Machine learning models powered by emergency medical services data enhance stroke triage in prehospital settings

Timely stroke diagnosis is essential for delivering life-saving treatments, yet current prehospital stroke assessment tools often lead to missed or delayed diagnoses. Emergency Medical Services (EMS) increasingly collect detailed data during transport, offering an opportunity to develop AI-based tools to support early stroke detection. In this retrospective study, we evaluated the availability and reliability of prehospital data compared to Emergency Department (ED) records. We tested the potential of machine learning to support EMS stroke triage. Our cohort included 4,333 patients across 8,221 ambulance encounters from 2015 to 2020 with stroke rate of 2.0% (64% severe strokes). Vital signs such as heart rate, respiratory rate, blood pressure, oxygen saturation, and GCS scores, recorded in over 88%, were generally higher than their ED counterparts. We trained and evaluated random forest, XGBoost, and sequential neural networks for detecting stroke and severe stroke. The XGBoost model performed best for stroke detection (ROC-AUC 0.843 [0.77–0.89], PR-AUC 0.293 [0.16–0.45]), while Random Forest performed best for severe strokes (ROC-AUC 0.826 [0.75–0.90], PR-AUC 0.186 [0.07–0.35]). Models were calibrated to improve reliability, and feature importance was assessed using SHAP to enhance interpretability. These findings highlight the promise of AI-based tools in improving prehospital stroke triage with real-time EMS data.