CP 4: Control Systems Engineering for Counteracting Notification Fatigue: An Examination of Health Behavior Change

Collaborating Investigator:

Dr. Daniel E. Rivera (PI), Arizona State University

Funding Status:

R01LM013107

NIH/NLM

3/6/19 – 2/28/23

Associated with:

TR&D1

Publications
Significance
Approach & Push Pull Relationship

TR&D1: Discovery

BayesLDM: A Domain-Specific Language for Probabilistic Modeling of Longitudinal Data

Authors:

Karine Tung, Steven De La Torre, Mohamed El Mistiri, Rebecca Braga De Braganca, Eric Hekler, Misha Pavel, Daniel Rivera, Pedja Klasnja, Donna Spruijt-Metz, Benjamin M. Marlin

Publication Venue:

IEEE/ACM international conference on Connected Health: Applications, Systems and Engineering Technologies (CHASE)

Publication Date:

September 12, 2022

Keywords:

Bayesian inference, probabilistic programming, time series, missing data, Bayesian imputation, mobile health

View Full Paper

Related Projects:

CP 3, CP 4

We have developed a toolbox for the specification and estimation of mechanistic models in the dynamic bayesian network family. This toolbox focuses on making it easier to specify probabilistic dynamical models for time series data and to perform Bayesian inference and imputation in the specified model given incomplete data as input. The toolbox is referred to as BayesLDM. We have been working with members of CP3, CP4, and TR&D2 to develop offline data analysis and simulation models using this toolbox. We are also currently in discussions with members of CP4 to deploy the toolbox’s Bayesian imputation methods within a live controller optimization trial in the context of an adaptive walking intervention.

Abstract:

In this paper we present BayesLDM, a system for Bayesian longitudinal data modeling consisting of a high-level modeling language with specific features for modeling complex multivariate time series data coupled with a compiler that can produce optimized probabilistic program code for performing inference in the specified model. BayesLDM supports modeling of Bayesian network models with a specific focus on the efficient, declarative specification of dynamic Bayesian Networks (DBNs). The BayesLDM compiler combines a model specification with inspection of available data and outputs code for performing Bayesian inference for unknown model parameters while simultaneously handling missing data. These capabilities have the potential to significantly accelerate iterative modeling workflows in domains that involve the analysis of complex longitudinal data by abstracting away the process of producing computationally efficient probabilistic inference code. We describe the BayesLDM system components, evaluate the efficiency of representation and inference optimizations and provide an illustrative example of the application of the system to analyzing heterogeneous and partially observed mobile health data.

TL;DR:

We present a a toolbox for the specification and estimation of mechanistic models in the dynamic bayesian network family.

A wide range of technologies associated with mHealth, such as smartphones, wearables (e.g., Fitbit, Apple Watch), and medical devices use alerts to inspire actions of users. Potentially useful alerts run a risk of alert fatigue, whereby individuals ignore alerts over time. For example, physical activity interventions use alerts to inspire activity. These notifications work initially, but their efficacy diminishes over time. CP4 seeks to develop novel solutions for reducing alert fatigue by applying advanced analytical procedures inspired by dynamical systems modeling and control systems engineering. Location, digital traces, and other digital data facilitate inferencing states when a person would desire or need alerts (which, in CP4 are referred to as “just-in-time” states). For example, a suggestion to walk (e.g., SMS saying, “Want to go for a walk?”) may only produce the desired outcome when a person’s state (e.g., low stress) and context (e.g., no meetings, nice weather) align such that the person appreciates the notification (“receptivity”) and can act on it (“opportunity”). Estimating the likelihood that a given moment is a just-in-time state requires not only data but also an approach to manage the multivariate, dynamic, idiosyncratic, and multi-timescale nature of the problem. Further, just-in-time notifications cannot be viewed in a vacuum and, instead, are often part of a more long-term process, such as sustained engagement in a health behavior, thus making it a multi-timescale problem. The purpose of this work is to develop a just-in-time state estimation strategy and to stage a multi-timescale controller for walking as a concrete use-case of a control systems approach to counteract alert fatigue.

Efforts in CP4 consist of using Model-on-Demand estimation (an instance-based machine learning methodology that integrates local and global modeling and has been widely adopted within the control engineering community) as the framework for a novel, just-in-time (JIT) state estimator applicable to mHealth. Its development will be initially tested on secondary data from the HeartSteps intervention v1 (described in CP3) that will be followed by an in-the-field validation experiment involving a micro-randomized trial. The validated JIT estimator will be linked with a daily timescale model created in previous work and a multi-timescale model predictive controller to support decision-making about when to send notifications and personalized, adaptive daily step goals to support accumulation of walking bouts into meaningful change. The final control system will be evaluated in a small cohort study.

CP4 will initially focus on the analysis of the HeartSteps v1 data set using a Model-on-Demand estimation procedure for inferring just-in-time states. The primary proximal outcomes in HeartSteps v1 is 30-minute step count, as recorded by a Jawbone fitness tracker. This device has the property that if the device is worn and no steps are detected, no step count is recorded. Thus, both a zero-step count and missing step count (e.g., due to lack of connectivity, device not being worn, device not being on, etc.) appear as missing data in the HeartSteps v1 data set. Initial treatment efficacy analysis performed using the HeartSteps v1 data set was based on zero imputation of missing step values, discarding the uncertainty due to missingness. The methods developed under TR&D1 Aim 1 will be applied to the HeartSteps v1 data set to model and represent the uncertainty due to missing step count information. While CP3 will test the clinical utility of uncertainty model in interventions in its MRT, CP4 will assess the utility of the TR&D1 in dynamical systems modeling over multiply imputed data sets. The two teams will work together to improve the missing step count models and analyze the impact of missing step count modeling on end-to-end inference of just-in-time states. The powerful modeling framework developed in CP4 will provide a rigorous and well-specified setting in which to create and refine uncertainty models through multiple push-pull iterations between the two teams.