We searched the databases recommended by Cochrane [37], including Medline (PubMed), ACM Digital library, PsycINFO, EMBASE, CINAHL, and Cochrane, to identify studies. In addition, we hand-searched the reference list of the selected publications to retrieve additional relevant publications. The search string included Adaptive, OR Flexible, OR Tailored, AND Internet, AND Interventions, AND Mental Health (see Multimedia Appendix 1 for detailed search string). Each term included medical subject headings, and the search was done on full-text papers. The search was limited to all papers published in English from January 2000 to January 2020. The database searches and subsequent review were performed by the same two authors (SKM and JDW) independently in a double-blind process.

We included studies in which the articles met the following inclusion criteria: (1) discussed an intervention delivered through the internet (web- or mobile-based), (2) attempted to provide adaptive (dynamic, tailored, flexible) interventions by using adaptive strategies, (3) targeted a mental health disorder defined by the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) [38], and (4) was published between January 2000 and January 2020. No data restrictions were imposed. The following exclusion criteria were used: (1) not written in the English language, (2) not a full-text paper or published in the form of a short paper, extended abstract, abstract or poster, (3) designed as nonempirical findings such as opinion papers, reviews, editorials, letters, agendas, or study protocols, (4) dealt with adaptive technology in any domain other than mental health, or (5) was not about adaptive technology.

The selection of studies took place in three phases based on the review of the title, keywords, abstract, and full text. Title and abstract screening were carried out blinded for author, journal, and date of publication. Any doubtful papers were included in the next phase, and disagreement was resolved through discussion. After identifying 3341 relevant papers in the initial database search, 372 duplicate papers were removed, and 2969 unique papers remained. In the screening step, the resulting list of 2969 papers was reviewed independently by the same two authors according to inclusion and exclusion criteria. By reviewing the title, abstract, and keywords, 105 eligible papers were retrieved. Two main reasons for the substantial exclusions were (1) the search engine returned the results containing any of the search terms, although they were logically connected, and (2) most of the papers were related to mental health without any reference to IDPT. Full texts were evaluated to determine the eligibility of the remaining papers. The full texts of the 105 eligible papers were assessed independently by the same authors. Any discrepancies between the authors regarding the selection of the papers were resolved through discussion. In total, 74 papers were excluded in this round, and the selection process led to the inclusion of 31 papers, as illustrated in Figure 1. The most common reason for exclusion in this phase was that the publication did not discuss an intervention delivered via the internet. Other publications were excluded because they focused on other types of health care interventions without clear information about IA, user adherence, or adaptive strategies.

Data from the included studies were extracted, verified, and tabulated for review by the authors. From the selected studies, we chose to obtain the main adaptive elements, adaptive strategies used, adaptive dimension, and actor involved in adaptation. Multimedia Appendix 2 provides a detailed summary of the analysis of the 31 articles in the review. All of the articles in the analysis are listed in the references [11-14,24,32,39-63]. We evaluated all of the relevant studies based on the adaptive IDPT model previously described [28]. As mentioned in the study [28], we extracted the core components of the adaptive reference model, including adaptive elements, adaptive dimensions, IA, and adaptive strategies. The rest of our results are based on these core components.

For purposes of transparency and reproducibility of our study, we have published the resulting data, code, and procedures on GitHub [64]. The GitHub repository includes raw articles extracted from database searches, keyword formulation documents, preprocessed article lists, and a literate programming script used for data preprocessing, analysis, and visualization.

A significant number of the included studies addressed depression (n=11) and anxiety disorder (n=7), followed by general mental health issues (n=8), such as well-being, mindfulness, and goal achievement. Furthermore, some studies reported the use of adaptiveness in other areas such as insomnia (n=2), social psychology (n=1), attention deficit hyperactivity disorder (n=2), posttraumatic stress disorder (n=2), suicidality (n=2), and substance misuse (n=1). The full list of types of mental health problems addressed in the relevant studies is presented in Table 1.

Based on our findings, the communication media used to administer internet-facilitated interventions to patients can be classified into three categories: web apps, mobile apps, and computer games. A significant number of the included studies were based on web apps [11,14,24,39,43,49,50,52,54,56,63], followed by mobile apps [13,32,41,49,50,55,60,61] and a game-based intervention [53]. There was only one paper that applied both web and mobile technologies for internet-based intervention. However, a lot of studies did not report the mode of delivery.

