Prediction Length of Stay in Intensive Care Unit in the Presence of Missing Data

Abstract

Patients’ length of stay (LOS) in intensive care units (ICU) is an important factor for managing limited ICU resources such as beds, staffing, medicines, and medical devices. The goal of this study predicts that the ICU length of stay of patients is more than 3 days or not with Support Vector Machine (SVM), Logistic Regression (LR), XGBoost classifiers. We retrieved the 37,600 ICU patients’ demographics data and last measured vital signs in their first 12 hours of stay from the MIMIC-III database. We filled the missing patients' data with three missing data imputation methods, namely k-nearest neighbor imputation (KNN), multivariate imputation by chained equations (MICE), and SoftImpute. Our results indicated that filling missing data with the Soft-Impute yielded the highest AUC score for all classifiers. We obtained the highest area under the curve score as 66.1% with the XGBoost classifier and Soft-Impute missing data imputation.