Introduction: Emergency departments (EDs) face the critical challenge of quickly and accurately triaging patients, a process essential for prioritizing care and allocating resources effectively. Current triage systems, while functional, often fall short in rapidly identifying critically ill patients, potentially delaying necessary interventions. This study aims to address these shortcomings by developing a more precise predictive model using machine learning algorithms, thus enhancing the efficiency and safety of triage operations. Methods: The study employed an exploratory, retrospective observational design to analyze 560,336 emergency triage records from the Fourth Military Medical University’s First Affiliated Hospital, collected between 2014 and 2023. The dataset included a mix of structured and unstructured data elements, including demographic information, vital signs, chief complaints, and triage levels. Our methodology encompassed data preprocessing, feature engineering, and text embedding using the Ernie-Health-Zh, a Chinese knowledge-enhanced pre-trained language model. We divided the data into a training set (70%) and a test set (30%) and developed models using Logistic Regression, Random Forest, and XGBoost algorithms. Results: The XGBoost algorithm emerged as the most effective, achieving a sensitivity of 0.9555 and an AUC value of 0.9956. It showed remarkably high classification accuracy across different severity levels of emergency conditions, with AUC values of 0.9900 for Level I, 0.9680 for Level II, 0.9839 for Level III, and 0.9993 for Level IV. Conclusion: The predictive model developed in this study significantly improves the accuracy of triage assessments in EDs, ensuring more effective and safer patient management. By enabling faster, more accurate identification of critically ill patients, the model optimizes the use of emergency resources and enhances overall patient outcomes. This advancement holds substantial potential to transform triage practices in emergency medicine, setting a new standard for emergency care efficiency.
Zhang et al. (Sun,) studied this question.