Predicting Emergency Severity Index (ESI) level, hospital admission, and admitting ward in an emergency department using data-driven machine learning.

INTRODUCTION

Emergency departments (EDs) are critical for ensuring timely patient care, especially in triage, where accurate prioritisation is essential for patient safety and resource utilisation. Building on previous research, this study leverages a comprehensive dataset of 653,546 ED visits spanning six years from Mater Dei Hospital, Malta. This dataset enables detailed trend analysis, demographic variation exploration, and predictive modelling of patient prioritisation, admission likelihood, and admitting ward.

METHODS

Two predictive models (Stage 1 and Stage 2) were developed using the Extreme Gradient Boosting (XGBoost) algorithm. In Stage 1, predictions were made at the triage level using basic demographic and presenting symptom data. Stage 2 incorporated critical blood test results (e.g., Haemoglobin, C-Reactive Protein, Troponin T, and White Blood Cell Count) alongside the demographic and symptom data from Stage 1 to refine and enhance predictions. Key steps in data preprocessing, such as handling missing values, balancing class distributions with SMOTE, and feature encoding, are discussed. Model evaluation employed comprehensive metrics, including AUC-ROC and calibration curves, to assess both performance and reliability. This enhanced description provides a clear roadmap of the model development process, reinforcing the study's rigor and contribution to advancing machine learning applications in emergency care.

RESULTS

The models demonstrated significant predictive capabilities. Key metrics showed improvement between Stage 1 and Stage 2. For example, patient prioritisation accuracy improved from 0.75 to 0.76, admission prediction accuracy rose from 0.80 to 0.82, and admitting ward prediction accuracy increased from 0.80 to 0.86. These enhancements underscore the value of incorporating clinical data to optimise predictions.

DISCUSSION

The integration of early predictions into ED workflows has the potential to improve patient flow, reduce wait times, and enhance resource allocation. By leveraging XGBoost's capabilities and integrating both demographic and clinical data, this study provides a robust framework for advancing decision-making processes in triage environments.

CONCLUSIONS

This research demonstrates the efficacy of machine learning models in predicting key ED outcomes, highlighting their potential to transform emergency care through data-driven insights.