Machine learning to refine theory: Participant engagement in mHealth interventions

Current technology allows for logging of specific actions within an app to track engagement, thus an opportunity to use strategies like machine learning emerges. Leveraging this depth of information enables refinement of our behavioral theories and constructs as it relates to mHealth interventions. Participant engagement is one construct that is assumed to lead to change in behavior during mHealth interventions. Engagement refers to a myriad of concepts within behavioral medicine research and generally refers to any action done to achieve the best health outcome from available services. Though many researchers propose engagement as an outcome or mediator, there is not a standardized operational definition for engagement when evaluating interventions. Despite mHealth apps requiring various types of engagement, little to no work has been done to conceptualize these types or their differential contribution to outcomes. At worst, trials do not report how often or to what extent an app is used when evaluating the effect on health outcomes. At best, few report how many days a participant opens the app, wears a device, or logs into an mHealth system. Machine learning was used in an ongoing mHealth multiple behavior change trial in college students over 6 months to identify actions that comprehensively define engagement such that specific actions or profiles may be linked to better health outcomes. Aligned with intended use of mHealth apps, engagement was divided into two primary categories: reviewing feedback and self-monitoring. We applied a head-tail classification to classify sub-types of engagement. Five sub-types of feedback engagement emerged: glances (< 7 seconds), check (7 – 17 seconds), brief review (17-38 seconds), detailed review (38-68 seconds), and deep review (> 68 seconds). For self-monitoring, three subtypes emerged: brief interact (3 - 29 sec), detailed interact (29-51 sec), and deep interact (> 52 seconds). This new operationalization enables identification of profiles that can be used to label participants and to predict outcomes. Future work will determine what type or combination of types of engagement predict important mechanisms and clinical outcomes in health behavior change research.