This narrative review was informed by structured systematic search, screening, and extraction procedures. A narrative synthesis was selected because the evidence spans clinical trials, qualitative inquiry, implementation research, and computational studies and because the review examines how engagement mechanisms are implemented and interpreted rather than estimating a pooled effect. A completed PRISMA-P (Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols) checklist is provided in Checklist 1.

The original search covered PubMed, IEEE Xplore, and Scopus from January 2018 through April 2025. Search concepts combined digital mental health, stress or distress, engagement, and adaptive mechanisms, including personalization, biometric feedback, EMA, JITAIs, gamification, conversational agents, LLMs, and reinforcement learning. Well-being terminology was also considered because preventative DMHIs may describe emotional support or coping without using stress reduction as their primary label. A further update covering May 2025 through August 2026 was conducted using the same concept blocks for the focused database searches and a supplementary journal search of recent peer-reviewed work. The review domains and illustrative search concepts are summarized in Table 1, and the full database-specific strings, dates, and the supplementary search description are reported in Multimedia Appendix 1.

Studies were included if they (1) described a DMHI aiming primarily to reduce stress or improve related affective outcomes within preventative mental health contexts; (2) reported at least 1 engagement-supportive mechanism, such as personalization, biometric feedback, JITAI, gamification, or AI or chatbot functionality, implicitly or explicitly; (3) presented empirical, qualitative, implementation, simulation, modeling, or system design evidence relevant to engagement; (4) were peer-reviewed; and (5) were published in English. Protocols, reviews, dissertations, nondigital interventions, and non–mental health interventions were excluded. A study was also eligible if it addressed 1 or more engagement-supportive mechanisms and met the broader stress-focused criteria; it did not need to include every mechanism or use the term “engagement” explicitly. Commercial applications, app store–only evaluations, and other gray literature were out of scope unless evaluated in a peer-reviewed study meeting these criteria. Two technical preprints from the original corpus were retained as a narrow exception because they addressed emerging AI and reinforcement learning architectures [25,37]; they were interpreted as design or computational evidence rather than clinical effectiveness evidence. No new preprints were added, and the evidence update did not broaden the eligibility criteria.

MM and SAP independently screened titles and abstracts and then assessed potentially eligible full texts in the original review. Disagreements were resolved through discussion with SG until consensus was reached. MM and SAP extracted data from studies identified in the original search using the structured 17-item form described below. For studies identified in the search update, MM conducted the initial screening and data extraction using the unchanged criteria, and SG verified the draft eligibility decisions, extracted data, and interpretations.

A structured 17-item form captured publication details, population, setting, intervention modality, duration, follow-up, study design, engagement definitions and measures, engagement-supportive mechanisms, AI or sensing components, key findings, and design implications. The same form was applied to the papers identified in the search update. Eligibility and mechanism coding were kept separate so that methodological studies could be included in the corpus without being misclassified as intervention-level engagement mechanisms.

The synthesis proceeded in 2 stages. First, descriptive mapping characterized study design, population, modality, duration, and follow-up. Second, overlapping mechanisms were coded as static tailoring, dynamic adaptation, conversational support, context-aware prompting, biometric-responsive adaptation, gamified reinforcement, narrative immersion, or social and co-design support. In this review, AI encompassed conversational agents and LLM-based systems; recommender systems; contextual bandits; reinforcement learning, rule-based, or fuzzy logic approaches; and simulation or modeling used for assessment, timing, content selection, or interaction. Higher-level labels such as JITAI or AI were classified according to their functional role in sensing, decision-making, or delivery.

Findings were synthesized across behavioral, cognitive, experiential, relational, and methodological evidence. Behavioral indicators were distinguished from subjective or interaction-oriented engagement processes, and associations were interpreted descriptively because interventions frequently combined mechanisms and rarely isolated their independent effects.

