“I Don’t Know Why I Should Use This App”: Holistic Analysis on User Engagement Challenges in Mobile Mental Health

The paper addresses the challenge of sustaining engagement in Mental Health Apps (MHAs), which is essential for effective long-term symptom management and relapse prevention. Engagement is conceptualized as a dynamic, multidimensional construct comprising behavioral, cognitive, and emotional components that evolve over time. Prior reviews identified barriers but often failed to link engagement challenges to specific app functions or propose solutions from both user- and system-centered perspectives. Motivated by this gap, the authors conduct a scoping review to systematically analyze user engagement factors in MHAs, define engagement based on established frameworks, and examine personalization via explicit and implicit methods. Recognizing the lack of standardized engagement metrics, the study surfaces recurring themes from reported findings rather than comparing engagement levels quantitatively. Research questions: RQ1 identifies barriers to sustained engagement in user- and system-centered approaches; RQ2 explores research directions to address these barriers, including advanced AI (LLMs) to improve personalization and interaction design.

The review synthesizes prior work on barriers to MHA engagement across five domains: (1) Technical and usability challenges (e.g., crashes, slow performance, poor onboarding, complex interfaces); (2) Personalization and content issues (static, generic content; limited adaptive interfaces; lack of culturally/contextually relevant material; repetitive text, limited multimedia); (3) Emotional, psychological, and social barriers (insufficient emotional support, lack of professional/peer interaction, poor crisis responsiveness); (4) Privacy, security, and trust concerns (opaque policies, default public settings, limited support, excessive ads, weak long-term strategies); and (5) Condition-specific challenges (depression-related low motivation, anxiety-related interface stress, schizophrenic usage decline post-initial engagement, irregular usage during manic or intense negative affect states). Prior reviews did not connect these barriers to specific MHA functions nor discuss comprehensive, user- and system-centered solutions or AI-integrated strategies.

Scoping review adhering to PRISMA-ScR and PRISMA 2020 guidelines. Inclusion criteria: peer-reviewed, English, accessible; empirical studies with human participants and sustained use; mobile app-based mental health interventions. Search strategy used keyword combinations for mobile technology and mental health across Google Scholar, ACM CHI/CSCW/IMWUT, ACM Digital Library, IEEE Xplore, Scopus (Aug 2023–May 2024), plus reference list reviews. Identification: 1,371 database records + 25 cross-referenced = 1,396; duplicates removed (n=129) → 1,267 records. Screening: fundamental filters excluded 106 not peer-reviewed, 14 non-English, 8 not accessible → 1,139 eligible. Empirical & sustained filter excluded 885 not empirical and 41 no sustained user study → 213 eligible. Inclusion: excluded 102 non-app-based interventions → 111 studies included. Coding: thematic extraction of specific functions from study content; bottom-up grouping into system functions and intervention approaches; inter-rater agreement (Fleiss’ Kappa 0.75–0.84, mean 0.80 for function derivation; 0.77–0.83, mean 0.79 for engagement-promoting features; 0.77–0.85, mean 0.81 for engagement barriers). A Sankey diagram and tables summarize mappings; Krippendorff’s alpha of .86 for inclusion screening; consensus and third coder used for disagreements.

