Optimizing ED patient disposition predictions through clinical narratives with advanced pre-trained language models

Timely identification of febrile patients requiring hospitalization remains a significant challenge in Emergency Departments (EDs), with delayed intervention potentially increasing mortality. This retrospective study presents a novel approach for predicting ED disposition decisions using pre-trained language models. We developed and validated a predictive model utilizing triage information from 25,405 febrile ED patient visits, including vital signs and self-reported symptoms, to identify patients requiring hospitalization or returning for admission within 72 h. The model employs an innovative methodology that transforms numerical vital signs into narrative text, allowing for comprehensive integration with clinical narratives through the GatorTronS architecture. Our approach demonstrates superior performance (AUROC: 0.9226, AUPRC: 0.8767, Macro F₁-score: 0.8446) compared to traditional machine learning methods and other biomedical language models. Analysis revealed significant demographic patterns, with hospitalization more prevalent among elderly patients, males, and those arriving by ambulance. By facilitating early identification of patients requiring admission, this model addresses critical challenges in emergency care, including reducing inappropriate discharges and optimizing resource allocation, ultimately improving patient outcomes and ED efficiency.

The online version contains supplementary material available at 10.1007/s13755-026-00431-4.