Self-determination theory (SDT) informed the interpretation of motivational and experiential engagement. Meaningful choice may support autonomy, appropriate feedback and achievable challenges may support competence, and socially responsive interaction may support relatedness [38,39]. Work on immersive digital environments provided further context on need satisfaction, motivation, enjoyment, and continued participation [40]. These were theory-informed interpretations and included studies were not assumed to have explicitly adopted or tested SDT.

This review analyzed published reports only and did not recruit participants, intervene in care, or collect or process identifiable personal data. Ethics approval was therefore not obtained. This decision was consistent with the University of Padua ethics self-assessment and JMIR guidance [41,42]. Informed consent and participant compensation were not applicable.

Figure 1 summarizes the selection process. The original search identified 2088 records; 179 duplicates were removed, and 1909 records were screened. After 1797 title or abstract exclusions, 112 full-text reports were assessed, and 22 studies met the original criteria. Eleven unique peer-reviewed candidate empirical papers identified during the evidence update were screened against the same criteria [43-53]. Ten studies met all criteria and were included in the review corpus [43-48,50-53]. The insomnia component trial [49] fell outside the stress or related affective eligibility boundary and was retained only as external interpretive evidence. Four reviews and meta-analyses [33-36] also informed interpretation but were not counted. The final review corpus comprised 32 studies.

The 32 included studies were published between 2021 and 2026 [8,9,12-14,24-27,37,43-48,50-65]. Of these, 30 appeared from 2024 onward, while 1 was published in 2021 [59] and 1 in 2023 [65], reflecting the rapid recent growth of adaptive, context-aware, and AI-enabled DMHI research. Dissemination spanned clinical, digital health, human-computer interaction (HCI), and computing journals and conferences, illustrating the multidisciplinary intersection of mental health, behavioral science, and technology.

Most empirical studies were conducted in high-income settings in North America, Western Europe, Australia, New Zealand, East Asia, and Israel. Culturally adapted interventions in Jordan provided evidence from an otherwise underrepresented regional context [57,63]. Target populations included adolescents, young adults, and students with stress-, anxiety-, or mood-related needs; caregivers and parents; employees and health care trainees exposed to occupational stress; people awaiting psychotherapy; users of youth mental health platforms or mental health chatbots; breast cancer survivors; and people with chronic physical illnesses, including chronic obstructive pulmonary disease [8,9,12-14,24-27,37,43-48,50-65].

Intervention modalities were correspondingly diverse. They included mobile apps with monitoring or guided coping strategies [13,14,43]; web-based and blended programs [8,12,47,52]; wearable-integrated systems using physiological or stress detection to trigger support [37,56]; narrative or gamified interventions [51,55,59,65]; chatbot- and LLM-based conversational systems [24-27,46,53,61,63]; context-aware JITAIs [37,43,44,48,56,64,65]; and simulation, modeling, or adaptive trial frameworks relevant to decision-making and optimization [45,50,52,54,62]. Across these modalities, engagement was examined through behavioral use and adherence, perceived relevance and burden, emotional or relational experience, contextual receptivity, and methodological optimization. Modalities and engagement-supportive mechanisms frequently overlapped within the same intervention.

Study characteristics and design distribution are summarized in Table 2. Randomized or microrandomized controlled trials were the most common primary design, accounting for 53% (17/32) of the studies [12,13,24,26,44-46,48,50-52,57,59,61,63-65]. Qualitative investigations accounted for 19% (6/32), including interview-based, experiential, diary-based, and reflective designs [9,14,27,53,56,58]. Design or implementation studies accounted for 9% (3/32) [8,25,37]. Mixed methods studies also accounted for 9% (3/32) [47,55,60]. Simulation or modeling studies represented 6% (2/32) [54,62], while 1 quasi-experimental study accounted for 3% (1/32) [43]. These mutually exclusive categories reflect each study’s primary methodological contribution, as some controlled trials also incorporated mixed methods components. Overall, 81% (26/32) of the included papers reported empirical studies involving conventional participant samples [8,12-14,24,26,27,43-48,50-53,55,57-61,63-65], while 19% (6/32) focused on design, reflexive inquiry, simulation, modeling, or reinforcement learning approaches without a conventional participant sample [9,25,37,54,56,62].

