Intensive Care Units (ICUs) are very complex settings where critically ill patients are constantly monitored. The analysis of ICU data is complicated by many factors, but one major concern simply comes from the size of the data. The Cox proportional hazards model has been applied to predict the onset of various types of adverse events (AEs) in patients admitted to ICUs like sepsis and septic shock. Our work aims to contribute to the research topic of predictive modeling using the Cox proportional hazards model by investigating whether a classifier built from the Cox model on "intelligently" re-sampled data can achieve similar performance to one trained on the whole, original data. By re-sampling we mean disregarding data points deemed uninformative with the goal of simplifying the training phase of the classifier built from the Cox model.