Data-Driven Mobile Health: System Identification and Hybrid Model Predictive Control to Deliver Personalized Physical Activity Interventions

Abstract

The integration of control systems principles in behavioral medicine involves developing interventions that can be personalized to foster healthy behaviors, such as meaningful and consistent engagement in physical activity. In this paper, system identification and hybrid model predictive control are applied to design individualized behavioral interventions using the control optimization trial (COT) framework. The paper details the multiple stages of a COT, from experimental design in system identification to controller implementation, and demonstrates its efficacy using participant data from Just Walk, an intervention that promotes walking behavior in sedentary adults. Mixed partitioning of estimation and validation data is applied to estimate ARX models for an illustrative participant, selecting the model with the best performance over a weighted norm balancing predictive ability with overall data fit. This model serves as the internal model in a three-degree-of-freedom Kalman filter-based Hybrid Model Predictive Controller (3DoF-KF HMPC) that provides “ambitious but doable” goals for initiation and maintenance phases of the physical activity intervention. Performance and robustness in a closed-loop setting are evaluated via both nominal and Monte Carlo simulation; the latter confirms the inherent robustness properties of the controller under plant-model mismatch. These results serve as proof of concept for the COT approach, which is currently being evaluated with human participants in the clinical trial YourMove (R01CA244777, NCT05598996).