Table 3 summarizes the presence or absence of the primary engagement-supportive mechanisms identified across the 32 included studies. Studies were classified using overlapping mechanism categories rather than mutually exclusive intervention types because many stress-focused DMHIs combined adaptive, conversational, contextual, motivational, or social components within the same intervention architecture.

Personalization reflected 2 conceptually distinct approaches. Static tailoring appeared in 28% (9/32) of the studies and used predefined characteristics, preferences, symptom profiles, cultural factors, or user-provided information to select or modify support [8,12,13,43,44,52,57,63,64]. Dynamic adaptation appeared in 44% (14/32) and used accumulating behavioral, contextual, physiological, or conversational information to modify intervention timing, content, allocation, or interaction over time [14,24-26,37,43,44,46,48,50,54,56,59,64]. Some interventions combined both approaches [43,44,64], illustrating that static and dynamic personalization are related but not interchangeable. Personalization did not uniformly improve engagement. In Rouvere et al [52], additional tailoring questions increased disengagement, while the effects of tailoring differed according to the information used.

Conversational support appeared in 31% (10/32) of the studies [24-27,46,53,54,56,61,63]. These interventions used chatbot-based or guided dialogue to provide psychoeducation, emotional responsiveness, reflective interaction, or contextually relevant support. Conversational studies reported acceptability and user experience outcomes as well as relational dimensions such as therapeutic alliance and perceived bonding. Shoshani et al [46] found that perceived therapeutic alliance was associated with engagement and symptom improvement, while Xu et al [53] identified conversational control; input fit; caring responses; useful advice; colloquial language; and a private, nonjudgmental setting as factors shaping users’ bonds with mental health chatbots.

Social or co-design mechanisms appeared in 28% (9/32) of the studies [8,9,44,47,50,51,58,63,65]. These included participatory design, blended care, culturally responsive guidance, peer or team interaction, support network activation, and selective human coaching. Valentine et al [47] associated sustained engagement with a youth mental health platform with validation, normalization, belonging, and the practical application of therapeutic insights. Chow et al [50] reported that adding human coaching increased the number of days of app use among participants whose use was low during the first intervention week.

Context-aware prompting appeared in 22% (7/32) of the studies [14,37,43,44,48,56,64]. These systems adjusted intervention delivery according to behavioral, contextual, physiological, or user-reported information. Receptivity and intervention timing were examined directly in several context-aware studies. Kor et al [43] used accumulating receptivity data to inform mindfulness prompt delivery, Elmer et al [44] found that prompts triggered by a stated need for support were rated as better timed and more helpful than prompts triggered by distress thresholds, and Dai et al [48] reported higher adherence with context-aware delivery than fixed delivery.

Biometric-responsive adaptation appeared in 13% (4/32) of the studies [37,48,56,59]. Physiological or emotion recognition data generally functioned as inputs to contextual assessment, prompting, or adaptive delivery rather than as independent engagement mechanisms. Gamified reinforcement appeared in 16% (5/32) of the studies [12,51,55,59,65], while narrative immersion appeared in 6% (2/32) of the studies [51,59]. Their implementation and findings varied: an idle breathing game increased practice frequency [55]; dynamic tailoring increased narrative presence without improving measured immersion, while play time and completed sessions were lower in the tailored condition [59]; and team competition produced early affective benefits that weakened over time [65]. In Barry et al [51], a co-designed narrative game reduced depressive symptoms relative to an active control over 12 months but did not improve anxiety; mechanism-specific engagement or narrative resonance outcomes were not reported.

