Analysis of length of stay for patients admitted to Korean hospitals based on the Korean National Health Insurance Service Database

Abstract

The aim of this study is not only to predict length of stay (LOS) for all patients admitted to medical institutions in South Korea, but also to unveil features that would significantly influence LOS. As our dataset, we used the 2016–2019 health checkup cohort DB data stored in the virtual server of National Health Insurance Service (NHIS) of Korea. Since this study presents the pioneer work of predicting LOS for all inpatients in South Korea to our best knowledge, we referred to extant literature completed elsewhere for feature selection with regard to LOS analysis. In this paper, we compared the prediction performance of inpatients' LOS between four machine learning models including XGBoost, Decision Tree, Random Forest, and Support Vector Machine. The results indicate that the XGBoost model shows the best performance with accuracy of 0.7472 and AUC of 0.88. Furthermore, using the XGBoost model, we discovered features that had a significant influence in predicting LOS with the use of feature importance. We found that the critical features in determining LOS were examination records which are consisted of age, the deductible insurance ratio and information with regard to one's main diagnosis and the number of sub diagnosis, the number of doctors and beds available in hospitals, cholesterol level, body mass index (BMI), and the admission month. The significance in this study is that we found the important common features that affect LOS for inpatients across medical department and we could make a generalization of our research result.