Related Experiment Video
Updated: May 26, 2026
10:05
High-throughput Detection Method for Influenza Virus
Published on: February 4, 2012
Comparison of natural language processing biosurveillance methods for identifying influenza from encounter notes
Peter L Elkin1, David A Froehling, Dietlind L Wahner-Roedler
1Mount Sinai School of Medicine, New York, New York, USA. ontolimatics@gmail.com
Annals of Internal Medicine
|January 4, 2012
Summary
Biosurveillance using whole encounter notes significantly improves influenza detection accuracy compared to relying solely on chief complaints. This enhanced approach is crucial for effective public health outbreak recognition and response.
Area of Science:
- Public Health Surveillance
- Infectious Disease Epidemiology
- Health Informatics
Background:
- Effective national biosurveillance systems are vital for rapid outbreak detection and coordinated response.
- Current systems like BioSense at the Centers for Disease Control and Prevention (CDC) utilize chief complaints but not comprehensive encounter note data.
- This limitation may hinder the accuracy and timeliness of public health surveillance.
Purpose of the Study:
- To compare the effectiveness of biosurveillance using data from the entire patient encounter note versus using only the chief complaint field.
- To determine if incorporating complete clinical data enhances the accuracy of influenza outbreak detection.
Main Methods:
- A 6-year retrospective case-control cohort study was conducted at the Mayo Clinic.
- Data from 17,243 individuals tested for influenza A or B between 2000 and 2006 were analyzed.
- Natural language processing techniques were employed to extract clinical features from free-text encounter notes for model development.
Main Results:
- Biosurveillance utilizing the whole encounter note demonstrated superior accuracy over using the chief complaint field alone.
- The normalized partial area under the receiver-operating characteristic curve was 92.9% for the whole note model versus 70.3% for the chief complaint model (P < 0.001).
- At a specificity of 0.4, sensitivities were 89.0% for the whole note model and 74.4% for the chief complaint model (P < 0.001), with a relative risk of 2.3 for missing influenza cases with the latter.
Conclusions:
- A biosurveillance model incorporating data from the entire encounter note is significantly more accurate for influenza detection than one using only the chief complaint.
- Case-defining signs and symptoms for influenza are frequently documented in health records, supporting the use of comprehensive data.
- The national biosurveillance strategy should be updated to include data from the complete health record for improved accuracy and effectiveness.
