Multimodal Deep Learning Boosts Emergency Triage Accuracy.

This study proposes a multimodal deep learning architecture designed to enhance the accuracy of emergency triage decisions. The model effectively integrates two distinct data types: textual descriptions of patients' main presenting complaints and numerical vital signs. By leveraging self-attention mechanisms, the architecture is capable of capturing intricate local and global relationships within and between these diverse features. The research addresses the critical need for comprehensive understanding of temporal structures and dependencies inherent in such complex patient data. The proposed model was developed and validated using a dataset of over 11,000 triage records from a Malaysian university hospital's emergency department. Experimental results indicate that the multimodal deep learning model significantly outperforms baseline models, showing improvements in accuracy, F1-score, and ROC AUC. This demonstrates its potential to provide more automated, efficient, and timely analysis of patient information, leading to more accurate prioritization of medical attention and potentially better patient outcomes.