Development and External Validation of Interpretable Partial Dependent Plot-based Triage Score for Emergency Departments

Abstract

Abstract Triage in an emergency department (ED) can help identify the urgency of patients’ treatment and allocate the appropriate resources. Interpretable machine learning methods could be a helpful tool for facilitating the triage process. However, existing related research used only conventional logistic regression methods. This study aims to develop and externally validate an interpretable machine learning model using a partial dependent plot (PDP). This retrospective cohort study included all adult ED patients of the Samsung Medical Center for development and Tan Tock Seng Hospital, from 2016–2020. The outcome of interest was in-hospital mortality after patients’ ED visit. We used the area under the receiver operating characteristic curve (AUROC) to assess the performance of the PDP score and other conventional scores, including the Korea Triage Acuity Scale (KTAS). Of the included 285,523 ED visits, 1.60% ended in in-hospital mortality. The PDP score achieved an AUROC of 0.821 in temporal validation and 0.833 in external validation, outperforming the KTAS score of 0.729. The PDP triage score was therefore superior to other scores for in-hospital mortality prediction. PDP is a generic, intuitive, and effective triage tool to stratify general patients who present to the ED.