There is a well-recognized global gap between the need for and provision of services to prevent, identify, and treat mental health conditions [1]. People who experience psychosis are among those with the highest need for support [2]; thus, interest and research into digital treatments for this condition have grown over the last decade. A recent meta-analysis of randomized controlled trials (RCTs) assessing apps designed for symptom reduction in patients with schizophrenia found statistically significant symptom reduction across positive (standard mean difference [SMD]=−0.205, 95% CI −0.388 to −0.022) and negative symptoms (SMD=−0.406, 95% CI −0.791 to −0.020) [3]. However, adherence to smartphone apps for mental health is notoriously low in both trial and real-world conditions. For example, a systematic review of 93 popular publicly available mental health and well-being apps found that only 3.3% (IQR 6.2%) of users were still accessing the app after 30 days [4]. Specifically relating to mental health apps, a meta-analysis of 70 RCTs found that adherence consistently declined with time [5]. One potential contributor to disengagement is insufficient attention to system design and how it affects users’ experience of apps [6]. Increasing researcher and app designer understanding of the role the technology is playing in digital interventions has the potential to shift patterns of user disengagement [7].

One area of system design that can be applied to app development is persuasive systems design (PSD). The PSD is defined as the use of computer-mediated products that aim to influence or change people’s cognition and behaviors [8]. A framework has been developed that transforms persuasive principles into software requirements and system features, namely primary task support, dialogue support, system credibility support, and social support [9]. For example, an app praising a user after they complete a session is a system feature of the PSD principle “dialogue support,” which pertains to the feedback an interactive system provides to its users to help them move toward their goal.

It has been hypothesized that persuasive features will increase alliance, adherence, and engagement, in turn leading to better treatment outcomes. Kelders et al [10] conducted a systematic review of web-based health interventions and found persuasive features accounted for a significant difference in adherence scores. However, more recent research on apps has produced mixed results. A review of 70 apps for health and wellness available to the public found a positive correlation between apps’ ranks of popularity (based on ratings and installation) and the apps’ persuasive features [11]. In contrast, Alqahtani et al [12] found no link between the number of persuasive features in mental health apps and efficacy or engagement. Interestingly, a recent meta-analysis looking at efficacy, engagement, and persuasive design in smartphone apps for depression and anxiety found persuasive features were linked to increased effect sizes; however, an increased number of persuasive features was also linked to rising attrition [13]. Thus, the relationship between persuasive design, efficacy, and engagement is multifaceted and opaque. It may be that for some client groups, persuasive design or over-application of persuasive design has the opposite of the intended impact, causing users to disengage. This requires further investigation.

There is evidence that motivation type [12] and psychological and demographic characteristics [14] impact mobile health users’ proclivity to persuasive design, meaning the characteristics of the user may be relevant to understanding how they relate to an app’s features. People with psychosis may have unique needs. For example, a review on the implementation of digital interventions with people with psychosis and bipolar disorder found that this client group may have cognitive impairments that make engagement with complex interventions challenging [15]. A related difficulty for people with serious mental illness is difficulties in social functioning [16,17]. Furthermore, in human-to-human psychosis interventions, the quality of the therapeutic alliance has been found to be associated with treatment outcome; thus, the quality of this client group’s interaction with the therapeutic agent (app) may be paramount to the intervention’s success [18]. Thus, understanding how apps are being designed for people with psychosis and the links between system design and app adherence could provide valuable information for improving these interventions going forward.

There is a gap in the literature for a review that looks in a granular manner at persuasive features within mobile apps designed for people with psychosis. This review aimed to answer the following questions: (1) what PSD features do apps aimed at people with psychosis have? (2) Which PSD features of smartphone apps aimed at people with psychosis are associated with app adherence and attrition rates? (3) What is the quality of apps aimed at people with psychosis?

This study’s protocol was preregistered on the Open Science Framework [19]. The data associated with this review are also available at the same location on the Open Science Framework. The electronic databases searched were PsycINFO, PubMed, and Google Scholar. Due to the fast-evolving nature of technology, database searches were limited from the year 2013 onward and searched in November 2023 and March 2025, respectively. Query strings were (smartphone* or “mobile phone” or “cell phone”) AND (“app” or “apps” or “application” or “applications”) AND (“schizophrenia” or “schizo” or “psychosis” or “psychotic”). To identify articles not found in the original searches, we reviewed the reference lists of included papers for relevant apps. The authors also used prior knowledge of apps to identify additional inclusions and linked these with their associated papers. Where papers could not be accessed, study authors were contacted.

