Explainable Machine Learning Model for Early Hospital Admission Prediction in the Emergency Department: A Retrospective Analysis

Emergency Department (ED) overcrowding represents a critical challenge in emergency care. Early admission prediction could optimize resource allocation, yet existing models rely on large datasets or black-box algorithms. The aim of this paper is to develop and internally validate an interpretable triage-only machine learning model for hospital admission prediction, following TRIPOD+AI 2024 standards. Using a stratified random sample of 1,000 ED visits from MIMIC-IV-ED, we selected 24 triage-based predictors and handled missing data through Multiple Imputation by Chained Equations with Predictive Mean Matching (MICE-PMM) combined with Rubin’s Rules pooling. Five algorithms were evaluated using Bayesian hyperparameter optimization and stratified 5-fold cross-validation. Given the statistically equivalent discriminative across all candidate models, an -regularized logistic regression model was selected to prioritize clinical interpretability and avoid the black-box nature of more complex algorithms. The model achieved an AUC of 0.796 (95% CI: 0.770-0.823) and Brier Score of 0.186 (25.6% improvement over null). A high-sensitivity threshold (p=0.354; sensitivity=0.878, specificity=0.547) and a balanced threshold (p=0.427; sensitivity= 0.779, specificity=0.653) were identified, yielding clinical actionable insights. Decision Curve Analysis demonstrated positive net benefit over treat-all across the full threshold range. Odds ratio and SHAP analyses identified triage acuity, ambulance arrival, and age as dominant predictors, providing evidence of the model’s clinical plausibility and face validity. This study shows that a rigorous, explainable model trained on a small representative sample can provide clinically actionable decision support, helping practitioners streamline admission decisions and optimize ED workflows.