluation. The Multiphase Optimization Strategy (MOST) [21] is a framework designed to tackle the limitations of the traditional intervention development and evaluation. Inspired by engineering and industry, it aims to prevent suboptimal allocation of resources during optimization by ensuring that intervention components are tested efficiently and purposefully, rather than through overly resource-intensive approaches. The MOST is guided by 2 core principles: continual optimization, which emphasizes iterative intervention refinement over time, and resource management, which prioritizes the efficient use of available resources across both intervention development and testing. By experimentally assessing individual and combined intervention components and applying explicit implementation constraints, the MOST balances effectiveness with the need for efficiency, economy, and scalability. In doing so, it provides a systematic and transparent approach to intervention optimization, illustrating that effectiveness, efficiency, affordability, and scalability are interrelated and can be targeted simultaneously. The MOST framework can be applied both to the development of new interventions and to the optimization of existing ones. It is structured into 3 phases: preparation, optimization, and evaluation [22].

The preparation phase lays the groundwork for optimization. The conceptual model is considered the “engine” that drives the intervention [23], that is, its development, its translation into a trial design, and the interpretation of the trial results. A strong model clearly specifies the relations among the key constructs, depicting the causal pathways through which the intervention components are expected to influence the intervention target (ie, the “primary” or “ultimate” outcome of interest). These causal pathways must include mechanisms of change. The mechanisms of change must be operationalized (and measured) as proximal outcomes expected to affect the “ultimate” outcome of interest, which often occurs later, thus sometimes also called a “distal” outcome. For example, in an intervention targeting insomnia (distal outcome), components, such as physical activity, light control, and journaling (intervention components), may reduce insomnia through lessening stress, improving the circadian rhythm, and reducing worrying (mechanisms of change). Ideally, each intervention component should target a specific mechanism of change to strengthen the interpretation of the optimization trial results and thereby facilitate the decision-making on which combination of components to retain. The intervention can be developed using the literature, but it is also beneficial to incorporate lived experience through co-design and input from end users and stakeholders [24]. Additionally, it can be beneficial to consider effect modification, that is, to assess person-specific variables which may impact the effectiveness of components (such as presence of comorbidity or demographic characteristics). This can help identify which components work best for whom. Next, an optimization criterion is defined. The optimization criterion specifies a (set of) key implementation constraint(s), within which the desired outcome (such as an effective intervention) should be achieved. This involves setting limits on resources so that effectiveness can be balanced against the efficiency, economy, and scalability. An example would be a behavioral intervention for lifestyle change that produces clinically significant change (outcome) while not exceeding 10 sessions (constraint). In a hypothetical scenario of unlimited resources, an “all active components criterion” can be used, allowing the inclusion of all components that demonstrate an effect. Finally, before proceeding to the optimization phase, the candidate components are pilot tested for feasibility, usability, and acceptability to ensure that each element can be delivered as intended.

For the optimization phase, an experimental method is used to determine which (candidate) components [1] produce meaningful effects, either independently or in combination. The choice of experimental design depends on the intervention design. For “fixed interventions,” where all participants in the same condition receive the same content, a factorial design is used. For “adaptive interventions,” where content is personalized according to participant input, different approaches are suitable such as a sequential multiple assignment randomized trial (SMART) or a microrandomized trial. Based on the results of the optimization trial, it is decided which (combination of) components to retain in the optimized intervention package. This decision-making process follows three steps: (1) identifying an effect size that a component should demonstrate alone or in combination to be considered important (“effect criteria”), (2) splitting components into those that qualify for further consideration (“screened-in set”) and those who do not (“screened-out set”) according to the effect criteria, and (3) selecting the best combination of screened-in components that meet the optimization criterion. If the trial does not produce interpretable effects, meaning that there is no difference in components on their effectiveness or other type of resource-intensiveness, a return to the preparation phase is necessary. Accordingly, multiple optimization rounds may be needed, which includes returning to the preparation phase for revision. However, even successfully optimized interventions can be further optimized in the future, in line with MOST’s continual optimization principle.

Once an optimized intervention has been achieved, it can be evaluated in a conventional RCT against a control condition. If effectiveness is demonstrated, the intervention is ready for large-scale implementation, having already been optimized with respect to key resource constraints in the preceding stage.

Digital (complex) interventions are particularly well suited to the MOST framework due to their capacity to precisely manipulate therapeutic components and the extensive opportunities for personalization. However, the extent and manner in which theoretical frameworks are applied differ across studies, highlighting the need for a clearer overview of current practice [25,26]. Therefore, the aim of this scoping review is to examine how the MOST has been applied to the development, optimization, and evaluation of DMHIs and to provide recommendations for future application. A scoping review methodology was selected instead of a systematic review because the scope, characteristics, and extent of the evidence were not yet well established and expected to be heterogeneous, making a broad mapping approach more appropriate for a broad research question than a narrowly focused approach [27].

