Predicting and Preventing Nonadherence to Mobile Health Interventions
EMBARGOED UNTIL 2028-07-11
Loading...
Author / Producer
Date
2025
Publication Type
Doctoral Thesis
ETH Bibliography
yes
Citations
Altmetric
EMBARGOED UNTIL 2028-07-11
Data
Rights / License
Abstract
Noncommunicable diseases (NCDs) are the leading cause of death worldwide, posing significant challenges due to their prolonged progression, high treatment costs, and loss of productivity. Their growing prevalence, driven largely by modifiable risk factors, highlights the urgent need for more accessible, scalable, and cost-effective prevention and management strategies. With the widespread adoption of smartphones, mobile health (mHealth) interventions have emerged as a promising approach to support the prevention and management of NCDs by promoting behavior change, improving health outcomes, and reducing healthcare costs. Consequently, the number of mHealth applications now exceeds 350,000, with approximately 200 new apps released daily.
However, despite growing evidence and wider availability, mHealth interventions are subject to nonadherence, where users fail to use these technologies as intended, and in its most severe form – churn – discontinue use entirely before therapeutic benefits can be realized. Nonadherence undermines the effectiveness of mHealth interventions, making it a major concern for researchers and developers seeking strategies to prevent its impact. However, the existing literature lacks clear conceptualizations and measures for adherence to mHealth interventions, and factors influencing adherence are not yet fully understood. Identifying intervention-related factors associated with adherence can aid designers in optimizing interventions. Measurable data points associated with adherence could also be leveraged in combination with machine learning (ML) models to predict nonadherence and thus open up pathways for targeted strategies to prevent nonadherence. However, it remains unclear which methods and features are most effective for predicting nonadherence and how accurately nonadherence can be predicted. Likewise, a research gap remains in understanding which strategies effectively promote adherence to mHealth interventions and the extent to which nonadherence can be prevented.
To advance research on adherence to mHealth interventions, we conducted six studies that focus on three related research questions: (RQ1) What factors are associated with adherence to mHealth interventions? (RQ2) To what extent can we predict nonadherence to mHealth interventions? and (RQ3) To what extent can we prevent nonadherence to mHealth interventions?
Addressing RQ1, Article 1 presents a systematic review of 99 mHealth intervention studies, synthesizing and categorizing factors associated with adherence across eight health domains. Identified key factors promoting adherence include personalized app content, tailored push notification reminders, user-friendly and technically stable app design, complementary humandelivered support, and social and gamification features. While various sociodemographic user characteristics were frequently associated with adherence, these findings were inconsistent across different health domains. Negative factors included limited technical competence, low health literacy, lack of prior experience with mHealth apps, privacy concerns, low trust in healthcare professionals, negative app expectations, and time constraints. Additionally, personal onboarding was linked to higher adherence compared to remote onboarding. Quantified adherence scores, derived as the ratio of intended use to actual use, from included mHealth studies averaged 56.0% (SD = 24.4%). Most mHealth interventions were evaluated in research settings, with studies conducted in real-world environments recording lower adherence scores. Correspondingly, studies that offered higher monetary incentives tended to achieve better adherence.
In Article 2, we conducted a secondary analysis of a publicly available mHealth intervention (MySwissFoodPyramid, n = 19,805), which promotes healthy eating through dietary selfmonitoring and nutrition literacy delivered via a conversational agent. 66.8% of included users adhered to the intended use of the intervention of completing a food diary by tracking dietary intake over three days, demonstrating relatively good adherence despite the absence of monetary incentives or human support. The analyses further revealed significant associations between adherence and the day and time of installation, tutorial use, reminder use, and conversational agent choice. However, these factors only accounted for a small portion of the variance in adherence, calling for more robust measures to predict nonadherence in mHealth interventions.
Addressing RQ2, Article 3 presents a systematic review that investigates applied ML algorithms and features for predicting churn in 50 mobile application studies to establish a foundation on effective methodologies to predict user behavior with data collected from mobile apps. The findings demonstrate a wide variety of ML algorithms – from logistic regression to custom deep neural networks – that can be applied to accurately predict churn probabilities (mean AUC = 0.83). The review also underscores the importance of behavioral features related to users’ activity and progress within apps, while contextual features such as sociodemographic user attributes did not significantly contribute to the predictive performance of prediction models.
In Article 4, we developed ML models based on these insights to predict early churn – defined as users unsubscribing in the initial week – in a public mobile weight loss intervention on data from 310,845 event logs (WayBetter, n = 1,283). Among eight evaluated ML algorithms and three feature sets, a random forest model that only used daily login counts achieved the highest F1 score of 0.87 on day 7, correctly identifying 93% of churned users within the first week. Our results suggest that churn prediction models based on login data can detect a substantial proportion of users churning in the initial week – even before they disengage.
In Article 5, we developed ML models for the prediction of nonadherence relative to the intended use of two mHealth interventions, one for nonspecific and degenerative back pain over 90 days (Vivira, n = 8,372), and another for hypertension self-management over 186 days (Manoa, n = 6,674 users). Our models identified an average of 94% of non-adherent users between weeks 2 and 13 in Vivira (mean AUC = 0.95), defined as completing less than eight therapeutic exercises per week. In Manoa, our models correctly identified an average of 86% of nonadherent users between months 2 and 6 (mean AUC = 0.82), defined as completing less than one blood pressure measurement week per month. Additionally, models predicting churn (users’ last login within program duration) achieved mean AUCs of 0.87 for both apps, correctly identifying 84-86% of churned users. Hence, our results demonstrate that past behavioral app engagement data predicts future nonadherence to mHealth interventions, enabling targeted prevention strategies.
Addressing RQ3, Article 6 assessed the effects of human-delivered web-health coaching and monetary incentives on adherence to a mHealth intervention for young adults with type 1 diabetes based on a secondary analysis of a 2×2 factorial randomized trial (SweetGoals, n = 300). The results show that both incentives (B = 0.13, p < 0.001) and coaching (B = 0.06, p < 0.01) significantly increased adherence to the mHealth intervention, measured as the ratio of weekly completed feedback in the app over the 25-week study duration. Although adherence declined over time, the positive effects of both components remained stable. Furthermore, an antagonistic interaction was observed, where the combined effect of incentives and coaching was less favorable than the sum of their individual effects (B = -0.03, p < 0.05). From an economic standpoint, incentives were more cost-effective than coaching in increasing adherence.
In summary, this thesis advances the understanding of adherence to mHealth interventions by identifying factors associated with adherence, demonstrating the potential of ML models to predict nonadherence, and examining the effect of two complementary intervention components on promoting adherence. The identified factors can guide strategies to optimize existing health offerings. The insight that nonadherence can be predicted with good accuracy opens pathways for targeted prevention strategies before users fully disengage and highlights a unique and underutilized advantage of digital interventions – the capacity to transform continuously collected app engagement data into actionable insights. Finally, robust evidence on the positive effect of humandelivered web-based health coaching and monetary incentives highlights their promise as effective nonadherence prevention strategies. Collectively, these findings support the development of more adaptive and impactful mHealth interventions and their potential to mitigate the rising prevalence and economic burden of NCDs.
Permanent link
Publication status
published
External links
Editor
Contributors
Book title
Journal / series
Volume
Pages / Article No.
Publisher
ETH Zurich
Event
Edition / version
Methods
Software
Geographic location
Date collected
Date created
Subject
Machine Learning; Predictive Analysis; Human–computer interaction (HCI); Digital Health; Behavioral Analysis
Organisational unit
03681 - Fleisch, Elgar / Fleisch, Elgar