Abstract
Emergency departments (EDs) face persistent challenges related to overcrowding, limited resources, and delays in triage, which can compromise timely patient care. Traditional triage systems rely largely on subjective judgement and rule-based criteria, leading to variability in prioritisation and limited adaptability to complex patient presentations. Recent advances in deep learning (DL) have shown potential to support more consistent triage decisions by leveraging both structured clinical data and unstructured patient complaints. However, much existing research focuses primarily on predictive accuracy, with limited attention to deployment readiness within clinical workflows.
This study investigates the use of DL models for ED triage classification and examines their preparedness for operational use through an MLOps-informed prototype approach, rather than full clinical integration. Three DL architectures were developed and evaluated: a Bidirectional Long Short-Term Memory (BiLSTM) model, a Convolutional Neural Network (CNN), and a Transformer-based model. The models were trained on an anonymised triage dataset comprising 1,267 real patient records, expanded through controlled data augmentation to 50,000 records to address class imbalance. Performance was evaluated using state-of-the-art classification metrics.
All three models demonstrated strong predictive performance under simulated conditions. The CNN achieved the most favourable balance between accuracy, computational efficiency, and stability across triage categories, attaining an accuracy of 94.0%, a macro F1-score of 93.9%, and a Matthews Correlation Coefficient of 0.925 on the augmented dataset. To move beyond retrospective evaluation, the study presents a simulated triage-support prototype that illustrates how the selected model could be embedded within a reproducible and maintainable workflow informed by MLOps practices.
Although the models were evaluated using historical and augmented data rather than live clinical inputs, the findings demonstrate the potential of DL-based triage systems to support more consistent, data-informed prioritisation decisions. External validation using real-world clinical data is required before deployment in operational emergency care settings. The study contributes methodological and practical insights to support future research and responsible exploration of AI-assisted triage, particularly in resource-constrained environments.