From Promise to Dropout: The Engagement Crisis Facing Digital Health

Digital health is rapidly moving from the margins of healthcare into mainstream treatment, promising lower delivery costs, continuous monitoring, personalized support and wider access to care. However, a major weakness threatens that promise: many patients simply stop using digital health applications before treatment is complete.

A new review, "Strategies to improve treatment adherence of digital health applications, rapid review and mixed-method analysis," published in AI & SOCIETY by Ulf Gerhardt and Cordula Kreuzenbeck, examines how serious the adherence problem has become and why patients disengage. Drawing on research published between 2017 and 2025, the authors combine quantitative evidence on dropout rates with qualitative analysis of the reasons behind them.

Across intervention groups representing 7,268 participants, 4,371 dropped out, producing a pooled dropout rate of 60.1%. Yet the median study-level dropout rate was only 18%, revealing enormous variation across apps, patient populations and clinical settings.

Digital-health adherence is not a single, uniform problem that can be solved with a better interface or more reminders. It is shaped by clinical context, user characteristics, product design, human support, infrastructure and health-system incentives. As digital therapeutics increasingly feed data into AI-enabled healthcare, failure to solve this problem could affect not only treatment outcomes but also the quality of the data on which future medical algorithms depend.

The 60% Dropout Figure Is a Warning

The study found unusually wide variation in dropout, with individual rates ranging from below 10% to around 90%. The mean study-level dropout rate was 32.1%, while the median was 18%, showing that several extreme values pulled the overall average upward.

One large real-world study alone had a particularly strong influence on the pooled result. When it was removed from the analysis, the pooled dropout rate fell from about 60% to 34.3%. The research does not establish that six out of ten users abandon every digital therapy. Instead, it reveals a sector in which performance is highly uneven.

The variation also complicates attempts to judge products using simple engagement benchmarks. An app providing daily rehabilitation exercises may require sustained repeated use, while another teaching coping techniques could achieve its purpose even if usage declines after patients have learned the required skills. The paper therefore cautions that lower usage does not always equal treatment failure.

It has implications for regulation and reimbursement. Health systems increasingly need to ask not only whether patients log in frequently, but whether they complete clinically meaningful tasks, achieve treatment goals and continue to benefit from the intervention.

The authors consequently advocate a more sophisticated concept of meaningful engagement: one focused on outcomes and goal-directed interaction rather than raw time spent using an application.

Patients Are Not the Only Ones Responsible

The review rejects the idea that dropout is primarily a matter of patient motivation. The researchers identify five broad groups of factors influencing adherence: patient-related factors, application-related factors, personal environmental conditions, technology-related barriers, and political or structural conditions.

Patient-level challenges include limited health or digital literacy, lack of time, low expectations of treatment, poor technological confidence and weak perceived benefit. But application design matters just as much. Poor usability, insufficient personalization, lack of interactive functions, limited motivational features and weak integration of personal or therapeutic support can all contribute to disengagement.

Social circumstances also influence whether patients stay engaged. Evidence reviewed in the paper links personal and professional support to better adherence, suggesting that fully automated treatment may not always be the ideal model. The finding challenges one of digital health's most attractive economic assumptions: that software can substitute for human contact at very low marginal cost.

In some cases, the technology may work best not as a replacement for clinicians, but as part of a hybrid care model in which physicians, therapists or multidisciplinary teams reinforce digital treatment. The study cites evidence that therapeutic alliance itself can influence adherence, strengthening the argument that digital medicine should be integrated into clinical relationships rather than deployed as a standalone product.

For developers, the message is that personalization, intuitive design, multimodal content, feedback, gamification and direct interaction are not cosmetic product features. They may determine whether clinically promising applications are actually used.

Digital Health Policy Must Move Beyond Approving Apps

The review exposes a broader governance problem. Digital therapies can receive regulatory approval and gain reimbursement pathways while still encountering obstacles that prevent sustained use.

Clinicians may lack knowledge about available applications. Prescribing and activation processes can be cumbersome. Healthcare professionals may have little time to explain digital treatment or insufficient financial incentives to support patients using it. Broadband gaps and technical problems can create further barriers, particularly outside well-connected urban environments.

For developing countries, these findings are especially relevant. Digital health is often promoted as a way to overcome shortages of specialists, expand access in remote areas and reduce pressure on health systems. Those benefits are real possibilities, but the study suggests that simply making an app available will not create equitable access.

Where connectivity is unreliable, smartphones are shared, digital literacy is uneven or health workers are already overstretched, adherence challenges could become even more severe. This is an analytical implication of the study rather than a directly tested finding, but it carries clear relevance for Global South digital-health strategies.

The policy lesson is that access and effective use are different objectives. Governments and insurers may therefore need to complement reimbursement with digital-literacy programs, patient onboarding, clinician training and technical support. Regulators could also require stronger post-market evidence on real-world adherence and effectiveness rather than concentrating primarily on safety and efficacy before approval.

The paper specifically argues that regulatory frameworks should pay greater attention to long-term adherence, usability, participatory development and real-world implementation. Such measures align closely with SDG 3 on health and well-being, but they also touch the broader development agenda around digital inclusion and inequality. Poorly implemented digital health may reproduce existing disparities instead of closing them.

Dropout Can Become a Data-Quality Problem

The adherence challenge becomes even more consequential as healthcare integrates artificial intelligence. Digital health applications can generate longitudinal streams of behavioral, physiological and clinical information. AI systems used for patient stratification, prediction, adaptive coaching or personalized recommendations depend on such data being sufficiently complete and representative.

If large or systematically different groups of patients disengage early, the resulting datasets may become biased. Older patients, people with lower education levels, digitally disadvantaged users or other vulnerable populations could be underrepresented in the data used to train or evaluate future systems.

The authors argue that dropout is not merely a clinical concern but an algorithmic one. The reliability and fairness of AI depend partly on who continues generating data and who disappears from the dataset.

The study itself has important limitations. It is a rapid review rather than a full systematic review or meta-analysis, and its methodology includes a restricted search, single-reviewer screening and no protocol registration. The underlying studies also vary substantially in size, design, disease area and definition of dropout, making direct comparison difficult. These constraints make the 60% pooled rate best understood as a warning signal rather than a universal benchmark.

That said, the broader finding is persuasive: digital medicine cannot be judged solely by what software is technically capable of doing. Its real value depends on whether patients can, will and need to use it under everyday conditions.