The “unit of interest” in this review was the smartphone apps themselves, as we were aware that multiple papers may be published relating to various aspects of the same app (eg, pilot studies, clinical trials, and qualitative papers). We linked together multiple papers where they related to the same app and included those most relevant for gathering the information for this paper.

Inclusion criteria for eligible apps were as follows:

Exclusion criteria were as follows:

Covidence was used for all stages of review, including screening and data extraction. After duplicate records were excluded, 2 reviewers (TT, DK, SD, and JP) independently screened each paper at both title and abstract and full-text review. All screeners completed training, and a consensus check on a small selection of papers was done to ensure consistency between screeners. The initial agreement between screeners was 81% at title and abstract review and 70% at full-text review. Where there were conflicting decisions between screeners, consensus was reached by discussion and with input from the senior author (PJ) when needed.

Relevant data from included studies were extracted in duplicate independently by 2 reviewers (TT or JP and DK or SD) using a predetermined data collection tool. Oinas-Kukkonen and Harjumaa [9], which lays out descriptions of persuasive features and examples of how the features can be defined, and their implementation was used as a coding guide for determining the persuasive features. Prior to extraction, reviewers piloted the data extraction tool and discussed discrepancies. During extraction, regular meetings allowed discussion of the coding guide and resolution of discrepancies. The review team extracted information on the apps, including the intervention name, a brief description of the app, treatment target, whether the app was automated (used without a human intervention component alongside) or blended (used with an accompanying human therapist), intended frequency of intervention, and intended duration of intervention. Blended interventions with human support may inherently include persuasive elements outside the app interface, which was out of scope for this review.

From the most recent empirical paper on the app, information describing the study was extracted (study design, location, sample size, participant mean age and diagnoses, attrition during trial, and available adherence data). Missing data were reported in the write-up. “Interrater reliability was assessed across all 22 included apps and 28 PSD features (616 coding decisions). Cohen Kappa was κ=0.95 (98.4% observed agreement).”

The reviewers had planned to assess the quality of the apps using the Mobile Application Rating Scale [20], a quality assessment tool for health apps with published psychometrics [21]. The Apple and Android app stores were searched, and the corresponding authors were contacted to request access to the apps. However, only 2 of the 22 included apps could be accessed; thus, assessing app quality could not be done meaningfully, and this was not completed.

The search identified 566 studies, including 234 duplicates. After title, abstract, and full-text screening, 30 [22-51] eligible papers reporting on 22 apps were included. The screening process is summarized in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses; Checklist 1) diagram (Figure 1). App names, treatment targets, and whether the app is automated or blended are detailed (Table 1). Further information on the apps, associated outcome study characteristics, and attrition and adherence rates is provided (Multimedia Appendix 1) and a matrix of persuasive features of each app (Multimedia Appendix 2). Across 22 apps, 21 categories of persuasive features were identified. Primary task support and dialogue support were the most common, with social support and system credibility support coded considerably less often.

Twenty-one categories of persuasive features were found across the 22 apps [22-51]. The most coded areas of persuasive features were primary task support and dialogue support, with social support and system credibility support being coded significantly less frequently. The maximum number of PSD elements in an app was 11, while the least was 2 (Multimedia Appendix 2).

The number of persuasive features and attrition percentage for each app showed no evidence of an association (Pearson r=0.07, 95% CI −0.4 to 0.5). However, this finding assumes equal weight for each PSD feature, which is likely unjustifiable. Certain features may be more strongly associated with attrition than others, if any relationship exists, but these data are insufficient for this level of analysis. Inconsistent adherence reporting across apps prevented comparisons regarding adherence-persuasive feature links.

App stores were searched, and corresponding authors were contacted to access all apps. However, only 2 apps were accessible, making a quality assessment impossible, as it would have required opening and using the apps to evaluate each criteria point.

This review assessed PSD features in mobile apps for people with psychosis, aiming to describe the features, links to adherence and attrition, and app quality. The categories of primary task and dialogue support were well-represented, while social support and system credibility support were underused, potentially impacting app efficacy and engagement, as both social connectivity (social support) and a user having ways to judge a system’s credibility (system credibility support) could impact motivation with this client group. No link was found between persuasive features and attrition; the relationship with adherence and app quality couldn’t be assessed.

