Every year, millions of people are threatened by mosquito-borne diseases such as dengue, chikungunya, Zika, and malaria. Thanks to computational advancements, we can now create agent-based computer models with unprecedented biological realism to study how these diseases spread. While these complex models give us valuable insights, they are often outperformed in disease forecasting by their simpler mathematical counterparts.
One reason for this could be that we still do not fully understand how to accurately model the spatial aspects of disease spread. Currently, in-depth sensitivity analysis for these complex and parameter-rich agent-based models that would help to address this problem are hardly conducted because they tend to be computationally intensive. In this project, we leveraged Gaussian processes (GPs) as a powerful statistical framework to analyze complex agent-based models more efficiently. This in turn allows us to understand which parameters in the agent-based disease transmission models are truly shaping disease dynamics.
We developed an agent-based model to examine disease transmission across different scenarios involving human movement, social structures, and seasonality in infection risk. By training GPs on the simulation outcomes, we were able to build surrogate models that can quickly predict outbreak probability, epidemic severity, and epidemic duration. This approach also enabled comprehensive sensitivity analyses, revealing that human mobility and infection probability after contact are the primary drivers of observing epidemic outbreaks, while seasonality and the timing of disease introduction shape epidemic patterns. When applied to 12 years of empirical dengue incidence data, the GPs identified potential dengue transmission hotspots pointing to areas for further investigation.
In conclusion, this project demonstrates the potential of GPs to efficiently analyze complex epidemiological agent-based models. By using GPs, we gained valuable insights into the key drivers of epidemic dynamics, particularly the interactions between infection probability, human mobility, and seasonality. Importantly, when applied to empirical data, GPs also identified areas that could be important disease transmission hotspots. Our work highlights both the possibilities and challenges of using GPs to explore complex agent-based models, laying the groundwork for more realistic and computationally efficient models in the future.
