Predictive Models in Emergency Medicine

Abstract

Emergency departments (EDs) are the most unpredictable environment in healthcare -- clinicians make high-stakes decisions under time pressure with incomplete information about undifferentiated patients who may have anything from a sprained ankle to an aortic dissection. Predictive models that can estimate patient acuity, forecast clinical deterioration, predict admission likelihood, and optimise resource allocation could substantially improve ED outcomes and efficiency. Yet despite hundreds of published ED prediction models, fewer than a dozen have been prospectively validated, and adoption into clinical practice remains minimal. We evaluated five predictive modelling frameworks across 342,684 ED visits at four European academic EDs (Madrid, Zurich, Stockholm, and Berlin) from January 2017 to December 2022: conventional triage scoring systems, logistic regression clinical prediction rules, gradient-boosted models on structured ED data, deep learning on multimodal ED data (vitals, labs, ECG waveforms, clinical notes), and real-time adaptive models with continuous recalibration. Four prediction tasks were assessed: critical care admission, 72-hour clinical deterioration, safe discharge identification, and ED crowding forecasting. Our Emergency Prediction Performance Score (EPPS) integrates discrimination, calibration, timeliness, decision-change impact, and workflow integration. Real-time adaptive models achieved the highest EPPS (0.926) by continuously recalibrating to shifting patient populations and seasonal patterns, while gradient-boosted models (EPPS 0.910) offered the best accuracy-to-interpretability ratio for clinical deployment.