Modeling Phone Engagement with Continuous Time Hidden Markov Models

• Activity Number: 443 - Latent Variables, Causal Inference, Machine Learning and Other Topics in Mental Health Statistics
• Type: Contributed
• Date/Time: Thursday, August 12, 2021 : 4:00 PM to 5:50 PM
• Sponsor: Mental Health Statistics Section
• Author(s): Benny Ren*
• Companies: University of Pennsylvania
• Keywords: Mobile Health; Hidden Markov Models; Latent Variable

Abstract

Hidden Markov Models (HMM’s) are well established tools that identify latent states in time series data. We propose a continuous time extension, with regression and random intercepts, to the standard HMM to model longitudinally collected mobile health data. We present an easy to implement expectation-maximization (EM) procedure to estimate the model parameters. In addition, we outline the potential clinical significance of each model parameter in a real-world dataset. We fitted our HMM with data from psychiatric study participants, in order to analyze individual activity profiles. Using accelerometer measurements as the main predictor and phone use as the outcome, our HMM was able to identify phone engagement as latent states for each hour of the day. Furthermore, by assigning hour of the day as the random intercepts, each intercept models the effect of hour of the day on phone engagement. As a result, we hypothesize that the variance of random intercepts is positively correlated with a structured routine and cross reference our model output with known clinical outcomes.