I. INTRODUCTION
Health-related data sets often contain many features relative to the number of cases, giving rise to the problem of how to select a few relevant features and discard irrelevant ones, before subjecting the data to machine learning protocols to increase predictive accuracy [1]. This ‘curse of dimensionality’ has been the subject of much debate in the machine learning literature [2], [3]. Current supervised machine learning programs incorporate a feature selection tool as part of the process. However, they use the existing classification outcome information within the selection process introducing bias into the process and result in models that may overfit the data and may not generalize [4]. An approach to feature selection that avoids the biasing problem, as the analysis operates independently of the classification outcomes, uses Item Response Theory (IRT). IRT has been applied in machine learning contexts, such as comparisons of collaborative filtering [5], natural language processing [6], identification of initial computer adaptive learning items [7], and classifier assessment [8], suggesting is has the potential to address some long-standing concerns of this research community. The purpose of the current study, however, is to demonstrate the use of IRT in feature selection using intensive care data and no-death/death outcomes for patients.