AI-ENHANCED EMERGENCY SEVERITY INDEX FOR COMPREHENSIVE ED TRIAGE

Background: Emergency Department (ED) triage systems, such as the Emergency Severity Index (ESI), are crucial for patient prioritization but have limitations in predicting outcomes and optimizing resource allocation. AI-enhanced triage systems offer potential improvements in accuracy, efficiency, and patient care.

Objectives: To develop an optimized AI-enhanced ESI score using the MIMIC-IV ED dataset, integrating patient history, real-time vital signs, model predictions, and traditional nurse-assigned ESI to create a more comprehensive, dynamic, and accurate triage tool.

Methods: We analyzed over 400,000 ED visits from the MIMIC-IV dataset. Using a customized autoML pipeline, we engineered features from triage variables, patient history, and temporal vital sign patterns. Multiple machine learning models, including XGBoost and MLPs, were trained to predict various outcomes. These predictions, along with the nurse-assigned ESI and real-time patient data, were then combined to create an optimized AI ESI score. Performance was evaluated using AUC, calibration metrics, and compared to traditional ESI scoring.

Results: Individual models demonstrated high performance in predicting key outcomes: hospitalization (AUC: 0.85), critical outcomes (ICU admission/mortality within 12 hours, AUC: 0.88), and ED reattendance (AUC: 0.75). These models, triage variables, and vital sign patterns were combined into a single AI-enhanced ESI score, which showed stronger correlation with critical outcomes compared to the nurse-assigned ESI. This combined score allowed for more precise patient ranking than the traditional 5-level ESI system, effectively refining the nurse-assigned ESI.

Conclusion: This AI-enhanced ESI system demonstrates significant improvements over traditional methods by integrating patient information with model predictions into a single, comprehensive, and dynamic triage score. It has the potential to optimize resource allocation, improve patient flow, and enhance overall patient care in EDs. Future work will focus on prospective validation, seamless integration into clinical workflows, and assessment of its impact on patient outcomes and ED efficiency.