Practical comparison of aberration detection algorithms for biosurveillance systems

Area of Science:

• Public Health Surveillance
• Epidemiology
• Biostatistics

Background:

• National syndromic surveillance systems are vital for public health.
• Effective anomaly detection methods are needed to identify disease outbreaks early.
• Existing methods require evaluation for optimal performance.

Purpose of the Study:

• To compare the performance of various anomaly detection methods for syndromic surveillance.
• To assess the sensitivity and timeliness of different statistical models in detecting multi-day public health signals.
• To identify strategies for improving anomaly detection in syndromic data.

Main Methods:

• Injected simulated multi-day signals into U.S. Centers for Disease Control and Prevention BioSense data for rash, respiratory, and gastrointestinal syndromes.
• Compared five control chart adaptations, a linear regression model, and a Poisson regression model.
• Evaluated methods based on sensitivity and timeliness at different alert rates, considering stratification and data normalization.

Main Results:

• Sensitivity ranged from 24-77% and timeliness from 3.3-6.1 days at 1-2% daily alert rates.
• Stratification by weekday/weekend/holiday and adjusting for total facility visits substantially improved performance.
• Linear regression generally outperformed control charts; Poisson regression showed best sensitivity for high-count data.

Conclusions:

• Stratification and adjusting for total visits are effective strategies to enhance syndromic surveillance anomaly detection.
• Method choice, particularly Poisson regression for high-count data, impacts detection accuracy.
• Further optimization of anomaly detection methods is essential for timely public health response.