These frequencies should be interpreted cautiously. The review intentionally foregrounded adaptive, technologically mediated, and engagement-supportive interventions; therefore, the distribution within this selected corpus should not be treated as an estimate of mechanism prevalence across all DMHIs. In addition, 3 studies [45,60,62] were not assigned a mechanism (Table 3). Slade et al [60] examined sensing compliance, user control, and data collection burden; Ying et al [62] contributed a computational mental health assessment model; and Newby et al [45] used response-adaptive allocation across trial participants rather than an intervention-level engagement mechanism. These studies remained eligible because they contributed methodological or computational evidence relevant to engagement and adaptive intervention design.

Figure 2 illustrates the co-occurrence of engagement-supportive mechanisms across the included studies. Dynamic adaptation most frequently co-occurred with context-aware prompting in 7 studies [14,37,43,44,48,56,64] and with conversational support in 6 studies [24-26,46,54,56]. These combinations reflected systems in which behavioral, contextual, or conversational information informed decisions about the timing, content, or form of support.

Biometric-responsive adaptation co-occurred with dynamic delivery in all 4 studies in which it appeared [37,48,56,59] and with context-aware prompting in 3 studies [37,48,56]. This pattern reflects the role of physiological or emotion recognition data as sensing inputs within broader adaptive systems. Context-aware prompting similarly tended to operate alongside personalization, recommendation logic, conversational support, or reinforcement learning rather than as an isolated feature.

Gamified and narrative mechanisms remained comparatively separate from AI-enabled conversational architectures, although Schwarz et al [59] combined dynamic narrative tailoring, biometric responsiveness, and gamified reinforcement. Social mechanisms spanned co-design, blended care, culturally guided support, team competition, support network activation, human coaching, and relational experiences of validation or belonging [8,9,44,47,50,51,58,63,65]. Overall, the co-occurrence patterns show that engagement support was commonly implemented through interacting sensing, decision, and delivery components. These descriptive relationships do not establish that combining more mechanisms necessarily produced stronger engagement.

AI-enabled systems were represented across controlled trials, qualitative studies, simulation and modeling research, and adaptive system design. AI served several distinguishable functions: direct conversational interaction [24-27,46,53,61,63]; within-person adaptation of timing, content, or prompting [37,43,44,48,56,64]; response-adaptive allocation across trial participants [45]; and back-end assessment, simulation, or modeling [25,37,54,62]. Therefore, AI functioned both as an intervention component and as methodological infrastructure for developing or evaluating adaptive systems.

This distinction was particularly important for Newby et al [45]. Its contextual bandit modified allocation probabilities across sequential participants according to distress severity, while the brief interventions received by individual participants remained standardized. The study consequently contributed evidence about AI-enhanced trial methodology rather than dynamically personalized intervention delivery. By contrast, within-person adaptive studies used user-level information to change when or how support was delivered [37,43,44,48,56,64].

The type of evidence also differed across AI implementations. Controlled conversational studies primarily reported completion, acceptability, therapeutic alliance, safety, user experience, or symptom-related outcomes [24,26,46,61,63]. Qualitative and diary studies examined emotional responsiveness, trust, perceived support, conversational fit, and concerns about accuracy or authenticity [27,53,56]. Simulation, modeling, and reinforcement learning studies examined prompt construction, adaptive assessment, recommendation logic, sensing burden, or optimization before or alongside real-world deployment [25,37,45,54,62].

Across the corpus, AI and algorithmic adaptation served distinguishable conversational, within-person adaptive, response-adaptive, and modeling functions. Shoshani et al [46] connected conversational alliance with engagement; Dai et al [48] compared context-aware delivery with fixed delivery; and Newby et al [45] demonstrated the methodological distinction between adapting trial allocation and adapting intervention delivery. Nevertheless, most studies evaluated AI as part of a multicomponent system, and few isolated its independent contribution to sustained engagement.

Patterns of mechanism adoption varied across study designs. Static tailoring was prominent in controlled, culturally adapted, and real-world personalization studies [12,13,43,44,52,57,63,64]. Dynamic adaptation appeared more often in AI-enabled, recommender-based, sensing-supported, or contextually responsive systems [14,24-26,37,43,44,46,48,50,54,56,59,64]. Context-aware prompting was concentrated in JITAI, recommendation, and reinforcement learning studies [14,37,43,44,48,56,64], while biometric-responsive adaptation remained limited to sensing-enabled adaptive interventions [37,48,56,59].

