*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Developing Transferable Deep Models for Mobile Health
Committee:
Dr. James Rehg, CoC, Chair, Advisor
Dr. Omer Inan, ECE, Co-Advisor
Dr. Judy Hoffman, CoC
Dr. David Anderson, ECE
Dr. Santosh Kumar, U of Memphis
Dr. Inbal Nahum-Shani, U of Michigan
Abstract: Human behavior is one of the key facets of health. A major portion of health outcomes are attributed to behavioral risk factors such as smoking, drinking, unhealthy eating, etc. One of the main visions of mobile health (mHealth) is behavior modification through targeted interventions, using machine learning to predict risk of adverse behavior. mHealth studies are conducted in different contexts based on the research aim, querying a different set of emotions, using different types of sensors. This results in two challenges to using machine learning models across different studies: 1) Covariate-space shift: where the input-space differs across studies, 2) Domain shift: where the data distribution varies. It is critical design methods to overcome these two shift issues to utilize all the available data for predictive modeling. This thesis is divided into three parts. First, a method to overcome covariate-space shift in EMA through a novel valence representation is developed. This enables data pooling across multiple studies and helps improve predictive performance when compared to using data from individual studies. Second, a physiologically-inspired self-supervised learning based method for domain adaptation for pulsative physiological signals (ECG, PPG) is proposed. The self-supervised method outperforms standard domain adaptation methods. Further, it allows models to be adapted without requiring target domain data during training. Finally, the next important challenge to deploying machine learning models in health settings is model explainability which is explored in the third part of the thesis. The problem of bridging the gap between counterfactual based explainability methods and domain-experts is studied. A novel Variational Autoencoder based counterfactual generation method is proposed to generate plausible, relevant, and convincing explanations of ML model decisions.
