Particle filtering in compartmental projection models
MetadataShow full item record
Simulation models are important tools for real-time forecasting of pandemics. Models help health decision makers examine interventions and secure strong guidance when anticipating outbreak evolution. However, models usually diverge from the real observations. Stochastics involved in pandemic systems, such as changes in human contact patterns play a substantial role in disease transmissions and are not usually captured in traditional dynamic models. In addition, models of emerging diseases face the challenge of limited epidemiological knowledge about the natural history of disease. Even when the information about natural history is available -- for example for endemic seasonal diseases -- transmission models are often simplified and are involved with omissions. Availability of data streams can provide a view of early days of a pandemic, but fail to predict how the pandemic will evolve. Recent developments of computational statistics algorithms such as Sequential Monte Carlo and Markov Chain Monte Carlo, provide the possibility of creating models based on historical data as well as re-grounding models based on ongoing data observations. The objective of this thesis is to combine particle filtering -- a Sequential Monte Carlo algorithm -- with system dynamics models of pandemics. We developed particle filtering models that can recurrently be re-grounded as new observations become available. To this end, we also examined the effectiveness of this arrangement which is subject to specifics of the configuration (e.g., frequency of data sampling). While clinically-diagnosed cases are valuable incoming data stream during an outbreak, new generation of geo-spatially specific data sources, such as search volumes can work as a complementary data resource to clinical data. As another contribution, we used particle filtering in a model which can be re-grounded based on both clinical and search volume data. Our results indicate that the particle filtering in combination with compartmental models provides accurate projection systems for the estimation of model states and also model parameters (particularly compared to traditional calibration methodologies and in the context of emerging communicable diseases). The results also suggest that more frequent sampling from clinical data improves predictive accuracy outstandingly. The results also present that assumptions to make regarding the parameters associated with the particle filtering itself and changes in contact rate were robust across adequacy of empirical data since the beginning of the outbreak and inter-observation interval. The results also support the use of data from Google search API along with clinical data.
DegreeMaster of Science (M.Sc.)
CommitteeStanley, Kevin; Mondal, Debajyoti; Page, Andrew; Neudorf, Cory
Copyright DateDecember 2018
machine learning, particle filtering, system dynamics, projection, outbreaks