Optimizing Individualized and Adaptive mHealth Interventions via Control Systems Engineering Methods

点击翻译按钮获取中文摘要

Background: While reinforcement schedules are, in general, understood to be valuable for behavioral change,17- 19 maintenance,20-22 and adherence,23,24 there is still much more research needed to better understand the mechanisms of action (MoA) for these processes, such that more robust digital health interventions can incorporate them, systemically and at scale. The Parent R01 approach for continuously adapting support includes two states: the initiation state in which it relies on a continuous reinforcement schedule and a maintenance state in which it shifts to the use of a variable reinforcement schedule. Primary purpose: To systematically study the dynamic MoA of reinforcement schedules, on behavior change, via computational modeling and rigorous secondary analyses. Hypotheses: We hypothesize that 1) idiographic “black-box” dynamical models can identify key measured social cognitive theory (SCT) constructs like self-efficacy, that are important to understanding behavioral maintenance; 2) the use of grey-box/semi-physical models, incorporating reinforcement schedules into models, will explain a larger portion of variance, and 3) Model-on-Demand with Simultaneous Perturbation Stochastic Approximation (MoD-SPSA) approaches will identify optimal model structure, and adjustable parameters in the estimation method that better fit non-linear assumptions. Methods: We will build on our previously validated SCT dynamical model, which is foundational to the Parent R01, but with added incorporation of key insights about reinforcement schedules incorporated into the model structures. Then, using the data generated from first 100 participants of the Parent R01 clinical trial (see research strategy for our approach to ensure trial integrity), we will conduct “black-box” auto-regressive dynamical models, which does not incorporate prior domain knowledge, save SCT variable selection. Next, we will conduct “grey-box” modeling, which is much like a dynamical structural equation model in that it incorporates prior domain knowledge into the mathematical model. Increased percent variance explained of steps/day of the grey-box modeling is indicative of the added value of prior domain knowledge about the MoA of reinforcement learning, thus a robust test of this dynamic MoA. Finally, we will use MoD-SPSA25 as an aid for identifying optimal features, model structure, and adjustable parameters in the estimation method, to examine potential nonlinear interactions and relationships. Implications: This research has a number of synergistic benefits including: 1) it will generate rigorous scientific evidence for better understanding the MoA, reinforcement schedules, for behavioral maintenance; 2) it will produce key novel ways to operationalize, dynamically, different reinforcement schedules for fostering behavioral maintenance via digital health interventions; and 3) for the Parent R01, this research will allow the approaches and techniques to be refined with a specific emphasis on improving behavior maintenance. Thus, this supplement remains within the original scope of the parent award and will maximize the impact of the knowledge gained from it, especially for advancing the behavior maintenance with the behavioral theory testing.