Explainable AI-Assisted Triage in Emergency Departments Using Multi-Modal Clinical Data

Overcrowding in emergency departments (EDs) increases waiting times and triage errors, straining healthcare systems and compromising patient safety. Existing AI-assisted triage systems primarily rely on structured data (e.g., vital signs, demographics), while neglecting unstructured information such as clinical text and images. To address this gap, we propose an explainable AI-assisted decision support system that integrates multi-modal clinical data. Using the Korean Triage and Acuity Scale (KTAS) dataset, we combine structured features with unstructured patient assessment text in a multimodal framework. Nurse-assigned scores are included to capture clinical judgement, with expert-reviewed triage levels as ground truth. After preprocessing, multiple classifiers were evaluated; Random Forest and AdaBoost achieved the highest performance(F1-scores: 89% and 87%; AUCs: 95.3% and 95.6%). RandomForest was selected as the final model and deployed via an explainable AI-powered web-based interface (Flask, HTML/CSS/JS,Docker). Findings demonstrate that explainable machine learning can improve ED triage accuracy, support clinical decision-making,and enhance transparency.