Abstract
OBJECTIVE: This study aims to develop and evaluate a transformer-based neural network model that leverages both vital signs and chief complaints to predict patient acuity more accurately and automatically. This study is a part of a major project, which envisions continuous monitoring in the emergency department, while this model provides a machine learning based tool to risk stratify patients. METHODS: This study utilized the public MIMIC-IV-ED dataset, containing patients' vital signs, chief complaints, and triage acuity levels. We developed multiple machine learning models, including a baseline model using only vital signs and a hybrid model that contains the transformer architecture and feed-forward neural networks, incorporating numerical and textual data types. A secondary analysis was performed after filtering inconsistent data points to test the model in an idealized scenario. RESULTS: Models incorporating chief complaints achieved significantly higher accuracy (over 70%) compared to baseline models that relied solely on vital signs (around 60%). After filtering out inconsistent data, the hybrid model's accuracy improved to over 90%. CONCLUSION: Integrating chief complaints is critical for improving the accuracy of AI-driven triage models. These findings highlight the potential for hybrid systems to enhance patient monitoring and prevent deterioration in ED waiting rooms. Future work should focus on incorporating more diverse datasets and time series data to further validate and improve model performance.