PREDICTING THE MORTALITY OF PNEUMONIA PATIENTS VISITING THE EMERGENCY DEMARTMENT THROUGH MACHINE LEARNING
Yeol Bae, Hyung Ki Moon, Su Hyun Kim · 2018
Background and Aims: Machine learning in the medical field is not yet widely used. The aim of this study is to compare the performance of pre-existing severity prediction models and Random forest(RF) based models for mortality prediction in pneumonia patients. Methods: We retrospectively collected the data of patients who visited the emergency department of a tertiary training hospital in Seoul, Korea from January to March 2015. The pneumonia severity index(PSI) and SOFA score were calculated for both group and the AUC for the mortality prediction was computed. For RF model, data were divided into a test set and a validation set by random split 100 times. The training set was learned in an RF model and the AUC was obtained from the validation set. The mean AUC was compared with the other two AUCs. RF model was built using python scikit-learn library 0.18 version. Results: Of the 443 people, 395 were enrolled and 41 of them were died. The AUC values of PSI and SOFA scores were 0.799 (0.737 - 0.862) and 0.865 (0.811 - 0.918), respectively. The mean value of AUC obtained by RF method was 0.916 (0.909 - 0.923) and there were significant differences statistically (p <0.001). The five major features used in the random tree model were vasoactive agent, platelet, lactic acid, GCS score, and albumin. Conclusions: Classification through machine learning may help to predict the mortality of patients visiting the emergency department.