Predicting hospital admission at emergency department triage using machine learning

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Emergency department (ED) triage is critical for patient management.
• Accurate prediction of hospital admission at triage can optimize resource allocation and patient flow.
• Existing models often rely solely on immediate triage information.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting hospital admission.
• To assess the added value of patient history data to triage information for admission prediction.
• To identify key variables for a low-dimensional predictive model.

Main Methods:

• Retrospective analysis of 560,486 adult ED visits from multiple facilities.
• Training and testing of logistic regression, XGBoost, and deep neural networks on triage data, patient history, and combined datasets.
• Variable importance analysis using information gain to create a parsimonious model.

Main Results:

• Models using combined triage and history data achieved higher AUCs (up to 0.92) than those using triage data alone (AUC 0.87).
• Machine learning algorithms reached peak performance with 50% or less of the training data.
• A low-dimensional XGBoost model incorporating ESI level, medication counts, demographics, and hospital usage achieved an AUC of 0.91.

Conclusions:

• Machine learning models effectively predict hospital admission using a combination of triage and patient history data.
• Patient history significantly improves predictive performance over triage information alone.
• The findings underscore the importance of integrating historical patient data into admission prediction models.