Prediction of Length of Stay Using Vital Signs at the Admission Time in Emergency Departments

Abstract

Length of Stay (LOS) prediction at the time of admission can give clinicians insight into the illness severity of patients and enable them to prevent complications and adverse events. It can also help hospitals to manage their facilities and manpower more efficiently. This paper first applies Borderline-SMOTE and multivariate Gaussian process imputer techniques to overcome data skewness and handle missing values which have been ignored by most studies. Then, based on our conversation with clinicians, patients are stratified into five classes according to their LOS. Finally, five machine learning algorithms, including support vector machine, deep neural networks, random forest, extreme gradient boosting, and decision tree are developed to predict LOS of unselected patients admitted to the emergency department at Odense University Hospital. These models utilize information of patients at the time of admission, including age, gender, heart rate, respiratory rate, oxygen saturation, and systolic blood pressure. Performance of predictive models on the data before and after imputation and class balancing are investigated using the area under the curve metric and the results show that our proposed solutions for data skewness and missing values challenges improve the performance of predictive models by an average of 13%.