Conversational support was represented in both controlled chatbot evaluations and qualitative studies of relational experience [24-27,46,53,54,56,61,63]. Gamified reinforcement appeared primarily in controlled or mixed methods evaluations [12,51,55,59,65], whereas social and co-design mechanisms were particularly visible in qualitative, participatory, implementation, and adaptive support studies [8,9,44,47,50,51,58,63,65].

These methodological patterns affected what could be learned about engagement. Controlled and microrandomized studies contributed behavioral indicators, receptivity data, and symptom-related outcomes; qualitative, diary-based, participatory, and reflective studies revealed trust, burden, cultural fit, power dynamics, and relational experience; and implementation, simulation, and modeling studies clarified scalability, decision logic, and optimization. The resulting evidence types were complementary but not directly interchangeable.

Intervention and observation periods varied substantially. Thirteen studies involved intervention or user study periods of approximately 4 weeks or less [13,24,26,43-45,52,53,55,56,59-61]. These included brief app and chatbot trials, a 4-week exergame, a 4-week gamified breathing deployment, an 18-day JITAI, a 22-day wearable LLM study, a 30-day sensing study, and a single-session personalization experiment. Seven studies lasted between 5 and 8 weeks [14,48,50,51,58,63,64], and 3 studies involved 9- to 12-week intervention or support periods [12,46,65]. Alvarez-Jimenez et al [8] was embedded in a longer implementation and tracked engagement for up to 24 weeks.

Eight studies did not report a conventional fixed duration or involved simulation, system design, computational modeling, autoethnography, qualitative accounts of previous use, or real-world platform engagement [9,25,27,37,47,54,57,62]. These studies should not be interpreted as fixed-duration interventions.

Follow-up or assessment beyond the immediate intervention or primary postassessment period was reported in 31% (10/32) of the studies [8,12-14,24,46,50,51,61,64]. Follow-up periods included 1 month [13,61], 8 weeks after a 4-week intervention [24], approximately 3 months [12,46], 6 months [64], engagement tracking for up to 24 weeks [8], and 6- or 12-month assessments in studies of adaptive coaching and narrative game-based support [50,51]. Most studies therefore provided limited evidence about sustained engagement beyond the immediate intervention period.

Tables 4 and 5 provide complementary summaries of the engagement-supportive mechanisms, AI-enabled systems, methodological approaches, and key findings represented across the included studies. Table 4 presents the engagement-supportive mechanisms identified in each study, the associated findings, and the resulting design implications. The mechanisms represent intervention characteristics, interaction processes, or adaptive functions intended to support engagement; they should not be treated as engagement outcomes themselves. Reported outcomes included behavioral indicators such as uptake, adherence, completion, retention, and receptivity; experiential and relational dimensions such as relevance, burden, immersion, trust, alliance, and emotional support; and qualitative or methodological findings relevant to intervention design.

Across Table 4, controlled studies most commonly contributed behavioral and symptom-related findings, while qualitative and participatory studies foregrounded the contextual, experiential, and relational dimensions of engagement. Adaptive timing, personalization, and conversational responsiveness were frequently associated with greater relevance, acceptability, adherence, or perceived support within individual study contexts [14,24,43,44,46,48,50,52,64]. These findings do not demonstrate that any single mechanism independently produced the reported outcome because most interventions combined several components. Gamified and narrative interventions also produced variable patterns across practice frequency, usage, immersion, affective outcomes, and longer-term symptoms [51,55,59,65].

