Oh, Wonsuk2021-05-172021-05-172020-09https://hdl.handle.net/11299/220133University of Minnesota Ph.D. dissertation. September 2020. Major: Health Informatics. Advisor: Gyorgy Simon. 1 computer file (PDF); xiii, 110 pages.Type 2 diabetes (T2D) is one of the fastest growing public health concerns in the United States. There were 30.3 million patients (9.4% of the US populations) suffering from diabetes in 2015. Diabetes, which is the seventh leading cause of death in the United States, is known to be a non-reversible (incurable) chronic disease, leading to severe complications, including chronic kidney disease, amputation, blindness, and various cardiac and vascular diseases. Early identification of patients at high risk is regarded as the most effective clinical tool to prevent or delay the development of diabetes, allowing patients to change their life style or to receive medication earlier. In turn, these interventions can help decrease the risk of diabetes by 30-60%. Many studies have been conducted aiming at the early identification of patients at high risk in the clinical settings. These studies typically only consider the patient's current state at the time of the assessment and do not fully utilize all available information such as patient's medical history. Past history is important. It has been shown that laboratory results and vital signs can differ between diabetic and non-diabetic patients as many as 15-20 years before the onset of diabetes. We have also shown in our study that the order in which patients develop diabetes-related comorbidities is predictive of their diabetes risk even after adjusting for the severity of the comorbidities. In this thesis, we develop multiple novel methods to discover T2D trajectories from Electronic Health Records (EHR). We define trajectory as an order of in which diseases developed. We aim to discover typical and atypical trajectories where typical trajectories represent predominant patterns of progressions and atypical trajectories refer to the rest of the trajectories. Revealing trajectories can allow us to divide patients into subpopulations that can uncover the underlying etiology of diabetes. More importantly, by assessing the risk correctly and by a better understanding of the heterogeneity of diabetes, we can provide better care. Since data collected from EHR poses several challenges to directly identify trajectories from EHR data, we devise four specific studies to address the challenges: First, we propose a new knowledge-driven representation for clinical data mining, second, we demonstrate a method for estimating the onset time of slow-onset diseases from intermittently observable laboratory results in the specific context of T2D, third, we present a method to infer trajectories, the sequence of comorbidities potentially leading up to a particular disease of interest, and finally, we propose a novel method to discover multiple trajectories from EHR data. The patterns we discovered from above four studies address a clinical issue, are clinically verifiable and are amenable to deployment in practice to improve the quality of individual patient care towards promoting public health in the United States.enThesis or Dissertation