BACKGROUND Out-of-hospital cardiac arrest has poor outcomes, and most patients do not survive. Emergency Medical Services must rapidly identify patients who may benefit from resuscitation and recognise when resuscitation is clinically inappropriate. Early prognostic information is captured in the computer-aided dispatch system. We developed and internally validated a Gradient Boosting Machine Learning model using this data to identify patients for whom resuscitation was not attempted. METHODS We conducted a retrospective study of 37,532 out-of-hospital cardiac arrest incidents in one UK ambulance service between 2020 and 2023. Event density was sufficient to support model development. A gradient boosting model was trained on dispatcher text summaries, with robustness assessed using stratified five-fold internal cross-validation, and performance compared with keyword-based triggers. RESULTS FastText embeddings provided superior text representation. For prediction of no resuscitation (prevalence 70.3%), mean cross-validated area under the precision–recall curve was 0.933 (95% CI 0.931–0.935), with good calibration. Compared with keyword-based triggers, the model demonstrated a 2.5-fold increase in recall at equivalent precision at a 95% probability threshold, and a 1.75-fold increase in recall with 2.7% higher precision at a 99% probability threshold. Random oversampling outperformed Synthetic Minority Oversampling Technique and class weighting for handling class imbalance (Brier scores 0.146, 0.156, and 0.186 respectively). Feature importance analysis identified semantically meaningful clinical drivers. CONCLUSION An internally validated machine learning model outperformed keyword-based triggers in identifying out-of-hospital cardiac arrest cases where resuscitation is not attempted, with potential to support more appropriate Emergency Medical Services triage.
McMorran et al. (Mon,) studied this question.