IA is concerned with the art and science of organizing and labelling components of web apps, intranets, software, and online communities to enhance their usability and accessibility. IA plays a vital role in web app development, and a good architecture can improve the ability of employees and customers to find information and decrease the app’s maintenance cost [28]. Finding the type of IA used in IDPT systems and their relevancy to treatment outcomes is one of the research questions of this study. However, a significant number of studies (20/31, 65%) did not report the type of IA they used in their IDPT system. Based on the IA of the intervention reported in the 31 reviewed articles, 4 (13%) studies reported the use of tunnel-based IA, and 3 (10%) studies reported the use of matrix IA and hierarchical IA. The full list of types of IA used in the relevant studies is presented in Table 2.

The analysis of Table 2 answers our first research question. Most of these IAs fall into four categories: (1) tunnel-based design, (2) matrix design, (3) hierarchical design, and (4) hybrid design. A previous study [65] showed that 90% of the available IDPT systems used a tunnel-based design, where users navigate sequentially to search for information. A tunnel-based design is analogous to watching TV series, reading textbooks, attending academic classes, or attending multiple clinical sessions. An argument for tunnel-based design is that the experience is less likely to overwhelm users with information and options [66]. The tunnel-based design is probably also the default IA design alternative in many projects and is the easiest to implement.

Adaptive elements are the main components that are personalized for the user. As reported in a previous study [28], the main adaptive elements can be intervention content, design, assessment tests, IA, content presentation, content complexity, content recommendation, user interface (such as navigation system, search engines), feedback, notifications/reminders/alerts, behavioral activities, exercises, and reporting/dashboards. We report the full list of adaptive elements found in the relevant studies in Table 3.

Numerous studies (9/31, 29%) reported adapting the content of the intervention. However, most of these studies did not explicitly report the type of content, level of complexity, or modality (audio, video, presentation, pictures, assignments, activities, and assessments). Knowledge of the modalities of the content and their associated complexity provides insight into how interventions could be adapted and personalized for patients.

Another notable observation is that several studies (11/31, 35%) used feedbacks as adaptive elements. Numerous studies described the process of adaptive feedback in different forms, including sending personalized motivational messages [43], tailored messages by therapists [11,13,54], and providing general support [14]. In contrast, a few studies aimed to adapt reminders and alerts by sending an email or SMS text message or making a phone call [24,48,50]. Only 2 studies targeted the adaptation of exercises [41,47] and 1 study targeted the adaptation of behavioral activities [60]. We identified a total of 7 papers (7/31, 23%) that adapted assessment tests or psychometric assessment tests [39,42,45,46,55-57].

Table 3 presents the list of the primary adaptive elements in the adaptive IDPT systems. The list includes (1) intervention content, (2) content presentation, (3) feedback messages, (4) assessment tests, (5) behavioral activities (sleep pattern), (6) reminder messages (email, SMS text messages, phone calls), (7) exercises, and (8) reporting (dashboard for the patients and the therapists). The central concept of adaptiveness is to create different levels of these adaptive elements and provide these elements based on a personalized profile. For example, if a person watches videos more than they listen to audio, read text, or view slides, then based on the principle of adaptiveness, it makes sense to present upcoming interventions in a video format.

The way an adaptive system changes its behaviors depends on a multitude of factors: (1) users’ data and preferences, (2) goals of the intervention, (3) measures, (4) adaptation actors, and (5) adaptation strategies. We refer to these aspects as the dimensions of the adaptive IDPT system [28]. Table 4 presents a list of adaptive dimensions extracted from the included studies.

The relevant studies were mainly grouped into two clusters based on the choice of adaptive dimensions: user preferences (13/31, 42%) or outcome measures (14/31, 45%). Only 1 study used a goal-based adaptive dimension. Among the studies using user preferences, some studies [32,42-46,62] used user context, while some studies used user location [32,62] to adapt interventions. The studies based on outcome measures used either psychometric tests or user behavior analysis based on interaction data to measure the performance outcome.

The adaptive strategy indicates the techniques used to tailor the intervention. In a recent study [28], four significant clusters of adaptive approaches were identified, namely rule-based adaptation, predictive algorithm-based (such as machine learning) adaptation, goal-driven adaptation, and adaptation through a feedback loop. Similar to the study [28], we identified the following adaptive strategies, presented in Table 5, in the reviewed studies.

