Fig 1.
(A) WNV compartmental model used in the analyses. (B) Example of a simulated time series with an upcoming epidemic, simulated by increasing R0 from 0.7 to 1. The shaded region indicates where R0 is above 1. (C) Example of a simulated time series with no upcoming epidemic, simulated with a stable R0 = 0.8. In (B) and (C), the time series represent a single simulation and were aggregated per month to improve the readability of the figure, but the daily time series are used in the analyses.
Fig 2.
(A) Number of infected birds A over time, for two stochastic simulations as examples of the perturbation-recovery experiments in a case of high (red) and low (blue) resilience. The points represent the observations of infected birds over time, and the dotted lines indicate the fitted lines used to calculate the return rate to the disease-free state (see Methods). (B) Average recovery time (solid line) and 95% confidence interval (dotted line) (in days) to the disease-free state after perturbing the system by introducing infected birds for different values of R0. (C) Average recovery time (solid line) and 95% confidence interval (dotted line) (in days) to the disease-free state after perturbing the system by introducing infected mosquitoes for different values of R0. For B and C, the recovery time is defined as the reciprocal of the return rate, assuming exponential decay, and is calculated over 1000 replicates.
Fig 3.
Prediction performance of the different univariate and multivariate indicators of resilience evaluated using the AUC.
(A) Prediction performance of the univariate indicators for all univariate time series. (B) Prediction performance of the multivariate indicators for the Anthro-equine scenario. (C) Prediction performance of the multivariate indicators for the wildlife scenario. (D) Prediction performance of the multivariate indicators for the hidden reservoir scenario.
Fig 4.
Performance of the best-performing, variance-based indicators of resilience in data-poor scenarios.
(A) The resolution is reduced by downsampling the original time series. (B) The observation probability is reduced by subsampling the number of infected using a Poisson distribution. (C) Effect of the reduction of resolution on the prediction performance. (D) Effect of the reduction of observation probability on the prediction performance.
Fig 5.
Prediction performance of the different species depending on how much they get infected.
(A) Effect of varying mosquito feeding preferences towards the hidden reservoir on the prediction performance. (B) Prediction performance of the univariate time series depending on how much the species get infected, quantified using the typical infection coefficient (i.e., the corresponding coefficient of the eigen vector associated with the dominant eigenvalue of the NGM).