Near Real-Time Syndromic Surveillance of Emergency Department Triage Texts Using Natural Language Processing: Case...

Area of Science:

• Medical informatics
• Public health surveillance
• Artificial intelligence in healthcare

Background:

• Adverse event detection post-immunization is crucial for public safety.
• Febrile convulsions are a known vaccine reaction in young children.
• Early identification of safety signals aids prompt intervention.

Purpose of the Study:

• To assess natural language processing (NLP) and machine learning (ML) for detecting febrile convulsions in emergency department (ED) triage notes.
• To determine optimal training data needs for ML models in this context.
• To evaluate real-time monitoring deployment requirements for ML models.

Main Methods:

• Developed a baseline pattern matching approach.
• Evaluated various ML models for febrile convulsion classification in ED triage notes.
• Compared model performance during training and on live ED data.

Main Results:

• Transformer-based models significantly outperformed standard neural networks.
• Transformer models demonstrated superior effectiveness in detecting febrile convulsions.
• The best models were resource-efficient and suitable for deployment.

Conclusions:

• NLP, especially large language models (LLMs), offers substantial benefits for syndromic surveillance.
• LLMs serve as highly effective classifiers for identifying adverse events.
• LLM text generation capabilities can improve training data quality and diversity.