Publikation
Developement of an AI-Supported Triage System for Safe Patient Allocation in Integrated Emergency Centers
Robert Simon Schiff; Natalie Kohler; Mareike Böckel; Ralf Möller; Mattis Hartwig; Sebastian Wolfrum
In: Abstracts of the 25th International Conference on Emergency Medicine (ICEM 2026). International Conference on Emergency Medicine (ICEM-2026), June 9-13, Hamburg, Germany, Germany, Pages 30-31, No. 46, ISBN 978-3-948023-45-4, Conventus Congressmanagement & Marketing GmbH, Jena, 6/2026.
Zusammenfassung
Background/Introduction: A core element of Germany's impending emergency care reform is the establishment of Integrated Emergency Centers (INZ) with a Central Assessment Unit. The use of validated initial assessment systems is crucial for safe patient steering to reduce overcrowding in Central Emergency Departments (EDs) by safely directing low-acuity patients to Outpatient Clinics (OPC). Objectives: This simulation study aims to present a new AI-supported hybrid initial assessment instrument, APONA-Triage. APONA (Assisted process optimization in the ED) is designed to proactively identify low-acuity patients to relieve the ED while simultaneously guaranteeing patient safety within the OPC. The features used are limited to established routine parameters, ensuring broad applicability.
Methods/Description: Predictive classification models were trained on 12,445 and validated on 3,734 (held-out) walk-in emergency patients from a university maximum-care provider, to predict case acuity (2024 data). Model training was based on parameters available at the time of initial assessment, including the routine CEDIS chief complaint, Manchester Triage System (MTS) Category, prior diagnoses, and reason for encounter. An optimal allocation strategy was determined using Receiver Operating Characteristic analysis with maximal Youden index, and was simulated using hybrid safety rulesets, including that high-risk patients (MTS 1+2) were always routed to the ED. Low-risk patients were allocated either at the OPC or the ED, per model predictions. Subsequent misallocation analysis refined the final APONA-Triage model by reducing the rate of intensive care interventions in the OPC group.
Results/Outcomes: APONA achieved high sensitivity (83.2%) and specificity (55.2%) for predicting outpatient treatment suitability at triage, using a maximum Youden-index of 0.24. By implementing APONA, the number of walk-in patients in the ED can be reduced by up to 42.7%. This includes 12.9% of patients initially referred to an OPC and required subsequent transfer to an ED. The rate of patients requiring immediate intensive care was low (0.25%).
Conclusion: Overall, our results suggest that APONA is a viable tool for INZs to significantly relieve EDs while maintaining a high patient safety. These findings underscore the necessity of physical and organizational proximity between ED and OPC to ensure optimal patient flow and safe recourse for misallocated patients.