A new prior for bayesian anomaly detection: application to biosurveillance

Area of Science:

• Epidemiology
• Biostatistics
• Computer Science

Background:

• Bayesian anomaly detection relies on prior probabilities, which can be difficult to specify.
• Effective anomaly detection is crucial for timely disease outbreak identification.

Purpose of the Study:

• To develop a meaningful and user-friendly Bayesian prior for disease outbreak detection.
• To achieve detection performance comparable to or exceeding standard frequentist methods.

Main Methods:

• A semi-informative prior probability for anomalous event patterns was developed.
• A univariate Bayesian anomaly detection algorithm was presented for clarity.
• The prior was derived from historical data to estimate baseline behavior.

Main Results:

• The Bayesian method demonstrated statistically significant improvements over a control chart method (frequentist) in disease outbreak detection.
• Performance was optimal when baseline periods avoided seasonal effects.
• The Bayesian algorithm exhibits linear time complexity due to a novel closed-form derivation.

Conclusions:

• A novel, expressive, and efficient prior probability for Bayesian outbreak detection was introduced.
• The proposed method is easy to apply and computationally efficient.
• The Bayesian approach matches or surpasses the performance of standard frequentist methods.