Multivariate resilience indicators to anticipate vector-borne disease outbreaks: a West Nile virus case-study

Abstract

Background and aim: To prevent the spread of infectious diseases, successful interventions require early detection. The timing of implementation of preventive measures is crucial, but as outbreaks are hard to anticipate, control efforts often start too late. This applies to mosquito-borne diseases, for which the multifaceted nature of transmission complicates surveillance. Resilience indicators have been studied as a generic, model-free early warning method. However, the large data requirements limit their use in practice. In the present study, we compare the performance of multivariate indicators of resilience, combining the information contained in multiple data sources, to the performance of univariate ones focusing on one single time series. Additionally, by comparing various monitoring scenarios, we aim to find which data sources are the most informative as early warnings. Methods and results: West Nile virus was used as a case study due to its complex transmission cycle with different hosts and vectors interacting. A synthetic dataset was generated using a compartmental model under different monitoring scenarios, including data-poor scenarios. Multivariate indicators of resilience relied on different data reduction techniques such as principal component analysis (PCA) and Max Autocorrelation Factor analysis (MAF). Multivariate indicators outperformed univariate ones, especially in data-poor scenarios such as reduced resolution or observation probabilities. This finding held across the different monitoring scenarios investigated. In the explored system, species that were more involved in the transmission cycle or preferred by the mosquitoes were not more informative for early warnings. Implications: Overall, these results indicate that combining multiple data sources into multivariate indicators can help overcome the challenges of data requirements for resilience indicators. The final decision should be based on whether the additional effort is worth the gain in prediction performance. Future studies should confirm these findings in real-world data and estimate the sensitivity, specificity, and lead time of multivariate resilience indicators.