The list of strategies includes rule-based strategies, predictive algorithm-based strategies, and recommendation-based strategies. As expected, a significant number of studies (20/31, 65%) used some form of rule-based adaptation mechanism. For example, some studies [42,45,46] used item-response theory [8] to tailor psychometric tests. The primary motivation toward adapting psychometric tests is to extract essential information from patients without asking them too many questions. The study by Van Gemert-Pijnen et al [56] applied user behavior analysis by inspecting web-log data and combined it with a rule-based engine to adapt the intervention. The same study concluded that pattern recognition can be a useful tool to tailor interventions based on usage patterns from earlier lessons [56]. With the hype of data science, several studies [12,40,44,49,50,58,60-62] attempted to use some form of predictive algorithm to adapt the intervention. While some studies did not report the outcome of the overall study [40,49], most of them reported that the use of predictive algorithms had a positive effect on the adaptation of interventions [12,44,50,58,60-62]. However, these studies concluded that further research is required to study the effectiveness of performance outcomes.

In general, this systematic review shows that tailoring interventions according to patients’ needs and preferences has a positive effect on user adherence and hence treatment outcomes. Several studies reported that the personalization of interventions [13,32,39,41,43,47,50-52,54,56,59] and assessments [42,45,46] increased user adherence. Similarly, a study by Coyle et al [53] reported that a personalized system provided a higher degree of user satisfaction. However, some studies reported a lack of noticeable improvement in treatment outcomes. For instance, a study by Batterham et al [48] reported that there was no significant difference between tailored and static versions in the effectiveness of treatment or adherence. However, the same study reported that participants in the tailored conditions were more satisfied than those in the control conditions [48]. Lillevoll et al [24] made a similar conclusion, reporting that tailoring of feedback and dispatching weekly email reminders did not improve the intervention outcomes or user adherence. Given the scenario with two different clusters of results, further research and clinical trials are required to comprehend how user adherence and personalization of interventions are correlated.

A total of 3341 studies were initially identified based on the inclusion criteria. Following a review of the title, abstract, and full text, 31 studies that fulfilled the inclusion criteria remained, most of which attempted to tailor interventions for mental illnesses. Approximately 68% (21/31) of the studies had a first author with a health care background. The most common adaptive elements were feedback messages to patients from therapists (11/31, 35%) and intervention content (9/31, 29%). However, how these elements contribute to the efficacy of IDPT in mental illness was not reported. The most common IA used was tunnel-based IA (4/31, 13%), while many studies (20/31, 65%) did not report the IA used. The rule-based technique was the most common adaptive strategy used in these studies (20/31, 65%). All the studies were broadly grouped into two adaptive dimensions based on user preferences or using performance measures such as psychometric tests.

Our findings show that web apps, mobile apps, and computer games are the primary platforms used to facilitate interventions. Apart from these, other communication media include robotics, virtual reality (VR) [67], augmented reality (AR), conversational agents, or chatbots [68]. A key finding in the literature study was that most of the IDPTs were made available on mobile apps [69] and web-based apps. This is as expected, as these are the most prevalent platforms used for personal computing. Smartphones contain a plethora of sensors and other data sources that inform aspects of users’ well-being, context, activities, behaviors, and intentions. However, only a few attempts have been made to provide IDPT using conversational agents [68], or VR or AR apps. As depicted in Figure 2, higher placement in the chain indicates higher computational complexity but lower prevalence. Conversely, lower placement in the chain indicates lower computational complexity but higher prevalence. The most obvious explanation for the selection of web and mobile apps for intervention platforms is their prevalence. It makes the most sense to develop for platforms that are being used the most.

Although a significant number of the studies failed to report which IA was used in their IDPT system, IA is still present in all software systems. Understanding the IA of a system helps a user to store, find, and interpret information readily, as IA is the design principle that is applied to making information discoverable and understandable. Finding the underlying IA of the IDPT system can help in making systems accessible and discoverable for end users, and knowledge about information design, structure, organization, and labelling can facilitate the development and evaluation phase. As explained in a previous study [70], IA consists of three major components: user, content, and context. Hence, IA helps to discover content for users based on their context. Assisting users with the correct piece of information they are looking for can increase user adherence by reducing bounce rate [71] and hence improve treatment outcomes. While our findings show most of the current adaptive IDPT systems are tunnel-based and don’t promote personalization, we promote the use of hierarchical IA or hybrid IA in order to present useful information according to patients’ needs and argue that the latter designs offer richer opportunities for adaptive IDPT systems.

The primary context for building adaptive IDPT systems is to assist patients suffering from mental health disorders to learn about and recover from their illnesses. An IDPT system provides this information by using different media and elements such as text, video, audio, pictures, presentations, feedback, reminders, reports, and others. These elements have their format, structure, metadata, volume, and dynamism (such as frequency of updates). It is essential to understand these elements (contents described by Pakkala et al [71]) and the associated attributes in order to deliver the correct content to the right user in their current context. For example, understanding the complexity of the text describing sleep disorders can help to personalize the content based on the educational background of the patients. A general assumption is that an educated person can comprehend more complex language and medical terms than a nonspecialist. For a nonspecialist, it is more valuable to present the same content in terms of animated videos or pictures.

