Development and validation of an automated triage system in the emergency department of Bordeaux university hospital

Overview & Synopsis

Triaging systems have been introduced as a tool to improve the patient workflow within an Emergency Department (ED). The triage scale is often a score between 1 and 5 and indicates the need for urgent care required by a patient. Although triage systems were first introduced to increase the objectivity within the triage process, it still relies heavily on a triage nurse's judgement. The use of an Artificial Intelligence (AI) system could be one manner in which bias is minimized to improve triaging and the overall ED process. The purpose of our study was to assess whether with clinical notes written by triage nurses or clinical data available or both we could classify each patient arriving at the ED in Bordeaux University Hospital and assign a triage score. The study included 296071 ED patient visits between 2013 and 2020. All information collected during triage was used in the analysis with XGBoost, LightGBM and clinical BELGPT-2 (GPT2 trained on French corpus and clinical notes) used to classify the triage score. Clinical BELGPT-2 outperformed all methods, obtaining a micro F1-score of 0.621 on all data, while 0.581 on unstructured text and 0.574 on structured, numerical data. Our results suggest the use of unstructured text significantly improves predictive performance. However, biases in triage nurse decision making affects the model's performance. A model, or triage system, that can implicitly account for the biases in triaging should be conceded, which consequently allows for the reliable assessment and prioritization of ED patients.