Chief complaint classification with recurrent neural networks

Area of Science:

• Public Health Informatics
• Computational Epidemiology
• Machine Learning in Healthcare

Background:

• Syndromic surveillance systems monitor population health using emergency department data.
• Current keyword-based methods for classifying unstructured text in these records are limited.
• Automated classification using machine learning can enhance speed and accuracy in outbreak detection.

Purpose of the Study:

• To implement and compare recurrent neural network (RNN) models (LSTM, GRU) against traditional bag-of-words classifiers (MNB, SVM) for syndromic surveillance.
• To evaluate model performance in classifying emergency department records using discharge diagnoses and chief complaints.
• To identify which conditions are most amenable to surveillance based on chief complaint data alone.

Main Methods:

• Four machine learning models were implemented: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Multinomial Naive Bayes (MNB), and Support Vector Machine (SVM).
• Models were trained on 3.6 million de-identified emergency department records to predict diagnostic code groups.
• Performance was primarily assessed using the F1 score, comparing predictions from discharge diagnoses and chief complaints.

Main Results:

• LSTM achieved the highest performance using discharge diagnoses (F1 score > 96.00), outperforming other models.
• GRU demonstrated superior performance using chief complaints (F1 score = 47.38), while MNB with bigrams performed worst (F1 score = 39.40).
• Certain conditions, like alcohol-related disorders, were predicted more accurately than others, such as influenza, highlighting variability in classification ease.

Conclusions:

• Recurrent neural network models significantly outperform traditional bag-of-words classifiers for automated syndromic surveillance.
• Deep learning approaches offer substantial improvements in classifying unstructured text data for public health monitoring.
• Further research can optimize RNN models for more accurate and timely detection of health threats from emergency department chief complaints.