Probabilistic Models for Exploring, Predicting, and Influencing Health Trajectories
Schulam, Peter F
MetadataShow full item record
Over the past decade, healthcare systems around the world have transitioned from paper to electronic health records. The majority of healthcare systems today now host large, on-premise clusters that support an institution-wide network of computers deployed at the point of care. A stream of transactions pass through this network each minute, recording information about what medications a patient is receiving, what procedures they have had, and the results of hundreds of physical examinations and laboratory tests. There is increasing pressure to leverage these repositories of data as a means to improve patient outcomes, drive down costs, or both. To date, however, there is no clear answer on how to best do this. In this thesis, we study two important problems that can help to accomplish these goals: disease subtyping and disease trajectory prediction. In disease subtyping, the goal is to better understand complex, heterogeneous diseases by discovering patient populations with similar symptoms and disease expression. As we discover and refine subtypes, we can integrate them into clinical practice to improve management and can use them to motivate new hypothesis-driven research into the genetic and molecular underpinnings of the disease. In disease trajectory prediction, our goal is to forecast how severe a patient's disease will become in the future. Tools to make accurate forecasts have clear implications for clinical decision support, but they can also improve our process for validating new therapies through trial enrichment. We identify several characteristics of EHR data that make it to difficult to do subtyping and disease trajectory prediction. The key contribution of this thesis is a collection of novel probabilistic models that address these challenges and make it possible to successfully solve the subtyping and disease trajectory prediction problems using EHR data.