Machine learning to improve predictive performance of prehospital early warning scores

Abstract

Abstract Early warning scores are used to assess acute patients’ risk of being in a critical situation, allowing for early appropriate treatment, avoiding critical outcomes. The early warning scores use changes in vital signs to provide an assessment, however they tend to identify a considerable number of false positive cases, especially among prehospital patients. We investigated the development and validation of predictive scores based on machine learning models among patients (aged ≥ 18 years) who used ambulances in the North Denmark Region from July 1, 2016, to December 31, 2020. The machine learning models were compared to standard early warning scores (NEWS2 and DEPT), on 7- and 30-day mortality and intensive care admission. The cohort of 219,323 patients was split into development (n = 175,458 (80%)) and validation (n = 43,865 (20%)) datasets to respectively develop and test the machine learning models. These models were logistic regression, random forest, Bayesian networks, and gradient boosting. The machine learning models outperformed NEWS2 and DEPT, with fewer false positives, reducing the number of patients needed to screen by nearly half, for 7 day mortality. This has the potential to reduce both under- and over-triage, improving the precision of the triage among prehospital patients.