Abstract Emergency department (ED) overcrowding and inefficient patient flow are significant operational challenges, often amplified by the volume of patients sustaining fall‐related injuries. This study predicts hospital admission for these encounters by leveraging unstructured clinical text: triage chief complaints (reasons for seeking care) and final emergency physician (EP) diagnoses. The primary objective is to develop a clinical decision support system that provides early, evidence‐based disposition predictions, serving as an operational alert mechanism to assist in the admission decision‐making process. We developed a “premodel” using triage data and a “postmodel” incorporating EP diagnoses. Analyzing 14,433 encounters, results across 12 algorithms demonstrated significant improvement when EP diagnoses were included. For top‐performing ensemble models, the receiver operating characteristic area under the curve—a measure of diagnostic accuracy—increased from 0.83 to 0.92, quantifying the value of integrating advanced natural language processing into this AI‐enhanced system. To foster clinician trust, we employed model interpretability techniques, specifically local interpretable model‐agnostic explanations and SHapley Additive exPlanations. These methods explain individual predictions by identifying influential clinical terms such as “fracture” and “head injury.” These high‐accuracy predictions provide critical inputs for future bed allocation models. This research underscores the potential of integrating text mining and machine learning to provide actionable insights that assist physician decision‐making and streamline ED outflow. Future research should focus on multi‐center external validation to assess generalizability across diverse healthcare settings.
Pai et al. (Tue,) studied this question.