A foundational triage system for improving accuracy in moderate acuity level emergency classifications.

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Clinical Decision Support

Background:

• Emergency department (ED) triage is critical but challenged by staff shortages and diagnostic ambiguity.
• Existing AI triage models struggle with classifying moderate acuity cases, impacting patient care.
• Developing AI to assist in accurate triage is crucial for efficient healthcare delivery.

Purpose of the Study:

• To develop a novel AI triage model, KUTS (Knowledge-based Uncertainty-Tuned Triage System), for improved classification accuracy.
• To enhance the identification of moderate acuity cases, which are difficult for human clinicians.
• To establish a foundational framework for advanced emergency triage systems.

Main Methods:

• Developed KUTS, a foundational classification model for emergency triage.
• KUTS integrates a knowledge prompt-tuning encoder and an uncertainty-based classifier.
• The model processes tabular data and chief complaints to assign patient acuity levels, tested on real-world ED datasets.

Main Results:

• KUTS achieved a 0.95 AUC score for moderate acuity cases, outperforming prior methods by 0.19-0.35.
• Across all triage levels, KUTS demonstrated superior performance with an average AUC of 0.96, exceeding previous models by 0.06-0.20.
• The model showed substantial improvements in classifying difficult and ambiguous patient acuity levels.

Conclusions:

• KUTS offers a robust framework and paradigm for emergency triage research.
• The model facilitates the development of more accurate and efficient patient triage systems.
• KUTS represents a significant advancement in AI-assisted emergency medicine.