Title:  Domain adaptation for modeling and predicting health-related outcomes

Committee: 

Dr. Rehg, Advisor

Dr, Inan, Co-Advisor     

Dr. Anderson, Chair

Dr. Hoffman

Dr. Nahum-Shani

Dr. Kumar

Abstract: The objective of the proposed research is to use domain adaptation techniques to study and address the issue of domain shift across different mHealth field studies. This will involve focusing on two sub-tasks: 1) Measuring the domain gap between datasets, 2) Aligning the datasets. We propose to use the two main modalities in mHealth: Ecological Momentary Assessment (EMA), and wearable sensor datasets collected in five different smoking cessation field studies. The overall goal of this thesis is to develop effective methods for addressing domain gaps issues when constructing predictive models from longitudinal mHealth data. This proposal describes two predictive modeling efforts in support of this goal. The first effort is to develop a method to model spatiotemporal progression of glaucoma with a continuous time hidden Markov model. The second is to predict the risk of non-compliance to EMA prompts in an mHealth field study. The proposed work on domain adaptation targets the development of techniques quantifying and addressing the domain gaps that arise between different mHealth studies.