Predictive analytics for hospital admissions from the emergency department using triage information

In this paper, we investigate the predictive factors of hospital admissions from the emergency department (ED) in various classification models and evaluate their performance. We analyze data from a major hospital in a metro area in the United States with an approximately 50,000 ED visits per year using logistic regression, artificial neural network, decision tree, random forest, support vector machine and extreme gradient boosting methods. The predictive accuracy of the models are evaluated in multiple experiments in which data volume on training and validation varied with multiple years of data. The data set also included a set of observations from a year when an influenza pandemic occurred, which made the data set statistically different. The extreme gradient boosting algorithm (XGBoost) gives the highest area under curve (AUC) statistic and it stands out as one of the fastest algorithms. However, a less sophisticated model, such as simple logistic regression model, also performs well in a reasonable computational time. We also evaluate performance of models with correct classification of patient admission from the emergency department. In our experiments, increasing data volume for model training and validation do not change performances of models significantly, since we have used a large number of observations in initial training and validation experiments. However, support vector machines (SVM) show notable decrease in AUC statistic while model accuracy of the XGBoost algorithm increases. In addition, increasing the data volume slightly increases the AUC statistic for the XGBoost algorithm (i.e., from 83% to 86%) while other methods show slight decreases. Using the most accurate and efficient predictive model for ED admissions would help hospitals design a decision support systems for efficient information flow from the triage at the ED to both outpatient and inpatient units in case of admissions, as basic administrative data with patient acuity information can provide predictions for bed capacity planning.