Abstract Background Emergency department (ED) crowding is a major global challenge that adversely affects patient safety and care quality. Conventional NEDOCS‐based approaches are largely reactive and provide limited support for proactive decision‐making. Methods We conducted a retrospective observational study using ED operational and clinical data from the Clinical Data Warehouse of Samsung Medical Center (2017–2025). A total of 292,033 ED visits were included. Time‐series datasets with 19 variables were constructed across five forecasting horizons ( t + 1 to t + 5). XGBoost‐based models were developed using two NEDOCS thresholds (≥141 and ≥ 101) with strict temporal validation. Model performance was evaluated using sensitivity, F1‐score, balanced accuracy, and AUC‐ROC. Model interpretability was assessed using SHAP. Results The nedocs1 model showed stable but conservative prediction behavior with low sensitivity (0.14–0.29). In contrast, the nedocs2 model demonstrated improved sensitivity (0.39–0.48) and F1‐score, particularly at t + 3. SHAP analysis revealed that waiting time and patient volume were key drivers, with increased contributions from hospitalization and acuity‐related variables at longer horizons. Conclusion Explainable time‐series machine learning enables early prediction of ED crowding with interpretable insights. The nedocs2 model showed superior performance for proactive operational alerting and may support timely resource allocation in emergency care settings.
Lee et al. (Wed,) studied this question.