Dae Lee

Medical errors have been cited as the third leading cause of death in the United States in 2013. Failure to rescue (FTR) is a subtype of medical errors and refers to the loss of an opportunity to save a patient’s life after the development of one or more preventable and treatable complications. Focusing on detecting early signs of deterioration may therefore provide opportunities to prevent and/or treat an illness in a timely manner, which may in turn reduce the number of FTR cases. When implementing a data-driven model to predict the risk of potential FTR onsets in a supervised setting, gold standard information for the target FTR onset is often not directly retrievable in electronic health records (EHR) so that it requires to manually annotate clinical observations with corresponding labels. This method acts as a bottleneck to scalability and the full utilization of the clinical observations available in EHRs for model training. In this dissertation, I propose a machine learning framework that can be used to derive a risk prediction model using proxy events of the disease of interest, the administration of relevant clinical interventions, as a noisy label via a distant supervision approach. Moreover, this study evaluated the effects of considering the temporal progression of FTR risk estimates calculated using myopic evidence. Lastly, a case study is presented to demonstrate that the proposed prediction models can be deployed to quantify the adverse effects of clinical interventions with regard to the target disease of interest. This dissertation demonstrates 1) the feasibility of using proxy events of the target disease as a label for supervised model training, 2) the performance improvement when temporal progression is considered in the risk prediction model design, and 3) the applicability of the proposed risk prediction model to quantify the adverse effects of clinical interventions regarding the target disease. Suggestions are also provided on how the proposed model could be further improved by integrating experts’ knowledge with the proposed framework.