In spite of many medical breakthroughs in the control of infectious diseases, the emergence of novel pathogens continues to pose a threat to both human and animal populations. Concerns for humans identified in recent years include new coronavirus strains capable of causing global epidemics similar to SARS in 2003 and the potential for the evolution of strains of pandemic-causing influenza virus. Forewarning of emergence has clear benefits. Ultimately, the goal is to determine whether the novel pathogens’s transmissibility is approaching the invasion threshold, beyond which it is capable of causing large scale outbreaks.
There has been a growth in interest, spearheaded in part by the ecology community, in early-warning signals: model-free methods by which critical transitions may be anticipated. Our work draws on this literature, developing two parallel approaches appropriate for emerging disease systems. The first is based on early-warning signals, the second is based on a likelihood ratio test.
An enduring challenges for early-warning signals has been moving from measurements to decisions on whether a transition is being approached. To address this problem, we apply a genetic algorithm — an efficient method for complex stochastic optimization—to simulated data to identify the optimum criteria.
We derive a set of early-warning signals by making use of the analytically-tractable stochastic Birth-Death-Immigration (BDI) process, the simplest possible model of emerging diseases with direct transmission. We also use the BDI process as the basis for likelihood ratio testing. Through comparison with simulated data, we test how well these distinct approaches hold up to immunological and demographic changes in model structure. We find that for all model conditions studied, the performance is informative for large populations (exceeding 106 individuals). For faster timescales of emergence, the likelihood ratio test performs better, whereas the opposite is the case for early-warning signals. Some early-warning signals, such as the autocorrelation, are shown to outperform others, such as the coefficient of variation.
The genetic algorithm identified the optimum thresholds and weights for the proposed signals under controlled, verifiable conditions. Application to more complex scenarios is ongoing. By presenting for the first time a method of evaluating thresholds, our work removes a key barrier to the development and deployment of early-warning systems for actual disease emergence.