Table 5 focuses specifically on AI-enabled systems and methodological contributions. It distinguishes studies in which AI functioned as enabling infrastructure (eg, assessment modeling, recommendation, reinforcement learning, response-adaptive allocation, or simulation) from studies in which conversational or adaptive systems directly shaped users’ experiences through dialogue, contextual delivery, emotional support, or personalized interaction. The table also shows how controlled, qualitative, implementation, and computational methods answered different questions concerning efficacy, feasibility, acceptability, safety, relational experience, system logic, and optimization.

Taken together, Tables 4 and 5 indicate that AI was most directly relevant to engagement when embedded within a broader engagement-supportive process than when presented as a standalone feature. They should not be interpreted as rankings of effectiveness because the studies used heterogeneous designs, outcomes, and observation periods and because most did not isolate the effects of individual mechanisms.

This narrative review synthesized engagement-supportive mechanisms across 32 heterogeneous studies [8,9,12-14,24-27,37,43-48,50-65]. Dynamic adaptation, conversational interaction, static tailoring, and social or co-design support were the most frequently represented mechanisms and were commonly combined with contextual or sensing components, although this pattern partly reflects the search focus.

Beyond these frequencies, 3 broader patterns concerned the fit between users’ needs, the system’s decisions, and the support delivered. Firstly, engagement-supportive mechanisms generally operated as layered systems in which sensing or user input informed a decision process and support was delivered through conversational, social, gamified, or other interfaces. Secondly, personalization separated into relatively static tailoring based on predefined characteristics and dynamic adaptation responsive to changing user information or context. Lastly, engagement was represented through behavioral, experiential, relational, and contextual evidence, alongside contributions to system optimization.

Across the context-aware and JITAI studies, the findings suggest that adaptive decision rules should account for users’ needs and receptivity rather than simply increasing the frequency or complexity of prompting. Adaptive recommendation, reinforcement learning, and wearable-triggered conversational systems illustrated potential to improve message relevance and reduce prompting burden, alongside limitations involving contextual interpretation and conversational depth [37,56,64]. Receptivity-informed timing and context-aware delivery were associated with acceptability or adherence [43,48], while prompts triggered by a stated support need were rated as better timed and more helpful than distress-triggered prompts [44]. However, the quasi-experimental timing study did not isolate the contribution of the timing strategy [43]. Considered together with the wider JITAI literature [29-31,66], these findings suggest that decision rules must represent both vulnerability and opportunity: a detectable stress state does not necessarily mean that a person is able or willing to act.

Biometric responsiveness was less common and generally functioned as enabling infrastructure rather than as a standalone engagement strategy. Physiological, behavioral, or affective signals contributed to stress detection, contextual timing, or adaptive delivery [37,48,56,59], but frequent mobile sensing could create fatigue, and users valued control over data collection [60]. Its potential to make support more contextually relevant while reducing unnecessary prompts must therefore be balanced against device and monitoring burden, privacy, and inference accuracy.

Personalization showed potential benefits alongside interaction costs. Tailored, culturally adapted, and guided interventions were associated with greater perceived relevance or acceptability, reduced stress, or improved emotional regulation [13,14,57,63,64], although digital augmentation did not consistently overcome engagement difficulties [12]. Demographic tailoring reduced disengagement, whereas tailoring based on perceived need or intended next steps did not, and collecting extra tailoring data increased disengagement [52]. Complementary evidence outside the stress-focused corpus showed that interface optimization and reminders improved engagement, whereas adaptive human support improved satisfaction but not engagement [49]. Collectively, these findings suggest that the value of personalization depends on both the information used and the effort required to provide it, qualifying the assumption that more complex adaptation necessarily outperforms simpler improvements in interface design, guidance, or reminders.

Conversational systems provided another pathway to engagement by creating responsive and potentially relational forms of interaction. Alongside favorable acceptability or mental health outcomes [24,63], higher rates of user-confirmed accurate reflections with generative responses [26] and short-term gains in self-care intentions and mental health literacy with topic-based chatbots [61] support assessing conversational responsiveness and self-care learning separately. User experience and alliance findings linked relational value to immediacy and availability [27]; conversational control; input suited to users’ preferred expression; caring responses; and a private, nonjudgmental setting [53], alongside cultural adaptation and human guidance [63]. Perceived alliance was associated with engagement, including an indirect path through engagement to symptom change [46]. These associative and qualitative findings do not establish a causal relational mechanism, and concerns about accuracy and authenticity [27] caution against equating rapport with clinical reliability or safety.

