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

Abstract

Early prediction of hospital admission at emergency department triage may support clinical decision-making, patient flow coordination, bed planning and resource allocation. However, external validation of admission prediction models remains limited, while missing-data handling, calibration, clinical utility, and interpretability are inconsistently addressed. We developed an interpretable triage-only admission prediction model using 1,000 MIMIC-IV-ED encounters. Missing data were handled using 15 MICE-PMM imputations within a leakage-resistant, patient-grouped 5×5 nested cross-validation framework. Seven classifiers were benchmarked under the same analytical pipeline and showed closely comparable discrimination, and L2-regularized logistic regression was retained as the final model based on its interpretability and direct coefficient-based odds ratios. Performance was assessed using AUROC, calibration, Brier score, and decision curve analysis, while interpretability was examined using SHAP and odds ratios. The finalized model was externally validated in an independent cohort of 118,349 MC-MED encounters. Internal validation yielded an AUROC of 0.805 (95% CI 0.779-0.832) and a Brier score of 0.181(95% CI 0.169-0.193). In external validation, AUROC was 0.783 (95% CI 0.779-0.786) with a Brier score of 0.185(95% CI 0.184-0.187). Triage acuity and age were the most influential predictors. Discrimination was similar between females and males, while calibration remained close to 1 across gender and age subgroups. Overall, the findings support an interpretable triage model whose performance remained stable across validation settings and whose outputs could inform early bed-demand, patient-flow, and resource planning, providing a foundation for future hospital supply chain optimization.