We used the Theory Coding Scheme (TCS) developed by Michie and Prestwich [65] to describe the theory basis of each intervention and to assess to what extent theories were utilized. This checklist has 19 items to verify whether an intervention mentions a theory or model, to what extent a theory or model is used in designing intervention features, and whether the theory is tested in pre/poststudy. Item 1 verifies whether a particular theory is mentioned in the studies, whereas items 2 and 4-11 measure to what extent the theory is used during designing interventions. Item 3 checks whether more than 1 theory is applied. Item 12a verifies whether the impact on adherence mediators was measured. The remaining items were left uncoded as they deal with reliability of methods in measuring and testing theory, and theory refinement, which is beyond the scope of this study. Each item was coded as 1 if applied and 0 if not present. As a last step and composite measure, the included eHealth systems were graded on the extent of theory usage as no, low, medium, or high usage. Studies were labeled as follows: “no,” if there was no mention of theory, or they mentioned a theory but there was no grounding of the intervention on it (TCS item 1); “low,” if it was explicitly stated that theory concepts were used to design the intervention (1 or more of the TCS items 2, 5, 8, and 11); “medium,” if all intervention techniques originated from theory concepts (TCS item 7 or 9); “high,” when all theory concepts were addressed, or used to select participants or tailor interventions (1 or more of TCS items 4, 6, and 10).

We investigated studies for the presence of discussions concerning how developmental theory concepts influence ART adherence among adolescents. This examination was particularly inspired by TCS item 1 which says “Theory/model of behavior mentioned—Models/theories that specify relations among variables, in order to explain or predict behavior are mentioned, even if the intervention is not based on this theory” [65]. Hence, we examined the presence of explicit discussions on how developmental changes (biological, social, psychological) would “explain or predict” (to use the exact words) ART adherence among adolescents.

For coding BCTs from intervention descriptions, BCTTv1 [52] was used. BCTTv1 is a well-established taxonomy with an extensive list of theoretical methods of behavior change, containing 93 BCTs grouped into 16 clusters. It has been widely applied to specify intervention techniques in various behavioral domains such as physical activity [66], alcohol use [67], and medication adherence [68]; as well as to identify the presence of BCTs in existing interventions [41,62]. To examine the use of BCTs in interventions, the prevalence of individual BCT across all interventions was calculated. Similarly, we also calculated the total number of BCTs per intervention. Considering the complexity of coding BCTs, coders completed a certified online training on BCT taxonomy (BCTTv1) [69].

Because BCTs are only descriptions of behavior change from a psychological perspective, in other words they are technology agnostic, we additionally coded the designed features, using the lenses of motivational design for mHealth developed by Geuens et al [57]. Geuens et al [57] explained how theoretical concepts of behavior change can be translated into design principles, by including also implementation instantiations of these principles through examples of mobile app features. These lenses of motivational designs provide implementation-level descriptions of design principles, also encompassing persuasive principles of [54-56]. The prevalence of individual design features among systems and the number of features per each system are calculated.

Because interventions differ substantially in measuring their impact, each outcome measure was coded (mediators, adherence, or health outcomes) as positive effect (+), no effect (0), or negative effect (–). Hence, we did not exclude any outcome measures or limit ourselves to direct measures of adherence such as medication intake or viral load. This inclusive approach is justified by the Information–Motivation–Behavioral Skills Model [47] of ART adherence, as it accommodates a broad range of outcome measures [47,70], and is widely adopted in ART adherence research [71-73]. This model conceptualizes ART adherence as a function of 3 mediators: information about HIV and ART, motivation to take medication, and behavioral skills required for taking medication. An increase in any of these constructs is theorized to result in improved adherence behavior which in turn produces favorable health outcomes.

A total of 10 different theories were mentioned (Table 3), of which 3 theories appeared commonly, namely, Social Cognitive Theory, the Information–Motivation–Behavioral Skills Model, and Motivational Interviewing. The complete theory coding sheet of included studies using the TCS is provided in Multimedia Appendix 3.

