GAM feature selection to discover predominant factors for mortality of weekend and weekday admission to the ICUs

Abstract

Abstract “Weekend effect” on mortality is a common controversial topic among many hospitals. Although the mortality of patients associated with weekend Intensive Care Units (ICUs) admission has been demonstrated slightly higher than that of patients admitted on weekdays in many studies, the underlying causal mechanisms and the potential factors are not clear at present. In this study, we extract medical record features from the database and propose a Generalized Additive Model(GAM) feature selection method to identify the main contribution features for analyzing this issue. The GAM feature selection system could rank candidate features by its importance, which turns out to be effective in reducing the complexity of a medical issue. The best lists of features are acquired by the weekday GAM model and the weekend GAM model separately. Fourteen out of forty-one features are identified for the reduced list of features. Both models’ reduced lists of features have ten identical characteristics, and the other four are different. The prediction accuracy with the reduced list of features is 79.90% for the weekday model and 78.56% for the weekend model. The contrast experiment has validated the feature ranking results. Furthermore, variables of the same feature classes are also different from weekday admission to weekend admission. We expect that the proposed GAM feature selection method could contribute to solving more medical issues in the future.