Interpretable Machine Learning for Early Detection of Critical Outcomes in the Emergency Department

Emergency departments (EDs) require effective approaches for quickly identifying patients at risk of critical outcomes (inpatient mortality or ICU admission within 12 hours). This study developed interpretable machine learning models using Gradient Boosting (GB) and Logistic Regression (LR) with 13 initial triage variables from the MIMIC-IV-ED database. GB performed slightly better than LR (Accuracy: 78.21% vs. 77.27%, AUROC: 0.887 vs. 0.863, AUPRC: 0.445 vs. 0.370). The Te2Rules method was used to extract 43 clinically interpretable rules from the GB model, with a overall fidelity of 98.90%. The use of the Rule Coverage Index (RCI) to categorize rules into high, medium, and low levels futher enhanced clinical utility. This study aims to strike a balance between predictive accuracy and interpretability, facilitating clinicians in early detection of critically ill presenting to the ED.