Of the 16 eHealth systems, 9 were grounded in theory. However, the extent of theory utilization (the extent to which interventions address particular theory-relevant constructs [65]) varies substantially. Four interventions were supported by more than 1 theory. Of these, 1 (Hightow-Weidman et al [93]) was guided by 3 theories, whereas 3 interventions (Tanner et al [79,80], Spratt et al [82], and Shegog et al [83]) combined 2 theories. Table 3 provides an overview of the specific theories used per intervention. The remaining 7 interventions were not grounded in theory. While 3 of them mentioned a particular theory and its constructs in relation to adherence, they were not utilizing it to inform the design of the intervention. The remaining 4 did not refer to theory at all.

Based on the degree of application of theories in designing interventions, that is, based on the TCS [65] (also see Multimedia Appendix 3), we grouped studies into 4 usage categories as stated in the “Methods” section: no, low, medium, and high (Table 4). As much as 7 of 16 interventions have no use of theory, 5 have low theory usage, 1 used theory moderately, and 3 used theories extensively.

Given the importance of using suitable theoretical foundation for interventions that address adolescents (see the “Introduction” section), we explored whether the studies included in our review adequately address this aspect in how they report their work. Hence, we entirely examined them for information concerning the influence of developmental changes (biological, social, psychological) on ART adherence among adolescents. For example, we explored whether the studies considered the question of which theories could work better for adolescents (ie, which theory would provide better coverage of factors associated with developmental changes). Our findings suggest that none of the theory-informed eHealth interventions explicitly discussed theory selection from a developmental perspective: none of the studies provided a discussion of why or how the respective theories were appropriate for adolescents, suggesting lack of consideration of developmental changes. However, they mentioned previous use of those health theories in behavior change research including in adolescent populations (see Multimedia Appendix 2), although they did not include explicit reasoning for their choice. This may suggest that, in general, developmental changes related to adolescence are currently not included in the choice of theoretical foundation for interventions to increase ART adherence among adolescents in a transparent way, leaving room for future work that explicitly draws from suitable theory to achieve better outcomes as suggested by [102]. Additionally, we searched for evidence that developmental changes were accounted for in the intervention design—whether any specific design features were related to these changes—but no such explicit association was found.

Across the 16 eHealth systems reviewed, a total of 42 BCTs were identified (Multimedia Appendix 4). The most popular technique was “Prompting/cueing,” which was used by 11 interventions. The second most frequent technique was “Social support (unspecified),” used in 10 interventions. The third was “Problem solving,” used in 8 interventions. Next, “Monitoring of behavior by others without feedback” was used 7 times, and “Information about health consequences,” “Credible source,” “Social support (emotional),” and “Instructions on how to perform behavior” were each used 5 times. “Demonstration of the behavior” has been used 4 times and the remaining techniques appeared 3 times or less.

The median number of BCTs used was 5. The intervention with the highest number of techniques was by Dworkin et al [97-99] utilizing 17 different ones. The lowest number of BCTs was found in Abiodun et al [101], using 1 technique. A complete list of details on how many techniques each intervention has applied is provided in Multimedia Appendix 4.

Most theory-based studies (7/9: Shegog et al [83], Belzer et al [84-86], Garofalo et al [96], Whiteley et al [77,78], Tanner et al [79,80], Hightow-Weidman et al [93], Dworkin et al [97-99]) linked at least one intervention technique to 1 theory (TCS item 8) and at least one theory-relevant construct to an intervention technique (TCS item 11). However, only half (4/9: Whiteley et al [77,78], Tanner et al [79,80], Hightow-Weidman et al [93], Dworkin et al [97-99]) linked all theory-relevant constructs to at least one intervention technique (TCS item 10), and only 5 out of 9 (Whiteley et al [77,78], Tanner et al [79,80], Hightow-Weidman et al [93], Garofalo et al [96], Dworkin et al [97-99]) linked all intervention techniques to at least one theory-relevant construct (TCS item 7). Two (Shegog et al [83] and Naar-King et al [91,92]) studies used theory constructs to tailor intervention techniques to recipients (TCS item 6), but no study selected recipients for the intervention based on theory constructs (TCS item 4).

