Random forest machine learning method outperforms prehospital National Early Warning Score for predicting one-day...

Area of Science:

• Prehospital emergency medicine
• Clinical prediction models
• Machine learning in healthcare

Background:

• The National Early Warning Score (NEWS) is established for hospital ward deterioration prediction but its prehospital utility is debated.
• Machine learning (ML) models show potential for superior short-term mortality prediction over traditional methods.
• Accurate prediction of prehospital mortality is crucial for timely intervention and resource allocation.

Purpose of the Study:

• To compare the accuracy of the National Early Warning Score (NEWS) against a random forest machine learning model for predicting one-day mortality in prehospital patients.
• To evaluate the added value of blood glucose measurement within the random forest model.
• To assess the performance of ML models using prehospital vital signs.

Main Methods:

• Retrospective analysis of 26,458 adult patients from electronic ambulance mission reports (2008-2015).
• Comparison of traditional NEWS with two random forest models: one using NEWS variables and another including blood glucose.
• Ten-fold cross-validation was employed for model training and validation.

Main Results:

• The random forest model using NEWS variables achieved an Area Under the Curve (AUC) of 0.858, outperforming NEWS (AUC=0.836) for one-day mortality prediction.
• Inclusion of blood glucose further improved the random forest model's predictive performance (AUC=0.868).
• A total of 278 patients (1.0%) died within one day of the ambulance mission.

Conclusions:

• Random forest algorithms trained with prehospital vital signs, particularly when including blood glucose, demonstrate superior accuracy in predicting one-day mortality compared to NEWS.
• Machine learning offers a promising advancement for risk stratification in the prehospital setting.
• Acknowledged risk of selection bias necessitates cautious interpretation.