Predictive accuracy of particle filtering in dynamic models supporting outbreak projections

Area of Science:

• Epidemiology
• Computational Statistics
• Mathematical Modeling

Background:

• Dynamic models require recurrent regrounding with new data, but effective configurations are under-explored.
• Computational statistics and data streams offer potential for real-time model updates.

Purpose of the Study:

• To explore the effectiveness of particle filtering for automatically regrounding dynamic models with incoming observational data.
• To evaluate various particle filtering configurations for accuracy in predicting future outbreak prevalence.

Main Methods:

• Employed dynamic models combined with particle filtering (condensation algorithm).
• Assumed observed incident cases followed a negative binomial distribution, updating particle weights with observations.
• Evaluated prediction accuracy across varying observation frequencies, negative binomial parameters, and contact rate evolution.

Main Results:

• Increased frequency of empirical data observations led to super-linear improvements in prediction accuracy.
• Optimal assumptions for negative binomial parameters and contact rate changes were robust across different observation frequencies and outbreak stages.

Conclusions:

• Particle filtering effectively projects outbreak evolution, with significant accuracy gains from frequent sampling.
• Results suggest potential for standardized guidelines for particle filtering in epidemiology.
• Combining basic models with particle filtering can provide strong guidance for anticipating infectious disease outbreaks.