Many of the systems were simplistic applications from a technical perspective developed on basic phones (ie, SMS text messaging [n=6] and phone calls [n=3]; Table 5). However, advanced applications designed for smartphones were also emerging (n= 3). These smartphone-based apps were designed in various forms—ordinary apps, serious games, and gamified apps. Already existing social media apps were also utilized. Electronic adherence monitoring devices (electronic medication containers that look like ordinary bottles or mobile phones, eg, WisePill) were also common, but mainly used in combination with other platforms and not as a standalone system. The primary purpose of these devices was to objectively measure adherence, except in 1 intervention (Spratt et al [82]) in which it was used to deliver reminders in the form of blinking lights and chime sounds. Other systems included web-based applications (desktop/laptop), remote videoconferencing, and desktop applications (Naar-King et al [91,92]). Such systems were, however, less frequent, each appearing just once.

Out of the 28 motivational design features proposed by Geuens et al [57], we found 24, with the most frequent features being “Reminders,” “Personalization,” and “General information”—appearing 11, 10, and 9 times, respectively (Multimedia Appendix 5). Information to educate adolescents about HIV and ART adherence was sometimes tailored to participants based on personal profiles—hence “Microtailoring.” “Instructions” on how to perform certain tasks such as taking medication and talking to providers were also detected. In some interventions, generic information and instructions were from health care expert sources (ie, “Expertise”). Asking patients to manually enter information (ie, “Logging”) about performance of their behavior (eg, whether they took medication or not) or outcome data (eg, viral load, CD4 count) was also common, although sometimes this was also done automatically, via “Tracking,” using electronic devices such as medication adherence monitoring devices.

The median number of design features was 4. The intervention with the highest number of design features (n=15) was Hightow-Weidman et al [93], while the interventions with the lowest number of features (n=1) were Puccio et al [90] and Abiodun et al [101] (Multimedia Appendix 5).

We adopted Johnson et al’s [74] method to categorize papers on methodological quality, computing the quality of evidence score, ranging from 5 to 15. Papers with a rating 8 or below are categorized as “weak evidence,” 9-12 as “moderate evidence,” and 13 and above as “strong evidence.” Six interventions scored strong on quality of evidence, 4 moderate, and the remaining 7 weak (Table 6). The quality ratings only pertain to the strength of empirical evidence for outcome effects, and do not judge the overall quality of the studies.

As mentioned in the “Methods” section, outcome measures were coded following Information–Motivation–Behavioral Skills Model’s conceptualization of ART adherence behavior as a function of 3 mediators: information, motivation, and behavioral skills [47]. Therefore, the measures corresponding to these parameters are knowledge on HIV and ART, personal motivation and social motivation, and ART self-efficacy, respectively. Direct measures of adherence behavior are ART medication adherence and appointment adherence. Similarly, measures of biological outcomes include viral load and CD4 count. The number of measurements for each outcome, effect type, and quality of evidence is summarized in Table 6. Of the 16 eHealth systems, 10 reported positive effect on 1 or more of the outcomes measured, yet 7 of these also reported no effect on 1 or more of other outcomes. As much as 6 out of 16 reported purely no effect in all the outcomes measured.

The nature of this review precludes any firm conclusions. A closer look at Table 6 suggests mixed results; we cannot unambiguously conclude that more extensive grounding of design features in theory related to health behavior or behavior change is associated with better (significant) effectiveness on outcomes with good evidence quality (compare studies with footnotes c and e). However, the results suggest that 10/16 (Whiteley et al [77,78], Tanner et al [79,80], Hightow-Weidman et al [93], Dworkin et al [97-99], Dowshen et al [94,95], Garofalo et al [96], Shegog et al [83], Belzer et al [84-86], Dulli et al [100], Abiodun et al [101]) of the included eHealth apps report a positive impact on ART adherence or on its mediators.

