Emergency Medical Text Classifier: New system improves processing and classification of triage notes

Area of Science:

• Public Health
• Medical Informatics
• Computational Biology

Background:

• Automated syndromic surveillance is crucial for early disease outbreak detection.
• Emergency Department (ED) data, including triage notes, offers valuable real-time information.
• Existing classification methods require enhancement for improved accuracy and efficiency.

Purpose of the Study:

• To develop and evaluate an improved automated system for classifying ED records into syndromic categories.
• To leverage the vector space model and pseudo-relevance feedback for enhanced classification accuracy.
• To support near real-time syndromic surveillance and disease outbreak early warning systems.

Main Methods:

• Utilized a vector space model with a pseudo-relevance feedback learning module.
• Constructed reference dictionaries using terms from standard syndrome definitions.
• Classified ED records based on cosine similarity, iteratively refining with feedback terms.
• Tested on Gastro-Intestinal (GI), Respiratory (Resp), and Fever-Rash (FR) syndromes across two datasets.

Main Results:

• The system demonstrated high sensitivity (Se) and specificity (Sp) across all tested syndromes.
• Results for Test Set 1: GI (Se 90%, Sp 71%), Resp (Se 97%, Sp 73%), FR (Se 100%, Sp 87%).
• Results for Test Set 2: GI (Se 88%, Sp 69%), Resp (Se 87%, Sp 61%), FR (Se 97%, Sp 71%).

Conclusions:

• The developed system significantly improves the syndromic classification of ED records.
• Pseudo-relevance feedback enhances classification accuracy (Se and Sp).
• The system's performance can be tuned to meet specific user requirements.