Patient Risk Assessment in Emergency Departments using Ensemble Learning

Abstract

Efficient patient allocation within a hospital’s Emergency Department (ED) is crucial in ensuring prompt and appropriate medical intervention, especially under high patient load conditions. In this study, we developed an ensemble-based machine learning framework to predict patient risk status during triage, using a newly curated medical dataset consisting of fundamental vital sign features such as heart rate, blood pressure, temperature, respiratory rate, and oxygen saturation (SPO2). To overcome the inherent class imbalance problem commonly observed in clinical data, the Synthetic Minority Oversampling Technique (SMOTE) was employed, which significantly improved model stability and predictive accuracy. Various ensemble algorithms were implemented and compared, including Gradient Boosting Classifier (GBC), Light Gradient Boosting Machine (LGBM), XGBoost (XGB), and Extra Trees Classifier (ETC). Among these, the Gradient Boosting Classifier exhibited superior overall performance in accuracy (87.21%), recall (60.12%), F1 score (67.11%), and AUC (77.42%), demonstrating its robustness and reliability for medical risk prediction. The GBC Classifier further achieved the highest sensitivity (60.12%), indicating its strong ability to correctly identify high-risk patients while maintaining excellent specificity (94.72%) for accurately classifying low-risk cases. This approach represents a promising advancement toward intelligent, automated risk-level assessment systems that can improve decision-making and enhance the judicious allotment of resources in emergency healthcare settings.