A novel framework for calibration and prediction of stochastic compartmental transmission dynamic models of novel pathogens
Final Report Abstract
Infectious diseases remain an important cause of morbidity and mortality and have large direct and indirect economic costs. The timely, well-designed response of public health authorities is important when responding to the risk of emerging infectious diseases. The use of stochastic epidemic models has become increasingly common for public health planning, especially when addressing questions relating to the emergence and elimination of infectious diseases, when chance events may be most important. This DFG funded project established a framework allowing for epidemic predictions based on surveillance data. The surveillance data is linked to a computational model of the disease that describes all key mechanistic processes which drive disease dynamics. This linking process requires the identification of reasonable values for key epidemic parameters such as the basic reproductive number or the average duration of infectiousness. Our framework allows for precise estimation of these parameters and accurate quantification of their uncertainty. Based on this parametrization of the computational model, we are able to run forward simulation which allows us to forecast the future number of cases or time and intensity of the peak. Our framework is not disease specific and can be applied to various kinds of diseases. In order to test our framework, we participated the Centers for Disease Control and Prevention (CDC) Influenza Prediction Challenge where well established research groups compete to accurately forecast the dynamics of influenza each year. We completed the challenge and were the best performers amongst newcomers. While we only used the official CDC data on influenza to calibrate our model, most other modeling teams also used indirect web-based data sources. We determine that our model’s performance was likely to increase with use of these additional data. Therefore, we also did further research on additional data sources and were able to quantify the gain associated with leveraging Wikipedia search data. We determined that these data did help and made the novel observation that the frequency of observation matters. Using daily resolved data (instead of weekly, which has been nearly universally adopted by modeling groups), lead to significant additional gains. We are confident that both results will help us improve future challenge participations as well as facilitate further research within the community.
Publications
- (2018) Use of daily Internet search query data improves real-time projections of influenza epidemics. Journal of the Royal Society, Interface 15 (147)
Zimmer, Christoph; Leuba, Sequoia I.; Yaesoubi, Reza; Cohen, Ted
(See online at https://doi.org/10.1098/rsif.2018.0220) - (2019) Accurate quantification of uncertainty in epidemic parameter estimates and predictions using stochastic compartmental models. Statistical methods in medical research 28 (12) 3591–3608
Zimmer, Christoph; Leuba, Sequoia I.; Cohen, Ted; Yaesoubi, Reza
(See online at https://doi.org/10.1177/0962280218805780)