• Sample characteristics: 111 empirical field studies; 80.36% from 2019–2024; 15 studies (13.39%) with AI (ML/LLMs). Target conditions: stress (n=40), depression (n=29), anxiety (n=17), comorbid anxiety/depression (n=13), PTSD (n=6), others (n=7). Participant sizes: excluding 4 open-access apps (619–13,421, median=4,438), remaining studies had 7–1,056 participants (median=64, mean=99, SD=133). Intervention duration: 5–390 days (median=42, mean≈68, SD=72). - Intervention approaches: User-centered (therapy content experience and self-assessment) vs. System-centered (system feedback and passive sensing). Counts: User-centered 85 therapy content (74%) and 30 self-assessment (26%); System-centered observed in 76 system feedback (95%) and 4 passive sensing (5%). - Specific functions: Psychoeducation (47), Mindfulness practices (29), Task management (8), Self-monitoring (12), Self-reports (17); System feedback functions: Notifications/reminders (40), Personalized feedback (24), Gamified reward systems (12); Passive sensing (4). - Engagement-promoting features: Integrated self-management (e.g., pairing stress literacy with stress trigger/response logs; mindfulness sensory focus with goal setting) and complex feature combinations for holistic engagement (emotional, physical, social). System-centered: personalized mental health support via data-driven features (tailored content, stress management, coaching; passive sensing) and adaptive features (progress tracking, stress alerts, gamified rewards). - Engagement challenges (n=4): Persistent participation fatigue; Less engaging functions; Privacy concerns; Less adaptive functions. Subthemes: • Fatigue: burden of continuous tasks; time constraints. • Less engaging: low attention; monotonous design; chatbot limitations; gamification competing with entertainment. • Privacy: surveillance apprehensions (image/log collection discomfort); loss of control (inability to review/delete passively captured data). • Less adaptive: frequent/uncontextualized notifications; evolving needs unmet (e.g., fixed module unlock periods); digital literacy barriers; diminished gamification effects when misaligned with mindfulness/intrinsic motivation. - Technical issues: 28 papers reported barriers limited to technical problems; 35 not explicit. Reported issues included smartwatch battery life, Bluetooth connectivity, app bugs/errors, WiFi/data access, chatbot technical errors.

Findings address RQ1 by mapping barriers to sustained engagement across user-centered (task burdens, attention/interest) and system-centered (privacy, adaptivity) approaches, showing how functional elements interact to create tensions that undermine intrinsic motivation. Addressing RQ2, the paper proposes balancing user/system strategies to satisfy autonomy, competence, and relatedness (Self-Determination Theory). Recommendations include: • Personalized emotional interaction design along adaptivity, continuity, and multimodality, enabled by LLM-based agents with memory (RAG, long-term memory) to detect subtle state changes, provide context-aware interventions, and foster ongoing relationships. • Collaborative design between HCI researchers and mental health professionals (MHPs) to digitize therapeutic content effectively and evaluate genuine efficacy given engagement constraints. • Flexible environmental adjustments to mitigate fatigue (task scheduling, caregiver-supported collaborative care). • Privacy and ethical enhancements: clear onboarding literacy, concise data transparency, federated learning, model unlearning; expert-guided safety for LLMs via RLHF, prompt constraints, risk screening and escalation. The holistic framing connects why barriers arise (structural imbalances) and how integrated design can rebuild intrinsic motivation, positioning MHAs as trusted partners rather than obligation-driven tools.

The study offers a comprehensive scoping review of 111 MHA studies, identifying four central engagement challenges: persistent participation fatigue, less engaging functions, privacy concerns, and less adaptive functions. It contributes a holistic framework that aligns user- and system-centered functions with barriers, and proposes five research directions: (1) interactions enhancing intrinsic motivation and emotional support; (2) collaborative design between HCI and MHPs; (3) flexible environmental adjustments; (4) privacy literacy and adaptive AI methods; (5) ethical safeguards for LLM integration. Emphasizing adaptivity, continuity, and multimodality as pillars of personalized emotional interactions, the paper highlights the potential of LLM-based agents to promote sustained engagement and effectiveness in future MHAs.

As a literature-based scoping review, findings may not generalize uniformly across all users or contexts. Individual differences (preferences, conditions, culture, prior experiences, digital literacy) can produce heterogeneous engagement patterns. Reported barriers (e.g., self-report fatigue, limited contextual personalization) represent recurring phenomena rather than universal truths. The review did not quantitatively compare engagement levels due to non-standardized metrics. Future work should incorporate personalized strategies reflecting latent user variables, context-aware interventions, and long-term motivational supports, with longitudinal designs to validate clinical efficacy and user satisfaction.