Classifying free-text triage chief complaints into syndromic categories with natural language processing

Area of Science:

• Computational linguistics
• Public health informatics
• Artificial intelligence in medicine

Background:

• Medical data often exists as unstructured free-text, hindering automated analysis.
• Chief complaints are crucial for syndromic surveillance but require encoding for AI systems.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) application.
• To classify free-text chief complaints into syndromic categories for biosurveillance.

Main Methods:

• Implemented an NLP text classifier trained on 4,700 chief complaints.
• Evaluated classifier performance using 800 chief complaints from a separate dataset.
• Applied the system to monitor the 2002 Winter Olympic Games.

Main Results:

• Achieved high Area Under the ROC Curve values for various syndromic categories (e.g., Rash=1.0, Hemorrhagic=0.99, Respiratory=0.99).
• Precisely extracted lower respiratory complaints (0.97) and fever with lower respiratory complaints (0.96).

Conclusions:

• A trainable NLP text classifier can effectively extract data from free-text chief complaints.
• This technology shows promise for accurate and automated biosurveillance.