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

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Emergency Medicine

Background:

• Identifying febrile patients needing hospitalization in Emergency Departments (EDs) is challenging.
• Delayed interventions for such patients can increase mortality.
• Current methods may not fully leverage diverse patient data.

Purpose of the Study:

• To develop and validate a novel predictive model for ED disposition decisions.
• To identify febrile patients requiring hospitalization or readmission within 72 hours.
• To improve patient outcomes and ED efficiency through early identification.

Main Methods:

• Retrospective study of 25,405 febrile ED patient visits.
• Utilized triage data including vital signs and self-reported symptoms.
• Employed pre-trained language models (GatorTronS) integrating numerical and narrative data.

Main Results:

• Achieved superior performance (AUROC: 0.9226, AUPRC: 0.8767, Macro F₁-score: 0.8446) compared to traditional methods.
• Identified demographic patterns: elderly, males, and ambulance arrivals more likely hospitalized.
• Demonstrated effective integration of vital signs and clinical narratives.

Conclusions:

• The novel language model approach accurately predicts hospitalization needs for febrile ED patients.
• Early identification can reduce inappropriate discharges and optimize resource allocation.
• This AI-driven tool enhances emergency care efficiency and patient outcomes.