Bayesian Structural Time Series for Mobile Health and Sensor Data: A Flexible Modeling Framework for Evaluating Interventions

The development of mobile health technology has the potential to contribute greatly to personalized medicine. Wearable sensors can assist with determining the proper treatment plans for individuals, provide quantitative information to physicians, or give individuals an objective measurement of their health. However, though treatments and interventions have become more targeted and specific, measuring the causal impact of these actions require more careful considerations of complex covariate structures as well as temporal and spatial properties of the data. Thus, emerging data from sensors and wearables in the near future will make use of and require complex models. Here, we describe a general statistical framework for sensor and wearable data that applies a Bayesian structural time series model to analyze and understand various behavior and health data collected in different environments. We show the wide applicability of this modelling framework, and how it corrects for covariates and biases to provide accurate assessments of intervention. Furthermore, it allows for a time dependent confidence interval of impact through its use of Bayesian estimation. We give three main examples, physical sensor data, environmental air sensors and longitudinal behavioral data to show the effect of various interventions through parameter estimation and comparison in pre- and post- intervention periods. The Bayesian structural time series model shows robust performance in a wide variety of tasks, further supporting its applicability to current and future mobile health and sensor data types.