Dynamical Communication for Adaptive Mhealth Interventions: Evidence from the HeartSteps Pilot

Abstract

Background: Tailored health communication is effective at encouraging health behavior change; however tailoring to-date has focused on static and personal, as opposed to dynamic and contextual, characteristics. In the context of just-in-time adaptive interventions (JITAIs), which aim to improve response by adapting intervention delivery to immediate context, contextually-appropriate communication may lower barriers to treatment buy-in and adherence. Increasing availability of passive sensing tools makes such tailoring possible, but also significantly changes the scale of tailoring required. In lieu of a small number of expert-authored messages, JITAIs require libraries of messages tailored for all combinations of important contextual variables. Expert-guided crowdsourcing pipelines provide one solution for scaling tailored message creation. Purpose: To leverage crowdsourcing tools to create a library of contextually-tailored activity suggestions for HeartSteps, a mobile application for encouraging walking and reducing sedentary behavior; and to assess participant satisfaction with the tailored messages following a six-week pilot study. Methods: We tailored activity suggestions to four contextual variables, all passively collected by mobile phones: location (home/work/other), time of day (morning/lunch/ afternoon/evening/after dinner), weekday/weekend, and weather (outdoor/indoor/outdoor snowing), for two types of suggestions: active (5-20 minutes of walking); and sedentary (movement to disrupt sedentary behavior). This resulted in 180 combinations, or ‘buckets,’ of tailored messages. Given that HeartSteps delivered tailored suggestions up to five times a day, each bucket required multiple activity suggestions to avoid aggravating message repetition. Message creation was crowdsourced through two sets of “microtasks” on Amazon’s Mechanical Turk (MTurk): (1) Turkers were asked to write immediately actionable activity suggestions for a given persona and context (“John is about to ride the bus home from work; it’s nice outside.”); (2) Different Turkers rated these messages (scale 1-5). Both tasks included brief summaries of health communication ‘best practices.’ Messages rated 3.5 or higher on average were edited further and included in the HeartSteps message library. A six-week pilot study of HeartSteps delivered activity suggestions on the mobile phone up to five times a day to 40 volunteer participants. Tailoring effectiveness was assessed qualitatively through 60-minute participant exit interviews. Results: MTurkers produced several hundred tailored activity suggestions rated 3.5/5 or higher. Although further editing was required to ensure alignment with health communication expertise, these messages provided the bulk of 500+ activity suggestions available through HeartSteps. Participant feedback at exit interviews indicated overall satisfaction with activity suggestions, however: (1) in spite of the library, most participants experienced message fatigue after 2-3 weeks; (2) participants requested more specific contextual tailoring, especially for location; and (3) participants requested personal tailoring (for activities or communication preferences) alongside contextual tailoring. Future work aims to better specify an MTurk pipeline that creates and replenishes libraries of deeply-tailored messages Conclusions: Exponential growth in passive sensing capabilities increases the potential for provision of health communication tailored to real-time context. Current expert-led creation of tailored communication, however, may not scale to accommodate increasing dimensions. The HeartSteps pilot provided early evidence that context-tailored communication is essential for effective JITAIs, and crowdsourcing pipelines provide one way of constructing large-scale libraries of deeply-tailored messages.