Abstract
Prehospital triage is critical for time-sensitive emergencies such as trauma, stroke, and acute coronary syndrome. Undertriage delays definitive care, while overtriage strains higher-level facilities. Existing triage tools based on vital signs and scoring systems have limited accuracy, but artificial intelligence (AI), machine learning (ML), and neural networks (NN) offer the potential to improve decision-making by integrating multiple data sources. This narrative review of studies indexed in PubMed and PubMed Central through August 2025 evaluated AI, ML, and NN models designed for prehospital triage or transport destination decisions across trauma, critical illness, stroke, dyspnea, cardiac emergencies, sepsis, and in-hospital studies preventing possible readmission. Across the conditions examined, ML models consistently outperformed traditional early warning scores and guideline-based tools. Trauma models achieved area under the curve (AUC) values between 0.75 and 0.93 and reduced undertriage to less than 10%. ML models predicted the need for critical care with an AUC of 0.908, and prehospital stroke algorithms reached AUCs above 0.98. NN, deep forest, and random forest models demonstrated an AUC of 0.88 in prehospital acute respiratory distress syndrome (ARDS) prediction. Additional studies demonstrated improved recognition of dyspnea-related serious events and acute coronary syndrome, while no validated models currently exist for prehospital sepsis. Despite promising results, most studies were retrospective, with limited prospective validation, generalizability, or evaluation of workflow integration. Future research should focus on prospective studies, diverse patient cohorts, integration into emergency medical service (EMS) workflows, model explainability, and rigorous comparisons with standard practice.