Dynamic Modeling and System Identification of User Engagement in mHealth Interventions using a Bayesian Approach for...

Digital behavior change interventions (DBCIs) have been found to positively impact health behaviors and are becoming increasingly important as an emerging topic for control systems applications. However, their effectiveness is heavily dependent upon user engagement with both the digital tool (e.g., mHealth app, wearable activity tracker) and the behavior change intervention (e.g., exercise activity planning). In this paper, engagement refers to the unique interactions of a participant with either of these components resulting in digital traces (e.g., app page views). Furthermore, engagement in DBCIs will change over the course of the intervention in response to an individual's environment, context, and psychological state. Intensive data collection enables modeling engagement in DBCIs as a dynamical system using fluid analogies, and applying prediction-error methods from system identification to estimate models. Missingness represents both a fundamental and practical concern in this application domain. This work addresses missingness using a novel Bayesian imputation method applied to data from the HeartSteps II physical activity intervention study. The benefits of this approach include the ability to impute missing data points more accurately than traditional methods and quantify uncertainty resulting from imputation and data scarcity; the latter is essential to the implementation of robust closed-loop interventions. The methods presented in this work provide insights into critical factors that impact engagement behavior over time and in context, ultimately benefiting the development of digital behavior change interventions relying on control engineering approaches.