The review shows that current eHealth systems to improve adolescents’ ART adherence refer to theory only lightly. Of the 16 included eHealth systems, only 4 interventions show extensive to moderate usage of theory, while the remaining have low or no usage (Table 4). This finding contrasts with the studies that have argued that grounding the design of eHealth systems in theory is associated with increased effectiveness [40-42]. In the context of HIV in particular, designing eHealth systems based on behavioral theory is associated with efficacy in improving adherence to HIV medication [30]. This is attributed to the fact that health behavior theories provide a comprehensive understanding of ART adherence behavior and its determinant factors, to inform the design of the intervention [44,103]. Theory underutilization may result in a limited understanding of the different moderators of ART adherence among adolescents, leaving several factors unaddressed by the intervention, which in turn reduces effectiveness [104].

Moreover, our findings indicate that the context of developmental changes and their impact on ART adherence are not explicitly addressed in the included studies. This is somewhat surprising as prior research has shown that factors associated with developmental changes affect adolescents’ health including medication adherence [105]. Adolescence presents a specific developmental stage marked by profound biological, psychological, and social changes [106-111] that may affect ART adherence behavior [20]. For instance, we did not find explicit mention of how the ongoing development of cognitive capacity [17], oppositional behavior associated with becoming autonomous [21], and conformity to peer pressure [22] and the synergy it may have with HIV stigma [112] can affect ART adherence among teenagers. Being informed of these developmental theory concepts might be useful in the choice of appropriate BCTs and design features. However, the absence of explicit referral to developmental theory does not automatically imply that interventions are not (implicitly) informed by developmental theories.

Almost half of the interventions applied only 4 or less BCTs, which may render them less effective in delivering intervention content: prior research on BCTs and health behaviors, (eg, [40-42,62]) has indicated that the number of BCTs applied in an intervention influences its effect, that is, interventions that employ more BCTs were found to have a larger effect on behavior than those that apply fewer BCTs [41]. This relatively low number of BCTs usage might be the result of a less comprehensive conceptualization of adherence and its determinants, or a less comprehensive conceptualization of adolescents’ developmental changes and their impact on ART adherence. For example, an intervention narrowing adherence to “consuming medication,” and focusing primarily on “forgetfulness” as main barrier, might end up employing the “prompts/cues” BCT only; “reminders” as the main technique can be seen in [90,94]. A deeper look through the perspective of developmental changes (ie, a reflection on why adolescents forget) might perhaps suggest another solution; for example, building adolescents’ skills on how to better integrate medications into daily life. In terms of breadth (covering a broad range of factors), reviews on barriers of adolescent’s ART adherence list many factors complicating adherence [9-11] that eHealth systems need to address. Moreover, we found that BCTs which might be relevant for adolescents such as demonstration of behavior (modeling), peer comparison, or incentivization were scarce.

We found that most included systems were limited to SMS text messaging or phone calling. Compared with more advanced information and communications technology systems, SMS text messaging and phone calls offer the advantage of being cheap to deliver. However, such systems also limit the implementation of more sophisticated features. For example, it is challenging to deliver intervention content enriched with engaging audiovisual content. SMS text messages alone might not be sufficiently engaging to adolescents, which could be a possible explanation for researchers reporting noneffectiveness [88,113,114]. A previous work noted that for adolescents “texts that say the same thing are boring” [77], and the authors recommended dynamically changing the content of the message so that adolescents do not get tired of reading the same messages [89].

As for the designed features, “reminders (notifications)” and “personalization (adapting color schemes and skins)” were the most frequently applied followed by displaying “general information.” The remaining features were scarcely used. Providing reminders for medication, enabling customization on system features, and educating adolescents about HIV and ART are still appropriate. However, interventions could be more effective if they additionally include “social” features such as principles grouped under “reward and incentives” and “social interactions” categories [57].

