JMIR Publications

ABSTRACT

Background:

Emergency Department (ED) crowding has resulted in delayed patient treatment and has become a universal healthcare problem. Although a triage system, such as the five-level Emergency Severity Index (ESI), somewhat improves the process of ED treatment, it still heavily relies on the nurse's subjective judgment and triages too many patients to ESI level 3 in current practice. Hence, a system that can help clinicians to accurately triage a patient's condition is imperative.

Objective:

This study aimed to develop a deep-learning-based triage system, using patients’ ED electronic medical records to predict clinical outcomes after ED treatments.

Methods:

We conducted a retrospective study using the data from open dataset from NHAMCS (National Hospital Ambulatory Medical Care Survey) from 2012 to 2016 and data from local dataset from NTUH (National Taiwan University Hospital) from 2009 to 2015. In this study, we transformed structural data into text form and used Convolutional Neural Networks combined with Recurrent Neural Networks and attention mechanism for accomplishing the classification task. We evaluated our performance using area under the receiver-operating characteristic (AUROC).

Results:

A total of 118,602 patients from NHAMCS were included in this study, in predicting hospitalization, the accuracy and AUROC achieved 0.83 and 0.87, respectively. On the other hand, the external validation was to use our own dataset from NTUH that included 745,441 patients, the accuracy and AUROC are similar to those mentioned above, namely, 0.83 and 0.88, respectively. Moreover, in order to effectively evaluate the prediction quality of our proposed system, we also applied the model on other clinical outcomes including mortality and admission to ICU, and the results showed that our proposed method is 3 ~ 5% higher in accuracy than other conventional methods.

Conclusions:

Our proposed method achieved better performance than the traditional method, and the implementation is relatively easy, including commonly used variables, and better fitting for real-world clinical setting. It is our future work to validate our novel deep-learning based triage algorithm with prospective clinical trials, and hope to use it to guide resource allocation in a busy ED once the validation succeeds.