Using Digital Surveillance Tools for Near Real-Time Mapping of the Risk of International Infectious Disease Spread: Ebola as a Case Study

Sangeeta Bhatia; Britta Lassmann; Emily Cohn; Malwina Carrion; Moritz U.G. Kraemer; Mark Herringer; John Brownstein; Larry Madoff; Anne Cori; Pierre Nouvellet

We used the ProMED/HealthMap data to perform a spatio-temporal analysis of the spread of the West African Ebola epidemic. We fitted a spatially explicit branching process model to the daily incidence data derived from ProMED/HealthMap feeds. The estimated model parameters were used to simulate the future outbreak trajectory over 4, 6 or 8 weeks. The model performs relatively well at short forecast horizon, i.e. up to two weeks. At a longer time scale, the model performance starts to deteriorate. A likely explanation for this is that our model assumes that transmissibility remains constant over the entire projection window. This assumption may not hold as we project over longer horizons, for example due to interventions being implemented. Model performance was also highly dependent on the phase of the epidemic in which projections were made. During the growing phase, the model tended to over-predict. This is likely due to interventions implemented throughout the growing phase to reduce transmissibility, leading to a reduced observed incidence compared to our model's expectation. In the declining phase on the other hand, our model tended to under-predict the observed incidence. This could be due to control efforts being relaxed too early as case numbers decline. Another contributing factor could be super-spreading, a phenomenon observed in many epidemics including Ebola epidemics [29, 30] whereby a small number of cases generate a large number of secondary infections, implying that when case numbers are small, epidemic trajectory may be difficult to predict and not well described by Poisson likelihood. Models using more complex likelihoods, e.g. using Negative Binomial distributions, 10 should be explored in future work, but will present additional challenges as analytical results underpinning the estimation of the reproduction number will no longer hold [20] . Such variability in model performance throughout an epidemic could have important implications if the model predictions are used to inform resource allocation. Model estimates should therefore be interpreted with caution, particularly as an outbreak is observed to be declining, and if the forecast horizon is long.