Identifying indicators of critical transitions in epidemiological data

A challenging problem in infectious disease modelling is assessing when a disease has been eliminated. Control campaigns have substantial economic consequences; as such there are high demands to reduce costs and reallocate resources. However, if campaigns are stopped prematurely it can result in disease resurgence and subsequently put control efforts back by decades. Early-warning signals offer a computationally inexpensive technique to monitor the progress towards elimination, using statistical indicators calculated on time series data. Early-warning signals are widely used in many fields to anticipate a critical threshold prior to reaching it. A system undergoes the phenomenon known as critical slowing down as it crosses through a threshold. Theory predicts that fluctuations away from the mean will recover more slowly as the system approaches a critical transition (Scheffer et al., 2009). This is key in infectious disease modelling to assess when the basic reproduction number is reduced below the threshold of one. Recent theoretical advances have shown indicators of critical transitions in epidemiology such as measuring the variance in synthetic disease data. Our work highlights several challenges when applying this theory in practice. One potential problem is known as 'detrending' the data, which can be difficult to achieve in a single time series (Dessavre & Southall et al., 2019). Accurately detrending the signal removes the mean to obtain the fluctuations, whilst preserving any statistical properties. We present a novel approach using a metapopulation framework to successfully detrend data using the mean of different geographical subpopulations. A second limitation is that often only incidence-level data is available publicly. However, current theoretical analyses of statistical indicators concentrate on prevalence data, instead of new cases. We demonstrate that indicators calculated on simulated incidence time series data exhibit vastly different behaviours to those previously studied on prevalence data (Southall et al., 2020). Inconsistencies in time series traits between different diseases systems and a variety of disease data types could lead to misleading results when applied to collected data. In this talk we present methods for dealing with the typical data collected and our results show promising methods for calculating early-warning signals of elimination on real-world noisy data.