Model comparison in Emergency Severity Index level prediction

•We model ESI prediction using data from 870 patients over one month in 2008.•Three methods are compared: regression, Bayesian networks and neural networks.•Saturated oxygen level and chief complaint are significant predictors of ESI level.•Models performed best using all data and had about 68% accuracy.•Naïve Bayesian networks are recommended over neural networks and logistic regression. Emergency Department (ED) triage is a process of determining illness severity and accordingly assigning patient priority. The Emergency Severity Index (ESI) is a 5-level acuity categorization system that aides in triage. This paper compared the capabilities of predicting ESI level using ordinal logistic regression (OLR), artificial neural networks (NNs), and naïve Bayesian networks (NBNs). Data were obtained from Susquehanna Williamsport Hospital for 947 patients over a one month period in 2008. It contained the assigned ESI level, chief complaint, systolic blood pressure, pulse, respiration rate, temperature, oxygen saturation level (SaO2), age, gender, and pain level. An OLR model was fit using a subset of these covariates. NBNs and NNs were modeled to relax the inherent assumptions of linearity and covariate independence in logistic regression. These three techniques were compared using incremental training dataset sizes between 50% and 100% of given data. All models were >60% accurate using the entire dataset for training. It was found that NBNs and NNs were robust to data size changes and all models had evaluation speeds of less than 0.5s. At this time the use of NBNs is recommended considering speed, accuracy, data utilization, model flexibility, and interpretability of the model.