No previous study has systematically documented the PSD features present in apps for psychosis. The finding that personalization, self-monitoring, and reminders were among the most common strategies is consistent with previous findings on physical activity, mental health, and health apps [11,12,53]. However, the existence of these features does not guarantee a positive impact on users; how these features are applied is crucial to their utility. For example, reminders were a prominent persuasive feature in the psychosis apps reviewed; however, most apps did not give users a choice in notification frequency. One of the postulates of persuasive systems is that the system should aim to be unobtrusive, and inappropriate or ill-timed reminders may have a reverse of the desired impact [54]. Persuasion depends on symbolic strategies to trigger emotions [55], and a blunt, unskilled approach such as sending countless reminders to use the app per day can provoke a negative emotional response, which is counterproductive. Similarly, self-monitoring is considered a fundamental feature of several evidence-based psychological therapies, including cognitive behavior therapy [56], dialectical behavior therapy [57], mindfulness [58], emotion-focused therapy [59], and acceptance and commitment therapy [60]. Thus, it is logical that psychosis apps would utilize self-monitoring to help users track their feelings, thoughts, and behaviors. However, the apps reviewed in this study limited data gathering to manual input, prompted by several daily notifications. Manual entry of personal data can be time-consuming and tedious [61] and may be a barrier to people with serious mental health difficulties. Thus, in the context of psychosis, it is unclear if self-monitoring promoted engagement in the manner intended or had a reverse to desired impact due to the frequency and high cost to the user of multiple manual inputs per day.

Research indicates that psychological and demographic variables influence susceptibility to persuasion [14]; thus, in planning these interventions, these variables should be considered. For example, in this review, reduction (simplifying tasks) and suggestion (suggesting actions to users) were among the top strategies used. However, in a review of mental health apps, these 2 strategies were used in fewer than 5 of 103 apps [12]. Thus, these strategies, which facilitate ease of app use, may be more widely implemented and necessary in a psychosis sample due to cognitive impairment and disability associated with this client group [15-17].

This study also found that apps relied heavily on the primary task support and dialogue support categories of persuasive design, with social support and system credibility support underrepresented. This means that attempts to be persuasive are unbalanced; there is emphasis on helping users to complete tasks and the interactiveness of the app, while techniques to make the app more trustworthy and credible to users and methods to leverage the impact of social forces to support behavior change are ignored. Social support strategies have been found to be more widely implemented in physical than in mental health apps [62]. However, social support is an important strategy for users who experience mental health issues because it is common to feel isolated or stigmatized, and digital interventions could benefit from intervening at the level of the social rather than limiting treatment targets to internal drivers for poor mental health. Lack of system credibility support is common across multiple types of apps [12,62]. A postulate of persuasive systems is that they should always be open to the user and reveal designer bias behind the system both to protect fairness and increase persuasiveness. We argue that system credibility support is fundamental in this time, where technology is developing at a rapid rate and users cannot be expected to understand a system’s functions without guidance. Users need to be assured of not only the effectiveness, reliability, and evidence base of the app’s content, but also that their data will be protected (privacy). Furthermore, without this in place, it is difficult to advance other features such as passive monitoring if users are not fully assured of the app’s credibility.

Our findings suggest that the number of persuasive features used in psychosis apps is not linked to attrition rates. However, this is not necessarily an indicator that the persuasive features were not assisting the user to achieve their purpose with the app. Some authors propose that disengagement is not a barrier to efficacy if a user has achieved their goals with the digital intervention at the point of disengagement [63]. For example, a recent meta-analysis looking at efficacy, engagement, and persuasive design in smartphone applications for depression and anxiety found increased persuasive features were linked to increased effect sizes and increased attrition rates [13]. However, there may be other factors explaining the disconnect between increased persuasive features and engagement in this review; for example, there is a lack of evidence on which persuasive features are the most effective at increasing adherence and reducing dropout. Therefore, it is likely that totaling persuasive features is insufficient. Using one appropriate feature is preferable to multiple inappropriate features, too many features from one persuasive category, or a combination of appropriate and inappropriate features [64]. Furthermore, although PSD is commonly used to inform app development, the model does not contain a guideline on operationalizing the features [9]. Consequently, the same persuasive feature can be shaped differently in different interventions. We found some apps with highly creative implementation of PSD, while others executed persuasive features in a minimal manner. This mirrors the results of Sporrel et al [65], who examined 29 physical activity apps and found heterogeneity in how the features were executed, which appeared to influence the effectiveness of the intervention. These diverse implementations mean that interventions with the same persuasive feature might evoke different user responses, rendering it difficult to draw a conclusion about the impact of a specific persuasive feature at the theoretical level. Furthermore, technological advancements mean that some apps used features that did not fit within the PSD framework [53]. Thus, the lack of clarity on how many persuasive features are optimal, the lack of guidance on the operationalization of features, and the lack of acknowledgment of emerging features highlight the need to extend the PSD model. Furthermore, extension of the PSD model could include standardized checklists or algorithms for the implementation of persuasive features as guidance to app developers and reviewers.