This review was conducted following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Scoping Reviews) guidelines. A completed checklist indicating where every item was reported can be found in Checklist 1. The review was not preregistered because the approach to organizing and synthesizing the evidence was refined iteratively as the scope and heterogeneity of the literature became clear. A systematic search was carried out in July 2024 (and updated in August 2026) in the databases Embase, CINAHL, Web of Science, PsycInfo, and PubMed. In all of them, abstracts, titles, and keywords were initially screened for the search terms: “multi-phase optimization strategy” OR “multiphase optimization strategy” OR “multi-phase optimization strategy” OR “most framework” (Multimedia Appendix 1). These search terms were derived from the systematic review of Landoll et al [28] of the preparation phase of the MOST framework. The current review extends this scope by additionally examining the optimization and evaluation phases, while narrowing the focus to DMHIs, enabling a field-specific synthesis of MOST implementation. The eligibility criteria consisted of (1) the paper used the MOST to develop or optimize an intervention, or the paper was part of a research project that did so; (2) the intervention targeted a mental health problem; and (3) the intervention was partially or fully delivered digitally. Interventions targeting children (aged <12 years) were excluded as well as theoretical examinations of the MOST framework, review papers, hypothetical applications to a new field, and simulated experiments. Preprints were eligible for inclusion.

After removing duplicates, 2 reviewers (AK, DEA) first independently screened the title and abstract of all search results to determine if they met the eligibility criteria. Following that, 2 reviewers (MK, CRvG) conducted full-text screening to further determine inclusion. Any conflicts or disagreements were resolved by jointly discussing the relevant papers and analyzing their adjustment to the eligibility criteria (MK, DEA, CRvG, AK). Additionally, once projects were selected for inclusion, supplementary forward and backward searches were conducted (MK) to identify papers within the projects that might have been missed in the initial search due to not referencing the MOST framework. Such cases typically occurred in publications preceding or following the optimization phase, such as intervention development and pilot testing or RCTs. This step was aimed to capture information relevant to the review, such as the theoretical background of the intervention, an existing pilot trial, or outcomes from an RCT that may have occurred either prior to a MOST preparation and optimization phase or following a MOST optimization phase.

Once the set of included papers was established, data were extracted using customized Excel spreadsheets (Multimedia Appendix 2). These spreadsheets collected extensive data on each paper. First, general information about the study was extracted (authors, publication year, location, type of article [empirical, protocol paper], part of which research project, etc), followed by general information about the MOST (what phase[s] did the paper refer to) and then detailed information about each of the MOST phases, including the activities recommended per phase. The unit of analysis was the underlying research project rather than the individual papers. Accordingly, multiple papers within a single research project were considered to represent one study, for instance, separate publications of the preparation and optimization phases, or a protocol paper describing an optimization trial and a subsequent paper reporting the trial results.

Additionally, a risk of bias assessment was conducted for the completed (MOST) trials. Two reviewers independently conducted the assessments based on the 5 domains of Cochrane’s RoB 2 tool [29], with disagreements resolved through discussion. The outcomes for each of the five criteria are provided in Multimedia Appendix 3, and detailed ratings are available on request.

Thirty-seven papers met the inclusion criteria, comprising 11 protocol papers and 26 empirical papers. These papers corresponded to 18 distinct research projects (see Figure 1 for the inclusion process).

Half of the projects were conducted in the United States (n=9), followed by the United Kingdom (n=3), Japan (n=2) and China (n=2), Canada (n=1), and Singapore (n=1). All papers were published between 2013 and 2026; the first paper, which explicitly mentioned MOST, was published in 2016 [30]. The most common primary outcome was depressive symptoms measured with the 9-item Patient Health Questionnaire. Table 1 shows an overview of the included studies and the associated papers, as well as the investigated interventions’ content and delivery. Seven interventions were entirely self-guided, one was adaptive, and one had a component with an adaptive element.

One study (Drink Less) tested the previously MOST-optimized intervention as a package within a 2-arm parallel group, double-blind, RCT against usual digital care, thus against an effective comparator [65]. At 6 months, the prespecified primary analysis did not demonstrate a statistically significant between-group difference in weekly alcohol consumption; a sensitivity analysis revealed a significant effect on the outcome of interest in favor of Drink Less, concretely a 2.00-unit greater weekly reduction of self-reported weekly alcohol consumption.