Machine learning models for predicting hospital admission in pediatric emergency departments: A systematic review.

Abstract

Pediatric Emergency Departments (PEDs) face overcrowding partially due to delayed hospital admission decision. Machine Learning (ML) models could early predict it. To systematically review and critically appraise the development, validation, quality, risk of bias, and applicability of ML models for predicting PEDs hospital admission. PubMed, Cochrane, Web of Science and Scopus were searched for studies published up to February 25, 2025. Studies that developed and/or validated predictive models for hospitalization from PEDs using ML methods were included. Case reports, reviews, meta-analyses, and non-English/Spanish studies were excluded. Main characteristics from the selected studies were extracted using a standardized form. Quality, risk of bias and applicability were assessed using the PROBAST + AI tool. Nineteen studies were included. Only one was prospective, nine multicentric, and seven included external validation. Final predictors for each model ranged from three to 6009, with most common including age, sex, chief complaint, arrival mode and triage category. The most frequent ML algorithm used was random forests (n = 9). Reported model performance varied widely (AUC-ROC 0.624-0.968) independently from sample sizes and algorithms used. Methodological inconsistencies, poor reporting, low quality and high risk were common. Only one received a highly favorable PROBAST + AI judgement. There was substantial heterogeneity and suboptimal transparency across the selected studies. ML models show promise in supporting early decision-making for hospitalization in PEDs, but studies frequently present methodological and reporting limitations. Future research should prioritize more rigorous designs, prospective external validation, and transparent reporting.