The limits on detecting epidemic resurgence from infection counts using reproduction numbers

What is it about?

The timely detection of epidemic resurgence (i.e., upcoming waves of infected cases) is crucial for informing public health policy, providing valuable signals for implementing interventions and identifying emerging pathogenic variants or important population-level behavioural shifts. Increases in epidemic transmissibility, parametrised by the time-varying reproduction number, R, commonly signify resurgence. While many studies have improved computational methods for inferring R from case data, to enhance real-time resurgence detection, few have examined what limits, if any, fundamentally restrict our ability to perform this inference. We apply optimal Bayesian detection algorithms and sensitivity tests and discover that resurgent (upward) R-changes are intrinsically more difficult to detect than equivalent downward changes indicating control. This asymmetry derives from the often lower and stochastically noisier case numbers that associate with resurgence, and induces detection delays on the order of the disease generation time. We prove these delays only worsen if spatial or demographic differences in transmissibility are modelled. As these fundamental limits exist even if case data are perfect, we conclude that designing integrated surveillance systems that fuse potentially timelier data sources (e.g., wastewater) may be more important than improving R-estimation methodology and deduce that there may be merit (subject to false alarm costs) in conservative resurgence response initiatives.

Featured Image

Why is it important?

We show that the uncertainty in estimating the reproduction number or rate of spread of an epidemic will usually be larger when detecting resurgent changes in epidemics (increases in infections) than equivalent changes that indicate control (decreases in infections). This is a simple consequence of resurgence often occurring when epidemic spread is small and infections are few, whereas control measures or interventions are commonly initiated when spread is larger and infections numerous. This asymmetry results in substantially more uncertainty around the estimates that can inform us about resurgent events. This uncertainty translates into a detection delay that is fundamental and limits our ability to speedily diagnose epidemic resurgences, often by as much as the mean generation time of the disease. This delay will only increase when practical data limitations are considered (e.g., lags in ascertaining cases) and supports early intervention as by the time we are confident that resurgence is occurring it may be too late.