Medical treatments are traditionally personalized in a manual process by healthcare practitioners. Personalization starts with a one-size-fits-all treatment adjusted for each patient in a lengthy trial and error process. Unfortunately, this process can result in unnecessary treatment, exposure to side effects, and patient loss of interest due to treatment ineffectiveness. Mobile health (mHealth) researchers have investigated ways to decrease ineffective treatment exposure by personalizing health interventions using Artificial Intelligence (AI). AI methods like contextual-bandits are often used for personalizing content (i.e., which health advice to provide) with promising results. However, content personalization approaches alone are underpowered by lacking personalization of time of treatment: an active component that delivers health advice in the form of alerts or reminders at appropriate times (e.g., time, location, and activity). State of the art work has shown that reminders alone can increase treatment adherence but have not resulted in behavior change yet.

In this thesis, I developed and tested a method for personalizing mobile health interventions' content and timing of treatment. I tested this approach in a real-world deployment (n=30, spring 2019) of a behavioral sleep intervention. I found that this personalization approach improved sleep duration, motivation to improve sleep-related behaviors, and adherence to sleep advice. In addition, I discovered that contextual factors and participant intrinsic characteristics have a significant effect on adherence to treatment. Building on these results, I implemented a machine learning classifier that predicts next-day adherence to treatment with promising performance.

Following up on the results from the sleep intervention, I deployed a larger (n=80) to investigate further the marginal effects of personalization of content and treatment timing. The intervention was deployed sleep days before the beginning of the 2020 pandemic. This intervention did not result in behavior change. In this part of my thesis, I investigate this 2020 deployment and the specific causes of the null intervention results. I compare the behaviors of the participants in the 2020 and 2019 studies using behavioral logs, phone usage, and sensor streams and surveys. I found that a lack of motivation caused by anxiety and stress induced by the pandemic and a drastic change in phone use and daily routines were the most likely reasons for the null intervention results. I close this thesis with recommendations on preparing for abrupt changes in their daily behavior and how they interact with computing devices used for intervention purposes.

In summary, this thesis contributes 1) A novel, effective, and sample efficient approach for the simultaneous personalization of content and timing of treatment using AI, sensors, and human feedback, 2) A deployment and test of a system using the personalization method mentioned above, 3) Findings on contributing factors that change adherence to treatment in the context of a behavioral intervention, 4) A machine learning classifier for the prediction of intraday adherence and 5) The development of a framework for understanding contributing factors that lead to null results during a pandemic and may generalize to pandemic-life events.

131 pages

Thesis Committee:
Anind K. Dey (Co-Chair)(University of Washinton)
Mayank Goel (Co-Chair) (CMU)
Robert Kraut
Carissa Low (University of Pittsburgh)
Tanzeem Choudhury (Cornell University)

Jodi Forlizzi, Head, Human-Computer Interaction Institute
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu