Machine Learning-based Prediction of ICU Patient Mortality at Time of Admission

In this paper we describe the application of machine learning techniques to the prediction of hospital Intensive Care Unit (ICU) patient mortality. A large dataset of over 58,000 ICU admissions from the MIMIC III database was used in the development, training and evaluation of a number of all-condition patient mortality predictive models. Evaluation results are presented, showing favorable performance comparing with existing studies, in the specific context that this work presents models making predictions for patients of all conditions as opposed to restricting to patients with a given condition or group of conditions, and the models are developed using only input attributes that are patient administrative data available at time of hospital admission. In this way, our models provide a unique utility for ICU mortality prediction in terms of their being applicable to all patients and at the earliest point in time of admission and utilizing a minimal and routinely collected set of patient administrative data.