Learning Evolving Patient Risk Processes for C. Diﬔ Colonization

Predictions of adverse events during hospitalizations can be used in programs aimed at improving patient outcomes. A patient’s risk for adverse events may be biased by temporal processes influenced by diagnostic and therapeutic activities, as well as by the overall evolution of the patient’s pathophysiology over time. Representing and reasoning about temporal process promises to enhance the accuracy of inferences about risk. However, understanding temporal influences is challenging for a number of reasons, including the large number of variables, the large class imbalance, and the difficulty of defining ground truth for risk over time. We explore such challenges in the context of predicting an inpatient’s daily risk of becoming colonized with Clostridium Difficile. We present and evaluate different methods for extracting risk processes from medical records. These results highlight the benefit of including a temporal dimension when modeling patient risk.