Machine learning models show variable accuracy predicting pediatric ED admissions in meta-analysis

UNLABELLED: Machine learning (ML) models have shown promise improving outcome prediction and early risk stratification in paediatric emergency department (ED) triage. This review aims to evaluate the diagnostic performance of ML in predicting hospital admissions from the data collected at triage in paediatric emergency departments (EDs). Searches were conducted in PubMed, Ovid, Scopus, and Web of Science. Two reviewers screened 264 abstracts after duplicate removal, excluding 239 not meeting inclusion criteria. Of the 25 full-texts assessed, 15 were excluded for outcome mismatch, leaving 10 for data extraction. Data were thereafter extracted including population characteristics, ML methods, and diagnostic metrics: area under the curve (AUC), sensitivity, and specificity. Most studies used retrospective cohorts from electronic records or national databases. Sample sizes ranged from 9,069 to over 2.9 million. AUCs ranged from 0.78 to 0.97, with top-performing models (AUC ≥ 0.94) using random forest algorithms and variables like age, heart rate, triage level. Meta-analysis of six studies showed pooled sensitivity of 0.78 and specificity of 0.76 (AUC = 0.84), though heterogeneity was high (I = 100%). CONCLUSION:  ML models have potential for paediatric ED triage. Standardized methods, explainable AI, and prospective validation are essential for clinical use. WHAT IS KNOWN: • Traditional triage in paediatric emergency departments may have limitations in accurately predicting hospital admissions. • Machine learning models are increasingly applied to improve risk stratification in clinical settings. WHAT IS NEW: • This review shows ML models can predict paediatric ED admissions with high AUCs (up to 0.97). • Random forest algorithms using vital signs and triage data performed best.