Artificial Intelligence for Emergency Department Triage and Risk Stratification: A Review

Abstract

Researchers increasingly investigate artificial intelligence (AI) as a tool to support triage and early risk stratification in emergency departments. However, improved predictive performance does not necessarily translate into safer decisions or better patient outcomes. This critical narrative review examines the evidence for AI in adult ED care, with particular emphasis on prospective evaluation, patient safety, and clinical implementation. The review compares conventional triage systems, early warning scores, and clinician assessment with machine learning, natural language processing, and large language models. It examines commonly used data inputs, prediction timing, model discrimination, calibration, external validation, and the limitations of outcome definitions. Attention is given to the distinction between predicting clinical deterioration, critical care need, short-term mortality, and hospital admission. Evidence is considered across retrospective development, external validation, silent prospective evaluation, clinician-visible implementation, and randomized trials. AI may help identify patients requiring closer assessment by integrating information from multiple sources. Nevertheless, its clinical value depends on data quality, timely predictions, generalizability, and healthcare professionals' ability to interpret outputs appropriately. Important risks include under-triage, automation bias, unequal performance across patient groups, and increased clinical workload. Prospective studies and randomized trials are therefore essential to establish whether AI-supported decisions improve patient outcomes compared with usual care. Future research should prioritize multicenter clinical evaluations, consistent outcome reporting, subgroup assessment, and continuous monitoring after deployment. AI implementation should also include clear accountability, human oversight, and procedures for reviewing adverse events. Evidence of predictive accuracy alone is insufficient to justify routine adoption. Clinical integration should depend on demonstrated safety and meaningful benefit for patients.