Relational engagement was not limited to interaction with chatbots. Clinical or peer support, social motivation, validation, normalization, belonging, and the practical application of therapeutic insights contributed to engagement [8,47,63,65]. Home-based delivery offered flexibility, whereas center-based programs provided social motivation [58], suggesting that delivery choices should account for practical circumstances and preferences for social interaction. Limited support availability and time constraints [44], cognitive overload and low motivation [47], and increased number of days of app use following targeted coaching among participants with low initial use [50] suggest that human support should address the particular barriers preventing participation. Relational issues also arose during intervention development: co-design revealed how power and competing priorities influenced whose perspectives informed the intervention [9].

Gamification produced variable findings depending on the form and purpose of the game elements. Gamified breathing increased practice frequency [55], dynamic tailoring increased narrative presence without improving measured immersion, while play time and completed sessions were lower in the tailored condition [59], and team competition produced early affective benefits that weakened over time [65]. A co-designed narrative game was associated with lower depressive symptoms than an active control at 12 months, but anxiety did not improve, and engagement quality or narrative resonance were not measured [51]. Considered alongside work on motivationally aligned gamification [19,20], these contrasts indicate that practice, narrative presence, immersion, and affect require separate assessment and that rewards or greater usage alone do not establish meaningful or sustained engagement [7]. The limited narrative evidence [51,59] does not yet establish whether immersion produces sustained use or whether immersion or sustained use mediates mental health outcomes.

AI-related evidence likewise differed in function and maturity. LLM prompting, architectural benchmarking, fuzzy logic assessment, and reinforcement learning studies contributed evidence about system design, assessment, and optimization [25,37,54,62] but did not necessarily demonstrate sustained engagement in everyday use. Response-adaptive allocation concerned research allocation across participants and not within-person adaptation of the standardized interventions they received [45]. The corpus consequently provides evidence of feasibility and experiential potential but limited evidence that AI itself independently improves sustained engagement. AI should therefore be understood according to its functional role within the intervention or research design rather than being treated as a single engagement mechanism.

These findings align with broader digital health and HCI literature showing that engagement cannot be reduced to usage metrics alone [2,5-7,33-36]. Behavioral activity, emotional investment, perceived usefulness, and relational value may diverge across users and contexts. Frameworks emphasize interactions among motivational, immersive, and social processes [7,38-40], but the corpus provided limited evidence on which combinations support meaningful or sustained engagement. Emotional safety, therapeutic credibility, burden, privacy, and contextual inference [28,67,68] may help explain why advances in sensing, adaptive delivery, and AI-enabled support have not been matched by evaluation of users’ experiences and continued participation over time.

SDT helps connect these mechanism-level findings without treating technology as intrinsically motivating. Autonomy may be supported when culturally or individually tailored content increases relevance, users retain meaningful control, and prompts fit their preferences or circumstances [12-14,43,44,52,57,64]. It may be undermined when additional questions, opaque inference, excessive sensing, or poorly timed notifications increase burden [43,44,52,60]. Competence may be supported through achievable practice, feedback, graduated challenge, and opportunities to apply skills, including well-integrated gamified or narrative experiences [51,55,59,65]. Relatedness is reflected in conversational alliance, nonjudgmental support, belonging, co-design, blended care, and selective human coaching [8,9,24,26,27,44,46,47,50,53,58,63]. These are theory-informed alignments, not evidence that SDT was directly tested across the corpus. Their explanatory value lies in distinguishing mechanisms that merely increase exposure from designs that may support volitional, competent, and socially connected participation.

