A Bayesian Network for Outbreak Detection and Prediction

Xia Jiang, Garrick L. WallstromAAAI 2006

Abstract

Traditional disease outbreak detection methods often provide simple alerts without context. This paper introduces a Bayesian network (BN) designed to predict an outbreak's size, duration, and current progress. By modeling the causal relationships between latent outbreak characteristics and observable proxy data, the system offers public health officials actionable intelligence for more informed decision-making.

Introduction: The Need for Predictive Intelligence

Public health officials require early and detailed information when an outbreak unfolds. Traditional detection methods, such as monitoring daily counts of over-the-counter (OTC) drug sales or clinic visits, typically rely on classical time-series analysis like ARIMA. However, these methods are "one-dimensional": they signal that an anomaly exists but fail to estimate its magnitude, duration, or current stage. Without these details, making high-stakes decisions—such as issuing a city-wide boil water alert—becomes difficult due to significant economic and social costs.

Comparison of new cases (a) and weekly OTC drug sales (b) during an outbreak.

Methodology: The Bayesian Network Architecture

The proposed solution is a Bayesian Network (BN) designed to reason under uncertainty by modeling the causal and probabilistic relationships between several key variables.

The Bayesian Network structure for outbreak detection and prediction.

The BN architecture incorporates:

  • Latent Variables: Outbreak start date ($O$), duration ($D$), and size ($S$).
  • Observable Variables: Total counts ($TC$) of a proxy variable, which is the sum of outbreak-related cases ($OC$) and background noise ($BC$).
  • Contextual Variables: Day of the week, cyclical influences, and hidden common causes ($H$).

A key innovation is the "general shape" assumption for epidemic curves (epi curves). The model assumes that while outbreaks vary in scale, they follow a consistent functional form. The authors developed a formula to scale a "standard" epi curve based on specific size and duration parameters being tested by the network.

Data and Validation

The study utilized real-world data from Cryptosporidium outbreaks in North Battleford, Saskatchewan (2001) and Milwaukee, Wisconsin (1993) to validate the "general shape" assumption. For testing, realistic outbreaks were injected into background OTC sales data from Cook County, Illinois (2004) using the HIFIDE simulator.

The model leverages expert knowledge—such as the historical frequency of specific outbreaks—as probabilistic priors. This helps the system distinguish between random data jitters and actual emerging crises.

Key Results

The "BayesNet" model was evaluated against classical methods like ARIMA and EWMA:

  • Detection Speed: For a 10% size outbreak, BayesNet detected the event in an average of 13.50 days, performing competitively with or better than classical algorithms.
  • Predictive Accuracy: Estimation of size ($S$) and duration ($D$) improved as the outbreak progressed. By day 21 of a 47-day outbreak, size estimates for 20% and 40% outbreaks were remarkably accurate.
  • Early Estimation: Even by day 13, the system provided rough scale estimates, offering officials a significant head start on resource planning.

Limitations and Future Work

While promising, the model has limitations. In early stages, small outbreaks can be confused with the initial ramp-up of larger ones because their peaks look similar to the start of a major event. Additionally, the current version is computationally intensive and limited to a single observable variable. Future research will focus on multivariate time series (combining OTC sales, hospital visits, and school absences) and learning more flexible epi curve representations directly from data.

Summary updated

BibTeX
@inproceedings{jiang2006bayesian,
  author = {Xia Jiang and Garrick L. Wallstrom},
  title = {{A Bayesian Network for Outbreak Detection and Prediction}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2006},
}
Figure