Of the 16 eHealth systems, 10 (Whiteley et al [77,78], Tanner et al [79,80], Hightow-Weidman et al [93], Dworkin et al [97-99], Dowshen et al [94,95], Garofalo et al [96], Shegog et al [83], Belzer et al [84-86], Dulli et al [100], Abiodun et al [101]) reported positive effect on 1 or more of the outcomes measured, yet 7 (Whiteley et al [77,78], Dworkin et al [97-99], Dowshen et al [94,95], Garofalo et al [96], Belzer et al [84-86], Dulli et al [100], Abiodun et al [101])out of these 10 also reported no effect on 1 or more of other outcomes. Six out of 16 (Stankievich et al [81], Spratt et al [82], Saberi et al [87], Linnemayr et al [88,89], Puccio et al [90], Naar-King et al [91,92]) reported purely no effect in all the outcomes measured. Overall, while evidence is mixed, the impact of these systems on ART adherence and its mediators has a positive trend. However, the nature of this review and diversity of eHealth studies preclude any firm conclusions based on our study findings (Table 6). Considering the relatively small number of digital interventions identified in this review and the mixed evidence on its impact, more digital interventions for adolescents’ HIV self-management should be evaluated to come to firm conclusions.

A clear understanding of theoretical insights enables designers to translate various BCTs into features appropriate to this specific audience; the way BCTs are implemented and presented to users matters in respect to their effectiveness. Moreover, implementation sophistication on top of applying relevant intervention content is paramount. In the paragraphs below, we detail the recommendations derived from this review.

Deciding on a particular theory is challenging for eHealth designers as a plethora of health theories exist, and there is not one theory specifically developed for adolescents. In this respect, protocols for developing behavioral interventions such as the IM (intervention mapping) [115] and capability, opportunity, motivation, behavior frameworks [116,117] may be of interest. According to IM, one should first consult existing literature to list reported causes of the target health problem (in this case, barriers/facilitators of ART adherence among adolescents). Next, concepts from this list should be linked to theoretical constructs (ie, the theory with better coverage of this list could be a good candidate). Moreover, use of a combination of multiple theories (such as Integrated Behavioral Model or any customized combination) might be considered. Complementing this approach, the practical study of the actual target group (eg, user-centered methods such as participatory design [118,119] with adolescents) might also help identify the most relevant determinants (eg, [120]) on which an intervention can focus [121].

Incentivization holds particular potential to improve adherence to HIV care [122-125] among adolescents. Moreover, as peer pressure among teen population is high [19], implementing social-based techniques involving peer role modeling and comparisons among peers might be effective. In this respect, Hightow-Weidman et al [93] reported that participants enjoyed “social-media-like” discussion boards that prompt peers to daily discuss on HIV matters and share experiences. However, any use of social media with adolescents should be done with utmost care, discussed further below.

Additionally, methods that help adolescents analyze barriers they face, and help them generate strategies to overcome these issues might be crucial. This is already addressed by some interventions that seek to develop problem-solving skills, for example, phone conversations with a professional adherence counselor [87] or a similar discussion with a humanlike character animated in computer software [91].

More advanced and engaging design features such as gamification and social connections are needed in eHealth systems. Although evidence on the efficacy of gamification specific to adolescents’ ART adherence is not synthesized yet, literature on gamification and adherence indicated its potential [126,127] and positive influence on health behavior change in general [74,128-132]. First, as mentioned above, adolescents might be more attracted to immediate rewards [17] (eg, points) and incentives associated with adherence achievements (eg, badges) than foreseeing future long-term health consequences. In addition to material rewards, virtual rewards in the form of gamification elements such as points, levels, leaderboards, and easter eggs might be interesting. Additionally, as adolescents are heavily influenced by peer pressure [13], integrating “social interaction” principles [57] might be important. Allowing adolescents to link and share via social media platforms, connect with others in a similar condition and demonstrating their success (social identification), know how other peers are performing (social comparison), compete or cooperate with peers on social interaction environments could be useful. However, we acknowledge that given the specific developmental characteristics of this age group and the dangers of issues related to HIV stigma (eg, concerns similar to privacy unraveling [133,134]), any such designed features should be stigma sensitive and introduced with the greatest care.

Finally, to enable using more advanced features, implementations such as desktop applications, mobile apps, games, or including gamification elements and stigma-sensitive social media–like features might be more appropriate to engage teenagers. As such applications are emerging, implementations in these platforms are warranted in future interventions.