Designers should specify what information an adaptive system collects, how that information affects its decisions, and how the resulting support is delivered. Receptivity and declared support needs should be evaluated alongside stress detection, and sensing and tailoring burden should be minimized. Users should be able to understand, control, or contest consequential recommendations. For conversational systems, this also requires attention to accuracy, privacy, explainability, perceived authenticity, emotional dependency, and safe escalation to human support [22,23,26,27,33,34,46,53]. Independent replication is important where intervention developers or commercially affiliated researchers participate in evaluations.

Evaluation should combine behavioral telemetry with experiential, relational, and implementation outcomes matched to the intended engagement pathway. Relevant outcomes include alliance in conversational systems, belonging, validation, and practical use in social platforms and repetition, competence, immersion, or social motivation in gamified systems. Mixed methods designs can connect use or sensing compliance with users’ accounts of burden, relevance, and perceived support [44,47,55,57,60,69]. Qualitative, participatory, autoethnographic, and duoethnographic approaches can reveal power, emotional friction, cultural fit, and trust that usage data may not capture [9,14,27,56,58,70-72]. Future studies should measure psychological need support and motivational quality explicitly if SDT is presented as a design rationale.

Controlled and longitudinal studies can estimate behavioral and clinical effects, with microrandomized trials examining proximal timing effects and factorial trials testing component effects. Sequential multiple assignment randomized trials address adaptive escalation, while response-adaptive designs concern allocation efficiency. Multiphase evaluation can connect modeling, feasibility testing, and controlled comparison with real-world implementation [73], where Proctor taxonomy and the RE-AIM (Reach, Effectiveness, Adoption, Implementation, and Maintenance) framework address acceptability, appropriateness, feasibility, adoption, and maintenance [74,75]. Consistent reporting of engagement definitions, intervention exposure, and adaptive decision rules, together with longer follow-up in routine settings, would improve comparison without reducing engagement to a single universal metric.

This is a narrative review informed by structured procedures and not an exhaustive systematic review or meta-analysis. The searches were restricted to English-language records and selected databases, and the mechanism-focused terminology may have missed relevant interventions. Excluding app store–only evaluations, commercial deployments, and other gray literature may also underrepresent engagement practices used in widely adopted consumer DMHIs [2,11].

Study designs, populations, and engagement measures were heterogeneous, and several interventions combined mechanisms. Mechanism counts should therefore not be interpreted as comparative effectiveness estimates. The narrative synthesis also does not establish that engagement mechanisms caused the reported behavioral or clinical outcomes when those mechanisms were not experimentally isolated.

Most intervention periods were brief; only 10 studies reported assessment beyond the immediate intervention or primary postassessment period, and the evidence remains concentrated in high-income settings. These limitations constrain conclusions about sustained engagement and generalizability. Although the studies from Jordan [57,63] contribute important culturally adapted evidence, the limited representation of lower-resource settings means that assumptions about connectivity, digital literacy, intervention burden, social support, and culturally appropriate engagement cannot be generalized confidently.

Two technical preprints were retained as narrow exceptions because of their emerging architectural contributions [25,37], but they were not treated as clinical effectiveness evidence. The insomnia trial [49] and 4 reviews and meta-analyses [33-36] were used only as external support and were not included in the study tables, mechanism counts, or denominators.

Engagement in stress-focused DMHIs can be understood as the fit among users’ needs and opportunities, the system’s sensing and decision logic, and the way support is delivered. Adaptive timing, conversational interaction, relational support, motivational design, and meaningful tailoring can contribute to that fit, but none should be assumed effective merely because it is technically sophisticated. The distinction between static tailoring and dynamic adaptation and between behavioral use and experiential or relational engagement should remain explicit in both design and evaluation.

The broader requirement is to build systems that are useful without being intrusive, responsive without being opaque, and scalable without weakening human oversight. Inclusive co-design, independent evaluation, and longitudinal mixed methods evidence remain priorities for translating technological promise into meaningful, sustained, and ethically responsible engagement in stress-management DMHIs.