Tracking viral warnings : Resilience indicators to anticipate mosquito-borne disease outbreaks

Summary

There are only a few ways to prevent the spread of diseases by mosquitoes. Most interventions rely on mosquito population control, which is costly and requires planning in advance. Therefore, it is important to predict future outbreaks in time to plan measures accordingly. This thesis investigates the use of resilience indicators—generic, model-free metrics derived from the theory of critical slowing down—to predict such outbreaks. Resilience indicators have been applied across diverse fields, from climate science to mental health, suggesting their broad applicability. They are calculated in long, high-resolution time series such as case reports, and warn for a loss of resilience and thus an upcoming outbreak when simple statistical indicators such as variance or autocorrelation display an increase over time. In this thesis, a review of the use of resilience indicators for infectious diseases showed that such indicators have been applied successfully to anticipate upcoming epidemics, but large data requirements often limit their use in practice. After reviewing compartmental models of West Nile virus (WNV), serving as a case study throughout this thesis, simulated data under different monitoring scenarios were generated to study the use of resilience indicators to anticipate future outbreaks. This thesis suggests alternative monitoring scenarios to allow for an accurate use of resilience indicators when the large data requirements are not met: (i) bursts of data and (ii) multivariate indicators. Using so-called bursts of data, i.e short, high-resolution time series, allows for detecting a loss of resilience and thus an upcoming epidemic, without continuous monitoring, which can be costly. We tested this method in different systems, to anticipate the onset of a depressive episode, a population collapse in an ecosystem and climate tipping points, showcasing the extent of the findings. Similarly, multivariate indicators take advantage of the different data types that can be monitored for mosquito-borne diseases: infected mosquito pools over time, and cases in humans and other hosts such as birds and horses for WNV. We found that combining these different types of data into multivariate indicators gives a more accurate warning than using only one type of data (univariate indicators) before an outbreak starts. This finding was more significant in data-poor scenarios, when the resolution of the data is reduced for instance. Finally, we compared the performance of resilience indicators with different machine learning approaches for anticipation. We found that resilience indicators can predict future outbreaks more accurately than machine learning approaches do. By suggesting new data collection strategies and integrating resilience indicators into early-warning frameworks, this work offers a pathway to more timely and effective public health responses to future mosquito-borne disease epidemics.