A0948
Title: Particle filtering models of pre-vaccine pertussis: Prediction, scenarios, outbreaks
Authors: Nathaniel Osgood - University of Saskatchewan (Canada) [presenting]
Xiaoyan Li - Topos Institute (United Kingdom)
Abstract: Particle filtering (PF) is a contemporary sequential Monte Carlo method supporting statistical filtering of nonlinear models with non-Gaussian measurement and process noise, given a time series of empirical observations. Previous research has demonstrated the effective application of PF to low-dimensional compartmental transmission models. Implementation and evaluation of PF application is demonstrated for more complex compartmental transmission models for B. pertussis, including models involving 1-32 age groups and with two distinct functional forms for contact matrices, using over 35 years of monthly and annual pre-vaccination data from the Canadian province of Saskatchewan. Following the evaluation of the predictive accuracy of these model variants, prediction, intervention scenario evaluation, and outbreak classification analysis are then performed based on the most accurate model. It is found that PF with relatively high-dimensional pertussis transmission models and reported case counts can support effective prediction of pertussis outbreak evolution and classification of outbreak occurrence in the next month (AUC 0.91) in the context of even aggregate monthly incoming empirical data. PF models are further found valuable in performing counterfactual analysis of interventions. With its grounding in an understanding of disease mechanisms and a representation of the latent state of the system, this technique further offers high explanatory value.
