To ensure a thorough exploration of the research domain, a systematic search strategy was devised. This strategy encompassed three key concepts: depression and anxiety, digital mental health interventions, and adolescents and young individuals. Using search queries outlined in Table 1, relevant publications were retrieved. The searches across databases yielded 2,295 results, from which 867 unique sources were identified for further screening following the removal of duplicates (refer to Figure 1).

Following the PRISMA statement and flowchart (43), our review process comprised four phases: identification, screening, eligibility assessment, and final synthesis. Initially, articles meeting the following criteria were considered for retrieval: (a) peer-reviewed journal publications, (b) written in English, and (c) published between January 2013 and July 2024. From this initial retrieval, 1,428 duplicates were identified and removed. Subsequently, the remaining articles underwent screening based on their title and abstract, resulting in the extraction of 236 articles. Non-empirical papers, such as study protocols, were excluded during this stage, along with articles not primarily focused on depression and/or anxiety interventions or studies that were not carried out as an intervention. In this study we framed DMHIs as interventions that deliver mental health support primarily through digital platforms such as mobile applications, websites or chatbots. To ensure that only digital interventions were included, we excluded studies focusing solely on traditional face-to-face therapy, paper-based self-help materials, or interventions without a digital component. Further eligibility criteria included the age range of the study population [10–24 years; adolescents (10–19-year-olds) and young people (15–24-year-olds)] and the presence of personalization features within the intervention, particularly those that could impact usability. Ultimately, 53 articles met all inclusion criteria and were included in the final synthesis.

Apart from adhering to the systematic study selection criteria, the review articles obtained during the initial retrieval were examined to identify any additional relevant studies. Through this, 14 additional articles meeting the inclusion criteria were identified and subsequently included in the final synthesis.

A total of 67 articles were incorporated into the final synthesis, and a comprehensive review of the full-text articles was conducted. The initial extraction of study data was conducted by a single researcher (VW), after which two reviewers (VW and AC) independently screened the data. Final decisions were made through the agreement of both VW and AC. The extracted information included sample characteristics, research design, concepts/technology of intervention utilized, application name, personalization variables employed and the country where the intervention was conducted.

During the title and abstract screening phase, two reviewers screened 867 records, reaching agreement on 851 records (98.2% agreement rate). Discrepancies were resolved through discussion.

In the data extraction phase, two reviewers independently extracted the personalization variables reported in the 67 included studies, after which the results were compared. For 49 studies (73.1%), there was full agreement on the extracted personalization variables. In 14 studies (20.9%), initial differences in interpretation were resolved through discussion. In 4 studies (6.0%), partial agreement was reached when multiple personalization elements were present and interpretations varied. All cases were reviewed and consensus reached through discussion.

This study attempts to classify the personalization variables investigated in DMHIs in the recent past. We adopt the conceptual framework proposed by Hornstein et al. (44) which provides a range of categories both in terms of what is personalized (content, order, guidance and communication) and the mechanisms of personalization (user choice, provider choice, rule-based and by machine-learning model) in DMHIs. However, their framework excluded our customization and interactivity dimensions of personalization as they limited their framework to mechanisms affecting the therapeutic content and structure. Yet, non-therapeutic elements, such as changing the appearance of an avatar or goal setting, could enhance user engagement and adherence. For example, Hollis et al. (45) argue that the capacity for a young person to personalize minor aspects such as the gender and appearance of a “guide” might influence their perception and connection with a digital health intervention. Additionally, Birk and Mandryk (46) highlighted that avatar customization could increase enjoyment, effort invested, and time spent in a game-based intervention. Furthermore, it is evident that the use of interactivity can enhance engagement with digital interventions (47). Therefore, we have extended the conceptual framework proposed by Hornstein et al. (44) with the consideration of non-therapeutic elements such as user-driven tailoring and interactivity. This better aligns the framework with Blom and Monk (27) to more fully address the functionality, interface and distinctiveness aspects of their definition of personalization.

The derived framework includes six potential dimensions of personalization. Both therapeutic content delivered during a session and non-therapeutic content, such as information on mental health resources, will be considered to define the “content” dimension. The order of the sessions included in the intervention and the guidance provided by the providers will also be considered as therapeutic dimensions. Additionally, therapeutic communication, such as prompts and mechanisms targeting the timing of the intervention, and non-therapeutic communication, such as maintaining a personalized contact list, will be considered to define the “communication” dimension. To represent aspects of application personalization, absent from the earlier framework, “interface” and “interactivity” will be included as two additional dimensions. The interface dimension is defined as the customization of non-therapeutic content of the intervention, such as the visual appearance of avatars and application background. The interactivity dimension includes the personalization of user preferences in terms of application control and input, such as what data to share with the application, personalized methods of input, saving favorites, and replaying users' personal goals.

In terms of mechanism of personalization, this study uses the same four categories as (44). While this study primarily focuses on digital interventions, the most common personalization mechanism—User Choice—will be considered. Provider Choice will represent the mechanism of providing personalized feedback or suggestions by system moderators, therapists, or professional experts. For automated personalization mechanisms, rule-based and machine learning approaches will be considered as distinct mechanisms applying static (designer/expert-specified) or data-derived criteria for personalization, respectively.

