Improving triage performance in emergency departments using machine learning and natural language processing: a...

Area of Science:

• Emergency Medicine
• Artificial Intelligence
• Data Science

Background:

• Emergency Department (ED) triage is vital for patient care prioritization, traditionally using the Manchester Triage Scale (MTS).
• Conventional triage methods are susceptible to human error, leading to under- or over-triage and inconsistent patient classification.
• Machine Learning (ML) and Natural Language Processing (NLP) offer potential solutions to improve ED triage accuracy and consistency.

Purpose of the Study:

• To systematically review and analyze studies on the application of ML and/or NLP algorithms for ED patient triage.
• To assess the performance and limitations of various ML/NLP models in classifying patient severity in emergency settings.

Main Methods:

• A systematic review was conducted following PRISMA guidelines across five major scientific databases.
• Studies published from database inception to October 2023 were included, focusing on ML/NLP methods for triage classification.
• The Prediction model Risk of Bias Assessment Tool (PROBAST) was used to evaluate the risk of bias in included studies.

Main Results:

• Sixty studies involving 57 ML algorithms were analyzed, with Logistic Regression being the most common.
• eXtreme Gradient Boosting (XGBoost), Gradient Boosting (GB) decision trees, and Deep Neural Networks (DNNs) demonstrated superior performance.
• Key predictive variables included demographics, vital signs (oxygen saturation, systolic blood pressure), chief complaints, age, and mode of arrival; however, significant bias risk was noted in classification models.

Conclusions:

• NLP integration enhanced ML algorithms' classification accuracy by utilizing nursing notes and structured clinical data.
• Feature engineering and class imbalance correction improved ML workflow performance, though Feature Engineering and Explainable AI (XAI) remain underexplored.
• This systematic review is registered in PROSPERO (registration number: CRD42024604529) and funded by CNPq, Brazil.