Computational applications to hospital epidemiology

Healthcare associated infections are a considerable burden to the health care system. The affected patients have their prognosis worsened and demand more resources from hospitals. Furthermore, the bacteria causing these infections are becoming increasingly resistant to antibiotics while also becoming more deadly and contagious. Contributing with knowledge for stopping these infections is, therefore, important. This thesis reports on two projects centered on data collected at the University of Iowa Hospital and Clinics. The first project consisted in analyzing data collected by sensors that reported the location and hand washing behavior of health care workers. After extracting meaning from these radio signals, I studied two socially and epidemiologically relevant tasks: the inference of contact networks, which can be used to study the spread of infections in the hospital, and the study of associations between social pressure and hand washing, learning that effectively workers in proximity to others wash their hands more, but also that not all workers are as influential. In the second project, I developed a data mining method for analyzing medical records aimed at tackling the problems of class imbalance and high dimensionality, and applied it to predicting Clostridium Difficile infection. The learnt models performed better than the state of the art and even improved prediction as the onset of symptoms approached. The main contribution, however, was in the information discovered: certain events in certain orders increased the risk of developing the infection, suggesting that reversing these orders could improve prognosis.