Quantifying the determinants of outbreak detection performance through simulation and machine learning.

Area of Science:

• Epidemiology
• Public Health Surveillance
• Computational Biology

Background:

• Effective disease outbreak detection is crucial for public health interventions.
• Existing biosurveillance algorithms have varying detection capabilities.
• Predicting detection performance aids in optimizing surveillance systems.

Purpose of the Study:

• To develop a probabilistic model for identifying and measuring factors influencing outbreak detection.
• To predict the detection performance of new and existing outbreaks.
• To enhance the configuration and efficiency of public health surveillance systems.

Main Methods:

• Simulated waterborne disease outbreaks overlaid on real-world gastroenteritis data.
• Analysis using biosurveillance algorithms with varied parameters.
• Development of a Bayesian network model using structure and parameter learning algorithms.
• Evaluation of model predictions via 5-fold cross-validation.

Main Results:

• The developed model accurately predicted performance metrics for common outbreak detection methods (accuracy > 0.80).
• Quantified the impact of outbreak characteristics (e.g., peak size) and biosurveillance algorithm parameters (e.g., alerting threshold, weekly pattern adjustment) on detection.
• Identified key drivers of detection performance in realistic surveillance scenarios.

Conclusions:

• A robust model was created to accurately predict disease outbreak detection influenced by outbreak and method characteristics.
• The model facilitates comparison of detection methods across diverse surveillance scenarios.
• Provides insights into critical outbreak and algorithm features driving detection performance, guiding surveillance system optimization.