This study’s findings provide possible guidance to app developers. The finding that, although reminders are a prominent feature, little choice of their frequency was offered indicates that app engagement may benefit from offering users choice when providing notifications. This review found that self-monitoring was a common feature; however, given the functional impairments this patient group frequently faces, passive monitoring may lower the burden on the user. To implement these more advanced features, user trust is crucial. Thus, given this review’s finding that system credibility features were not common, developers could benefit from including trust-building features such as highlighting the team, evidence, endorsements, backend and data storage of the app system, and real-world institutions behind the app product. This client group faces social isolation, and this review found there was little focus on social support features; thus, developers can consider integrating peer-to-peer interactions or content that gives an indirect indication of peers and others using the app (eg, leaderboards) into apps for psychosis.

This review has several methodological strengths, including a preregistered protocol, a systematic search strategy, and 2 independent raters throughout study selection and data extraction. However, several limitations should be noted.

The most significant constraint on interpretation is that only 2 of the 22 apps could be tested directly, rather than through reading their described features in their associated research papers [24,25]. This means the review may not fully capture the persuasive features present in each app, particularly those that are less tangible or harder to articulate—such as system credibility support—which researchers may be less likely to describe explicitly. The reported prevalence of feature categories should therefore be interpreted with this in mind, as it may reflect limitations in how apps are described in the literature as much as the features themselves.

Identifying persuasive features is also an inherently interpretive process; review teams may conceptualize and operationalize features differently. Relatedly, this review’s feature-counting approach assumes equivalent value across different persuasive features, which may not hold in practice; however, a more nuanced weighting of features was beyond the scope of this analysis.

The absence of a formal risk of bias assessment is a further limitation. Although a quality appraisal of the apps was planned, it was not feasible given their inaccessibility. Additionally, attrition and adherence data were either unreported or reported inconsistently across studies, preventing any meaningful analysis of the relationship between persuasive features and these outcomes. Together, these constraints limit conclusions about feature effectiveness, and further empirical work is needed.

Finally, while non-English articles were not explicitly excluded, studies would only have been identified if their abstracts were available in English, meaning relevant research published in other languages may have been missed.

Our findings indicate that in psychosis apps there is an uneven use of persuasive features, with most apps relying on primary task support and dialogue support features, while social support and system credibility support remain underrepresented. This may be impacting the efficacy and engagement of these interventions. Psychosis apps may benefit from incorporating more features that leverage the persuasive impact of having users interact (social support). There is a human instinct to respond to social forces, and thus incorporating these domains into psychosis apps may allow developers to engage nonengagers. Incorporating features that emphasize system credibility and trustworthiness to the user is also crucial. As technology advances, there is an increasing need for designer openness about their positioning, bias, and how data are stored and used, and further technological advances can only be widely applied if users have confidence in the systems that gather their data. Future research could explore expanding the PSD model, as this review indicates it no longer encompasses all features apps are using, and although the underlying principles may remain sound, the way the model quantifies and describes features is becoming outdated with rapid technological advancements. Furthermore, there is a lack of clarity on how many persuasive features are optimal, how to operationalize features, and how persuasive categories work together. It may be that certain features are more effective in this context than others, and emphasis should not be on increasing numbers of persuasive features but on those most likely to influence adherence, attrition, and efficacy. In future research, a positive deviance approach could be used to examine the features of the most engaging and effective apps, while correlational or experimental research designs could explore the impact of different features on adherence and attrition. Qualitative interviews with users may shed light on what features are experienced as most impactful by users and contribute to understanding their emotional valence and possible impacts on effectiveness.