Epidemic forecasts during outbreaks of emerging infectious diseases (EIDs), such as Zika, could be useful if they were available in real time. Without historical epidemiological data available to train forecasting models for novel EIDs, producing actionable forecasts models that can be used by decision makers in real time remains a challenge. To account for this lack of historical data, we used a semi-mechanistic TSIR model tailored to vector-borne pathogen transmission to account for susceptible depletion, human mobility (from aggregated and anonymized call data records), and stochasticity over the course of the epidemic and assessed its feasibility as a real-time forecasting tool in the context of the 2015-2017 Zika epidemic in Colombia. We incorporated previously estimated temperature-R0 relationships for dengue as prior distributions for spatial R0 values in our model, which we updated, along with other parameter values, over the course of the epidemic. We used a Bayesian data assimilation algorithm that iteratively proposed new model fits, made forecasts under different model assumptions, and proposed new model fits based on new data. We then assessed forecast accuracy using forecasting targets, such as peak incidence, and posterior parameter estimates using independent epidemiological estimates, such as timing of the introduction of Zika virus to Colombia.
Results/Conclusions
The spatial and temporal accuracy of forecasts depended both on which model we used and the magnitude of the observed epidemic in each department, resulting in a trade-off between more accurate forecasts in departments with large epidemics and somewhat less accurate forecasts in departments with small outbreaks. The inclusion of human mobility data into our spatial model allowed for the forecasting algorithm to accurately predict when the first Zika cases would occur in many departments weeks before they actually occurred, which could provide actionable results to decision makers if implemented in real time. Our posterior parameter estimates of the timing of the introduction of Zika virus to Colombia align with independent estimates from phylogenetic analyses. Our results indicate that actionable, real-time forecasting would have been feasible in the context of the 2015-2017 Zika epidemic in Colombia, especially early in the epidemic in departments with high suitability for mosquito-borne disease transmission. The capacity of our forecasting model to update model parameters, such as R0, in real time has implications for forecasting EIDs more generally.