# EDT-MCFEF: a multi-channel feature fusion model for emergency department triage of medical texts

## TL;DR

This paper introduces EDT-MCFEF, a new AI model that improves emergency department triage by analyzing medical texts more accurately and efficiently than existing methods.

## Contribution

The novel EDT-MCFEF model combines multi-channel feature extraction and fusion for enhanced triage of medical texts.

## Key findings

- EDT-MCFEF outperformed seven benchmark models on two medical text datasets.
- The model handles imbalanced datasets effectively and shows strong generalization and robustness.
- It provides favorable interpretability and supports data privacy through its design.

## Abstract

Triage is a pivotal function within the operational framework of an emergency department, as it directly influences patient outcomes and hospital efficiency. However, traditional triage methods frequently depend on human judgment, which is susceptible to high subjectivity and low efficiency.

To address these issues, this paper presents a novel emergency department triage algorithm. The proposed EDT-MCFEF (Emergency Triage Algorithm Based on Multi-Channel Feature Extraction and Fusion) addresses numerous shortcomings of conventional triage methodologies. The model employs a hybrid masking approach and RoBERTa (Robustly Optimized BERT Approach) to facilitate feature enhancement and word vector processing of text. Moreover, the model employs a convolutional neural network (CNN) and a multi-headed attention (MHA) mechanism to extract text features from multiple channels, effectively capturing both local and global features. Furthermore, this paper introduces a multi-channel feature fusion method, which integrates local and global features and achieves comprehensive learning and optimization of feature information through dynamic weight adjustment.

The objective of this model is to enhance the accuracy and efficiency of emergency department triage, thereby providing scientific and technological support to the emergency department. In this paper, two medical text datasets are employed for experimental validation: a self-built emergency department triage dataset and a medical literature abstract dataset. The emergency department triage dataset consists of 28,000 English-annotated samples from 11 clinical departments, while the medical literature abstract dataset is a publicly available dataset (https://huggingface.co/datasets/123rc/medical_text). The experimental findings demonstrate that the proposed model exhibits superior accuracy to seven benchmark models utilized in this study on both medical text datasets, indicating its efficacy in handling imbalanced datasets. This suggests enhanced generalization and robustness. In addition to its strong classification ability, the model exhibits favorable interpretability through its multi-channel design, and the hybrid masking strategy supports data minimization and privacy protection, aligning with ethical AI principles. This approach holds promise for integration into clinical decision support systems for improved triage accuracy. The models and the self-built dataset presented in this paper are available at https://github.com/Yiii-master/EDT-MCFEF.