Artificial Intelligence-Driven Clinical Decision Support Systems for Emergency Department Triage and Acute Care Prioritization: A Systematic Review of Accuracy, Safety, and Implementation Readiness

Abstract

Background: AI is being increasingly investigated as a clinical decision support tool for ED triage and acute-care prioritization. Machine-learning, deep-learning, natural language processing (NLP) and large language model systems have shown good predictive performance but concerns remain about under-triage, over-triage, generalizability, workflow integration and implementation safety. We systematically reviewed the accuracy, safety and implementation readiness of AI-driven clinical decision-support systems for ED triage.

Methods: This review was conducted according to PRISMA 2020 recommendations. PubMed/MEDLINE, PubMed Central, Embase, Scopus, Web of Science and Cochrane Library were searched (citation chaining, reference-list screening and recent systematic reviews were also used). Of the 1076 records identified, 312 duplicates were removed and 764 records screened; 79 full-text reports were assessed and 65 excluded; fourteen studies were included in narrative synthesis. Meta-analysis was not performed because populations, AI architectures, triage systems, reference standards and outcome metrics are highly heterogeneous.

Results: Conventional machine learning, gradient boosting, neural networks, deep learning, natural language processing and large language models were included in the studies. AI systems generally outperformed conventional triage for critical illness prediction (hospitalization), intensive-care admission and mortality; reported AUROCs for early mortality were 0.962 and for 30-day mortality 0.925. Direct triage systems also had some success: KATE had 75.7% ESI assignment accuracy; prospective GPT-4 and GPT-4o studies had κ = 0.899 and multicenter κ = 0.833 with F1 = 0.897 agreement with expert triage but safety outcomes under-triage/over-triage subgroup performance and patient-important outcomes were inconsistently reported.

Conclusions: ED triage systems powered by AI have great potential as supervised clinical decision support tools but not yet enough evidence to justify autonomous deployment. Prospective multicenter validation, standardized safety metrics, subgroup analysis, workflow evaluation, governance and post-deployment monitoring are needed before widespread implementation.