전자의무기록 데이터와 머신러닝 기법을 활용한 중환자실 사망 예측모델 개발

Abstract

In this study, a mortality prediction model was developed and evaluated using Electronic Health Record (EHR) data and the XGBoost (eXtreme Gradient Boosting) machine learning algorithm. The research dataset comprises EHR records of a total of 5,211 patients who were admitted to the medicine and surgery intensive care units of a tertiary medical institution from January 1, 2017, to February 29, 2020. Models were developed based on the accessibility of variable data, their high correlation with mortality, and their diversity. The research results indicated that variables such as respiratory pattern, pulse rate, RASS (Richmond Agitation-Sedation Scale), and post-surgery admission status showed relatively high importance compared to other variables. Models demonstrated higher predictive power compared to APACHE II, Random Forest and logistic regression models (AUC .93 ~ .98, accuracy .91 ~ .95). The model utilizing various variables showed the highest predictive power, but even a model based solely on accessible physical examination records demonstrated sufficient predictive capability. It was confirmed that patient observations recorded by nurses holds high importance and accessibility in mortality prediction. Also using simplified variables with machine learning techniques can efficiently predict mortality. It is expected that healthcare professionals can reduce the burden of record-keeping for mortality prediction and receive more accurate information about patients.