The definitions and types of personalization dimensions and mechanisms are summarized in Figure 2 and Figure 3, respectively.

The chosen studies have employed diverse methodologies for user data collection, with randomized controlled trials (RCTs) emerging as the prevalent choice, closely trailed by mixed-method studies which employed both qualitative and quantitative research designs. Qualitative study designs were also employed to investigate user experiences with various approaches such as semi-structured or focus-group interviews or user surveys. At lesser frequency, pilot studies were conducted to assess the feasibility and acceptability of the interventions, followed by user-centered study designs to incorporate user feedback and open trials. Apart from these prevailing research designs, longitudinal studies, pre-post studies, co-designs, usability testing, exploratory, experimental, uncontrolled trials and large-scale evaluation studies (of acceptability and utility) were also observed. Figure 4 illustrates the distribution of methodologies adopted across the selected articles.

Among the studies concentrating on a single application or intervention, Cognitive Behavioral Therapy (CBT) emerged as the most frequently utilized therapeutic technique. Additionally, mood tracking, positive psychology, psychoeducation, acceptance-based therapy, self-assessment, mindfulness, games or gamification, and behavioral change and activation were each employed in more than two instances within the selected articles.

Some apps were used more than once in the selected interventions. Both the apps “The Grow It!” (48–51) and “BlueIce” (52–55) were used in four interventions. “Thought Spot” (56–58) was used in three interventions. “Reachout.com” (59, 60), “ibobbly” (61, 62), POD Adventures (63, 64), MoodHwb (65, 66), Sleep Ninja (67, 68), Rebound (69, 70) and JoyPop (71, 72) were each utilized in two studies.

The largest number of studies was conducted in the United States of America (USA), followed closely by Australia, and the United Kingdom (UK), and then the Netherlands and Canada. Conversely, only six studies were conducted outside of North America, Oceania, and Europe regions. Figure 5 depicts the distribution of studies categorized by country.

When selecting participants for the studies, only five studies specifically targeted indigenous or First Nations individuals (61, 62, 72–74). Eighteen studies focused on recruiting university students (52, 56, 57, 75–89), and six studies aimed to recruit secondary school students (63, 64, 90–93). One study was conducted with sexual minority men (94), while another study exclusively recruited females (71). One study recruited “black or biracial seventh- and eighth-grade adolescent females” (95). The remaining studies targeted general adolescent or youth populations.

This section examines the personalization aspects in the selected articles, focusing on dimensions and mechanisms as defined above. It discusses the types of dimensions and mechanisms reported, as well as the frequency of their application in the interventions.

The findings of the study characteristics showed where the DMHIs included in this review were conducted, the methods they used, the techniques they employed, and the characteristics of the participants. The selected studies for this review were mainly conducted in America, Australasia, and Europe. In recent years, studies that used personalization in DMHIs for depression and anxiety in adolescents and youth have been limited to only four studies on the Asian continent. When compared to America, Australasia and Europe, Asia exhibits lower or decreasing age-standardized incidence rates of depression in adolescents/youth aged 10–24 years (115). This trend may have influenced America, Australasia and Europe regions to increase their research on adolescent mental health compared to Asia. However, in light of total population, Asia and Africa need to be considered as under-studied in this domain.

With respect to the characteristics of the participants, only one study considered sexual minority men. Otherwise, rainbow users were unaddressed even though it is reported that transgender and gender-diverse youth experience considerable mental health disparities when compared with their cisgender peers, including higher rates of depression, anxiety, and suicidality (116). Although this review does not cover all DMHIs for depression and anxiety in adolescents and youth, the lack of research on personalized user experiences on DMHI engagement for gender-diverse young populations is notable. Comparatively, the presence of studies with indigenous and First-Nations young people in the findings shows at least some attention to DMHIs for these communities. Furthermore, only one study specifically investigated the impact of DMHIs on Black or biracial females (95). Black and multiracial adolescents face the highest risk of suicide among US college students (117) but are the least likely to access preventive therapy (118). Research suggests that mental health providers may be reluctant to acknowledge their role in perpetuating racial disparities in care (119). Therefore, the findings of this review provide further evidence of this reluctance, highlighting the need for future research to improve DMHIs tailored to multiracial adolescents and youth.

Almost all the apps and websites that incorporate personalization among the reviewed interventions were non-commercialized. There were no findings addressing the effect of personalization of commercialized mental health apps such as Headspace, Calm or Replika on youth depression and anxiety. The barrier to the usage of commercialized apps in scientific interventions could be their high subscription fees and freemium models or licensing constraints. Also, having strategies like freemium negatively affects user experiences specifically for vulnerable users (120). Therefore, interventions tend to use either a free version of commercialized apps (121) or shift their focus to non-commercialized apps.

