Show simple item record

dc.contributor.advisorLiu, Shan
dc.contributor.authorGong, Jue
dc.date.accessioned2019-08-14T22:34:42Z
dc.date.available2019-08-14T22:34:42Z
dc.date.submitted2019
dc.identifier.otherGong_washington_0250E_20350.pdf
dc.identifier.urihttp://hdl.handle.net/1773/44326
dc.descriptionThesis (Ph.D.)--University of Washington, 2019
dc.description.abstractFor many chronic diseases, an individual patient may experience a wide variety of progres- sion pathways. Personalized medicine needs tools to predict the trajectory of an individual patient’s disease progression, which can in turn enable clinicians to optimize the sequence of treatments. The objective of this thesis is to design artificial intelligence methods to sup- port clinicians in making smart treatment selections in chronic disease care. To achieve this objective, we develop algorithms optimized for an individual patients’ demographic profiles, past medical history, and response to current treatment by utilizing electronic health record (EHR) data of a large population. Developing a personalized treatment plan is a difficult sequential decision-making problem that seeks to improve overall health outcome, efficiency, and reduce unnecessary cost. One challenge is to understand the complex disease progression in a heterogeneous population. Another challenge is the lack of adaptive treatment strategies for diseases with partially observable health states. We develop innovative methodologies for personalized treatment selection to mitigate these challenges. First, we model the heterogeneity in disease trajectories of a population by detecting the subtypes of a chronic disease from longitudinal treatment data using an artificial neural network. Then we propose a framework called the partially observable collaborative model (POCM), to learn the individual disease progression model under various treatment options when the true health state is hidden to the decision maker. Next, utilizing the learned individual models, a personalized treatment plan can be derived by solving a partially observable Markov decision process (POMDP). We further extend this framework to mitigate the risk of reduced performance of POMDPs with uncertainty in transition dynamics by finding robust policies. Mental health is an understudied disease area that may greatly benefit from optimiza- tion in personalized medicine. Using simulated data informed by the Mental Health Research Network’s EHR, we apply the proposed methods to simulate the treatment of chronic de- pression. The contributions of this thesis include a novel framework for learning personalized disease progression model, a robust and adaptive treatment selection method, and an ap- plication on chronic depression treatment optimization. This thesis helps to advance the development of artificial intelligent decision support tools for chronic disease care.
dc.format.mimetypeapplication/pdf
dc.language.isoen_US
dc.rightsCC BY-NC-SA
dc.subjectdepression
dc.subjectmedical decision making
dc.subjectpartially observable Markov Decision Process
dc.subjectrobust optimization
dc.subjectIndustrial engineering
dc.subjectOperations research
dc.subject.otherIndustrial engineering
dc.titleOptimizing Personalized Treatment Selection for Partially Observable Chronic Conditions
dc.typeThesis
dc.embargo.termsOpen Access


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record