A Comparison of Feature Selection Techniques for First-day Mortality Prediction in the ICU

First-day mortality prediction is a critical task in the Intensive Care Unit (ICU), as it can help clinicians identify which patients are at the highest risk for death and thus may need more intensive care. Many heuristic-based metrics for first-day mortality exist. However many signals exist in the Electronic Health Record (EHR) that are not used in these metrics. For the implementation of these metrics, it is important to keep the number of required signals as small as possible so that the user can quickly receive an assessment of their patient. In this paper, we leverage techniques from classical machine learning to compare sets of signals from a patients record (like demographic information, lab values, and vital sign measurements) to find the minimum feature set that best informs first day mortality. We compare several feature selection techniques to identify various feature sets with differing number of features. We found that Elastic Net was the overall best performing method and was able to reach the same performance as the current state of the art with less than half the features. This suggests that an optimal feature set is clinically meaningful.