Feature Representation for Predicting ICU Mortality

Effective predictors of Intensive Care Unit (ICU) Mortality have the potential to identify high-risk patients earlier, improve ICU resource allocation, and create more accurate population-level risk models. Machine learning practitioners typically make choices about how to represent features in a particular model, but these choices are seldom evaluated quantitatively. This study compares the performance of different representations of clinical event counts from the MIMIC III database in a logistic regression model to predict post-36-hour ICU mortality. The most common representations are linear (normalized counts) and binary (yes/no). These, along with a logarithmic representation and a new representation termed Hill, are compared using L2 regularization. Results indicate that the introduced Hill representation gives a higher area under the reciever operating curve (AUC) for the mortality prediction task than the log, binary and linear representations. The Hill representation thus has the potential to improve existing models of ICU mortality.