A machine learning model for predicting short‐term outcomes after rapid response system activation

Abstract

The rapid response system (RRS) has become the standard international system for preventing death from unexpected cardiac arrests. 1-4 Recently, the development of better prediction models for adverse events has become an active research area related to RRS. Machine learning models show better prediction than the widely used early warning scores (EWSs), such as the National EWS (NEWS) or the modified EWS (MEWS). 5-11 Although developing better track-and-trigger systems is important for improving the quality of in-hospital emergency systems, saving patients from preventable death requires adequate treatment provided by the medical emergency team (MET) and rapid response team (RRT). The importance of accurately assessing severity and providing care levels that match that severity has been highlighted. 12,13 Ideally, once the RRS is activated, patients should receive the best treatment. However, maintaining the quality of RRT/MET response at a high level can be Abstract Aim: Maintaining rapid response team (RRT) response quality is difficult. A system that supports RRT assessment could potentially contribute to medical safety. Although rapid response system (RRS) triggers have been well-studied, studies on the prediction models of short-term prognosis after RRS activation are scarce. We aimed to develop a model to predict short-term outcomes after RRS activation using machine learning. Methods: This retrospective cohort study used the In-Hospital Emergency Registry in Japan, a multicentre RRS online registry. We collected data on patient demograph-ics, treatment before RRS, RRT calls, and physiological parameters. The outcome was death within 24 h after RRS calls or unplanned transfers to an intensive care unit. To develop the eXtreme Gradient Boosted Tree Classifier (XGB) and Random Forest (RF) algorithms, a logistic regression (LR) algorithm was used. For model comparison, receiver-operating area under the curve (AUC) was evaluated and compared with those of the National Early Warning Score (NEWS) and Modified Early Warning Score (MEWS). Results: 5414 cases were included in the study. The outcome occurred in 28.4% of the cases. The XGB model showed the highest AUC (0.798) compared to the RF model (0.796), LR model (0.785), NEWS (0.696), and MEWS (0.660). The most weighted feature in the XGB model was doctor activation, followed by hypotension as the activation criteria and usage of oxygen. Conclusions: We developed the first machine learning model for short-term prognosis after RRS. This model has the potential to support decision-making by RRT. K E Y W O R D S early warning score, machine learning, medical emergency team, rapid response system