personalized messaging interventions
Physical inactivity is part of a constellation of lifestyle factors – with smoking and diet – that contribute to weight gain in early adulthood. Risk factors that compromise cardiovascular health begin to accumulate during the transition into adulthood. Interventions that prevent decreases in physical activity (PA) during this period can reduce long-term chronic disease risk. Text message interventions have shown a consistent positive effect on PA but efforts to increase those intervention effects via tailoring, targeting or personalizing have not realized their potential. New approaches have emerged for tailoring interventions based on treatment responses or contextual factors (e.g., stepped care, just-in-time adaptive interventions) but they apply a single decision rule uniformly for all participants. Behavior is complex and multiply determined so it is possible that treatment responses are idiosyncratic, necessitating personalized decision rules. Building on interest in precision medicine, we propose a method to develop personalized adaptive messaging interventions using intensive longitudinal data (from wearable sensors and momentary weather indices) and tools from control systems engineering (system identification and robust control synthesis). In preliminary work, we developed a computational model of physical activity responses to individual text messages. The greatest barrier to implementing that approach in interventions is that the computational models required for predictive modeling of PA dynamics have a high degree of uncertainty and are too complex to run efficiently on smartphones and other wearable devices. We propose to solve that problem by (1) developing a dynamical model of physical activity based on historical responses to messages, recent behavior, location-specific weather, and temporal features, and (2) evaluating the acceptability and feasibility of more versus less aggressive adaptation strategies for personalizing an intervention controller. To accomplish these aims, we will recruit young adults to participate in a PA messaging intervention and develop a computational model of responses to different messages under different conditions. A model-based controller will be developed to (a) optimize message timing, frequency, and content selection, and (b) achieve specified behavior change goals under varying conditions. We will then deploy that controller with an independent sample of young adults to determine how more versus less aggressive adaptation strategies over the next six months impact user experience. This study will contribute a model-based intervention controller and an acceptable adaptation strategy to use in a personalized adaptive messaging intervention for increasing PA. If successful, it will increase both PA and user engagement by selecting and timing messages to maximize effects and minimize burden. This approach can be applied to develop personalized interventions for other behaviors relevant for preventing weight gain, preserving cardiovascular health, and reducing chronic disease risk.