JMIR Publications

ABSTRACT

Background:

Triage is the process of accurately assessing the patient's symptoms and providing them with proper clinical treatments in the emergency department (ED). While many countries have developed their triage process to stratify patient’s clinical severity and thus distribute medical resources, there is still some limitation of the current triage process. Since triage level is mainly performed by the experienced nurse based on the mix of subjective and objective criteria, mis-triage often happens in the ED. It can not only cause adverse effects on patients but also impose an undue burden on the healthcare delivery system.

Objective:

Our study aims to design a prediction system based on the triage information, including demographics, vital signs, and chief complaints. The proposed system can not only handle the heterogeneous data including tabular data and free-text data, but also have interpretability for better acceptance by the ED’s stuff in the hospital.

Methods:

In this study, we have proposed a system comprising three subsystems, and each of them handles one mission including triage level prediction, hospitalization prediction, and length of stay prediction. We use a large amount of retrospective data to pre-train the model, and then fine-tune the model on the prospective dataset with golden label. The proposed deep learning framework is built by TabNet and MacBERT.

Results:

The performance of our proposed model is evaluated on our own collected data in NTUH, with 902 patients are included. The models achieved promising results on our collected dataset, with 63%, 82%, and 71% accuracy on triage level, hospitalization, and length of stay prediction, respectively.

Conclusions:

Our system demonstrates an improvement compared to other machine learning methods in three different medical outcomes. With the pretrained vital sign encoder and the re-pretrained MLM MacBERT encoder, our multi-modality model can gain a deeper insight into the character of the EHRs. Additionally, by providing interpretability, we believe the proposed system can assist the nursing staff and the physician in making their medical decisions.