Personalization has been suggested to be effective in increasing adherence to DMHIs among youth (122, 123). This review investigates what has been personalized and how the personalization has been applied in recent DMHIs for adolescents and young people with depression and anxiety. The review adapted the conceptual framework of Hornstein et al. (44) and modified it to categorize both the therapeutic and non-therapeutic personalization elements discovered in the selected interventions. Considering non-therapeutic personalization elements, Hornstein et al. (44) “content” and “communication” dimensions have been extended and two new personalization dimensions, namely “interfaces” and “interactivity”, have been added. Conversely, the definitions of personalization mechanisms have been taken as defined by the original framework.

Personalizing therapeutic content was largely evident in this review. Yet, facilitating change in the order of the sessions or modules was weakly addressed. Among the sixty-seven articles, only three (89, 106, 107) have reported enabling personalization of the order of therapeutic sessions of modules (4%). A similar observation was made by Hornstein et al. (44), who also found that 4% of the interventions enabled personalization of the order. Nevertheless, even though this aspect of personalization is underexplored, tailored sequencing of modules specific to the type of mental health disorder, especially when interference such as comorbid conditions and stressors occur, has been found to be effective with youth (124). Further, some study participants have argued for the convenience of tailored sequencing of modules to one's needs as the interventions could become less attractive if they already understood the concepts but still have to follow the order of modules to complete (125). Therefore, more thorough investigation is needed to understand the possibility of applying personalization of the module sequences in DMHIs for youth and why it has not been widely applied.

The review found that personalization of DMHI interfaces was mainly focused on customizing the appearance of avatars such as clothing or hairstyles. However, offering users the ability to personalize other characteristics of an avatar such as gender or incorporating gender-diverse characters in game-based interventions could enhance user engagement, as discussed in recent studies (45, 126). However, only one study provided users with the ability to customize their avatar's gender (111), though the options were limited to a binary selection. Yet, employing gender as a characteristic to personalize the experience was also not visible in the selected studies for this review. The reluctance to employ gender in therapeutic or non-therapeutic personalization dimensions could be caused by the recent recognition of gender as a spectrum instead of binary and the recognition of the importance of inclusivity and diversity in research and design (127, 128). Using gender as a variable to personalize user experience may inadvertently exclude or marginalize individuals who do not identify with the provided options.

The lack of generative AI in the retrieved studies is notable. Machine learning was widely used to personalize therapeutic and non-therapeutic elements in the studies. Yet, the study incorporating the greatest range of AI personalization (89) was still a largely template-and-rule based method. While the rapidly growing field of generative AI holds significant potential for widespread use in digital mental health interventions—ranging from mental disorder detection and counselling support to therapeutic applications, clinical training, decision-making support, and goal-driven optimization—its inherent limitations underscore its role as a supplementary tool rather than a replacement for mental health professionals (129). Further, generative AI could be used in DMHIs to provide personalized experiences to users with conversational support, mood predictions, and risk assessments (130), and it could assist mental health practitioners in re-defining mental illnesses more objectively and personalizing the treatment based on user characteristics (131). Despite the potential benefits of AI in this domain, lack of sufficient data to train the models (132), ethical challenges such as biased data (e.g., instinctive and expressive quality of clinical text data, connecting mental disorders to specific ethnic groups, etc.) (133) and lack of established standards to guide the use of AI in healthcare settings (134) could be among the reasons for hesitation in AI use with DMHIs. Ameliorating these challenges with standard practices to produce clinical data (135) and use AI in healthcare (136) would be beneficial for an enhanced user experience.

Automated mechanisms such as rule-based and machine learning in this review mainly take into account user's symptom profile, mood state and responses to pre-defined questionnaires. There is a lack of evidence on using user characteristics such as culture, age, ethnicity and personality. However, these characteristics were evidently employed in DMHIs in general (40, 137, 138). Although culture has not been directly recognized as a personalization variable, Thabrew et al. (139) observed positive feedback on using features to increase cultural appeal such as introductory “karanga” (Māori ceremonial welcome call). Hence, culturally appropriate content could be adapted to enhance user engagement. Further, this review being narrowed in its scope to adolescents and youth could be pointed out as a reason why age is rarely recognized as a variable that was used to personalize the user experience. However, one study in this review (114) used customizable iconography, allowing users to choose between a straight-lined or cartoon-styled design. Within the 11–19 age range, the straight-lined style was aimed at older adolescents, while the cartoon-styled design targeted younger adolescents. Such segmentation was also observed in the study by Fleming et al. (138), which found that perceptions of gaming contexts differ between younger and older adolescents as younger adolescents tend to prefer interactive and gamified interventions while older adolescents preferred straightforward and concise designs. Therefore, DMHIs for adolescents and youth could still use age as a variable to personalize their engagement. Personalizing user experience based on user's ethnicity may not have been evident in this review because, similar to gender, ethnicity or race is a complex social construct with significant variation within and across racial groups. Therefore, treating ethnicity as a monolithic variable oversimplifies individuals' experiences and identities, ignoring the diversity and intersectionality of race with other aspects of identity. Further, a recent study (40) witnessed that understanding the user's evolving personality traits can help gauge their receptiveness to targeted content, taking into account factors like competitiveness or openness.