| Description |
Temporal reasoning denotes the modeling of causal relationships between different variables across different instances of time, and the prediction of future events or the explanation of past events. Temporal reasoning helps in modeling and understanding interactions between human pathophysiological processes, and in predicting future outcomes such as response to treatment or complications. Dynamic Bayesian Networks (DBN) support modeling changes in patients' condition over time due to both diseases and treatments, using probabilistic relationships between different clinical variables, both within and across different points in time. We describe temporal reasoning and representation in general and DBN in particular, with special attention to DBN parameter learning and inference. We also describe temporal data preparation (aggregation, consolidation, and abstraction) techniques that are applicable to medical data that were used in our research. We describe and evaluate various data discretization methods that are applicable to medical data. Projeny, an opensource probabilistic temporal reasoning toolkit developed as part of this research, is also described. We apply these methods, techniques, and algorithms to two disease processes modeled as Dynamic Bayesian Networks. The first test case is hyperglycemia due to severe illness in patients treated in the Intensive Care Unit (ICU). We model the patients' serum glucose and insulin drip rates using Dynamic Bayesian Networks, and recommend insulin drip rates to maintain the patients' serum glucose within a normal range. The model's safety and efficacy are proven by comparing it to the current gold standard. The second test case is the early prediction of sepsis in the emergency department. Sepsis is an acute life threatening condition that requires timely diagnosis and treatment. We present various DBN models and data preparation techniques that detect sepsis with very high accuracy within two hours after the patients' admission to the emergency department. We also discuss factors affecting the computational tractability of the models and appropriate optimization techniques. In this dissertation, we present a guide to temporal reasoning, evaluation of various data preparation, discretization, learning and inference methods, proofs using two test cases using real clinical data, an open-source toolkit, and recommend methods and techniques for temporal reasoning in medicine. |