The adaptive dimension provides the context for an adaptive IDPT system to tailor its behavior. A common way to tailor the behavior of the IDPT system is based on input regarding user preferences, measures (psychometric tests, user behavior analysis, and others), and goals, as shown in Table 4. In the context of user preferences, a study by Yardley et al [72] outlined that users can feel overwhelmed by the quantity and complexity of information presented to them. Hence, it is advisable to meet the needs of different users based on their needs and preferences. Personalization is critical mainly because users have different cognitive skills, educational backgrounds, content format preferences, comprehension capabilities, and other qualities. A similar conclusion was made by Sundar and Marathe [73]. Their study identified two types of users: those influenced more by the affordance of agency (power users) and those influenced by the relevancy of the resulting content. The results of the study revealed that nonpower users preferred personalized content, whereas power users rated content quality higher when the website had a customizable feature. Similar to user preferences, other dimensions of adaption are performance measures such as the use of psychometric tests (eg, Patient Health Questionnaire-9 [74] for depression), user behavior analysis, and others.

Adaptive strategies provide a mechanism to present the right content to the right people based on their needs and preferences. While our review findings reveal that the rule-based adaptive strategy is the most widely adopted practice, other strategies, especially machine learning, are becoming highly prevalent. Given the premise that we can capture every digital footprint of a user, resulting in a complex and comprehensive data set, there is a possibility of using sophisticated machine learning or deep learning algorithms on the one hand, but it also raises an essential question about privacy on the other. In general, to build an adaptive IDPT system, it is crucial to understand which adaptive strategies can be used. Based on the selected strategies, one needs to collect and store the data. No matter which adaptive strategy is chosen, the adaption in an IDPT system is an iterative cycle where data is collected and preprocessed; preprocessed data are then analyzed and, based on the results of the analysis, an action is taken to tailor the intervention. However, how the data are analyzed and the result is extracted affect the way an IDPT system is developed, the choice of IA, the method of data storage, and other parameters.

Although the integration of health care systems has emerged as a policy for several health care agencies, there is a large gap between current policy, program implementation efforts, and evidence for health care integration. The results of our review led us to list the following challenges in current IDPT practice.

It is not easy to predict how technologies will develop over time and whether these technologies will continue adapting to clinical use. However, based on the results of this systematic review, we outline some implications and future directions in the field of IDPT system development and innovation.

An immediate future task involves the creation of a conceptual framework for adaptive IDPT systems. In addition to this, we envision the development of domain-specific language that can model such an adaptive IDPT system. Furthermore, it is imperative from the review that there is a need for a comprehensive visual dashboard for therapists and patients where they can receive the intervention, monitor their symptoms, and manage their illness.

Given that the health ICT literature is quite diverse and extensive, the current study focused exclusively on internet-delivered interventions for mental health morbidities. Notwithstanding this limitation, this paper highlights the significance of the continued study of this intervention method. Another limitation is that our literature exploration only encompassed articles in the English language; therefore, it is plausible that some research conducted in other parts of the world and published in other languages were missed. A third limitation pertains to IDPT apps developed by industry that were not accessible for review. Hence, we have less knowledge about the adaptive elements involved in their architecture.

Adaptive psychological interventions tailor the type of content or tasks to individuals based on their needs and preferences in order to improve saliency and intervention efficacy. This systematic review describes the investigation and analysis of existing studies about adaptive psychological intervention delivered through the internet. The study outlines the main elements used in the process of adaptation, the IA used in the adaptive systems, the main dimensions of adaptation, and the main adaptive strategies. Based on these findings, we envision the development of a conceptual framework that researchers and clinicians can utilize to build adaptive models of several health care interventions.

The findings of our review indicate the use of web-based and mobile apps to deliver mental health interventions, such as for depression (most studied), anxiety, and others. However, a number of these studies did not report the IA used in their system, and those that did report mostly used tunnel-based systems. Similarly, several studies used rule-based adaptive strategies to adapt intervention based on performance measures such as psychometric tests. Feedback messages, reminders, and support were the most used adaptive element. Further study is required to explore the role of IA, adaptive elements, adaptive dimensions, and adaptive strategies in building a successful IDPT system. Knowledge about these core elements of the adaptive IDPT system can serve to create a conceptual framework that can be used for several health care interventions.