Abstract

Determining how severe the condition of a patient in the ICU can help doctors evaluate how to allocate their resources. This paper attempts to take the data from the first 48 hours of a patient’s time in the ICU and make the prediction of mortality. We use data from the PhysioNet 2012 challenge and the model was evaluated on the test set provided. We use the patient’s data and global averages to replace missing data and use an RDF based encoding to preprocess the data. Through randomized training sets and random oversampling, we train a quorum of neural networks that each learn slightly differently. Then through a threshold based voting scheme, we predict the mortality of a sample. This model earned an event 1 score (as defined by the PhysioNet challenge) of 0.538 which would have placed 1st in the competition.