Crowding in real-time: an artificial intelligence-based model to manage crowding in the emergency department (CRAMMED Study).

Abstract

Background: Emergency department (ED) crowding is a global challenge with adverse effects on patient outcomes and staff well-being. Traditional crowding scores, such as the National ED Overcrowding Study (NEDOCS) score, have limited ability to predict imminent crowding. The aim of this study is to develop and validate a real-time, artificial intelligence (AI)-driven prediction model using a long short-term memory (LSTM) neural network to predict ED crowding, using ambulance diversion status as the reference standard and comparing performance with the (modified) NEDOCS score.

Methods: In this single-centre, retrospective cohort study, we extracted parameters related to crowding from electronic health records and workforce data. Ambulance diversion status served as the reference standard for crowding. The LSTM model was trained and evaluated with walk-forward validation and performance was assessed using the area under the receiver operating characteristic curve (AUC). Model calibration was assessed using Brier score and expected calibration error (ECE).

Results: The optimal LSTM model used the previous hour of data to predict one hour ahead, yielding an AUC of 0.84 (95% CI 0.81-0.86). Key features contributing to the predictive quality of our model included the number of pre-notified arrivals, patients in the yellow and orange triage categories, and the number of ED supervisors. Compared with the modified NEDOCS score (AUC of 0.73 (95% CI 0.70-0.76)), the LSTM model showed higher discriminative performance in internal validation within this single-centre study.

Conclusions: Our AI-driven prediction model demonstrated good predictive performance for ED crowding in internal validation. These findings highlight the potential of real-time prediction models to support ED clinical workflow.