Optimal control prevents itself from eradicating stochastic disease epidemics

The resources available for managing disease epidemics – whether in animals, plants or humans – are limited by a range of practical and financial constraints. Optimal control has been widely explored for optimising allocation of these resources to maximise their impact. The most common approach assumes a deterministic, continuous model to approximate the epidemic dynamics. However, real systems are stochastic and so a range of outcomes are possible for any given epidemic situation. The deterministic models are also known to be poor approximations in cases where the number of infected hosts is low – either globally or within a subset of the population – and these cases are highly relevant in the context of control. Hence, this work explores the effectiveness of disease management strategies derived using optimal control theory when applied to a more realistic, stochastic form of disease model. We demonstrate that the deterministic optimal control solutions are not optimal in cases where the disease is eradicated or close to eradication. The range of potential outcomes in the stochastic models means that optimising the deterministic case will not reliably eradicate disease – the required rate of control is higher than the deterministic optimal control would predict. Using Model Predictive Control, in which the optimisation is performed repeatedly as the system evolves to correct for deviations from the optimal control predictions, improves performance but the level of control calculated at each repeated optimisation is still insufficient. To demonstrate this, we present several simple heuristics to allocate control resources across different locations which can outperform the strategies calculated by MPC when the control budget is sufficient for eradication. Our illustration uses examples based on simulation of the spatial spread of plant disease but similar issues would be expected in any deterministic model where infection is driven close to zero.Author summary: Improving epidemic outcomes in situations with limited resources for control is important across human, animal and plant health. An example of this type of optimisation is deciding which infected plants to remove first to reduce ecosystem or agricultural losses from an invasive plant disease when the budget for control is limited. A common approach to these optimisations is to model the epidemic as a totally predictable (deterministic) process and modern algorithms allow these optimisations to scale to large, real-world problems (e.g. allocating resources across a landscape). However, when it is possible, the agencies responsible for allocating resources generally prefer solutions where the disease is removed entirely from the system or from a subset of the population. We tested solutions from deterministic optimisations against an epidemic model which includes some of the random effects we would expect to see in real applications. We showed that, in cases where eradication is feasible, the optimisations which use deterministic models are not optimal. They do not account well for the range of outcomes that can occur and so they can be outperformed by much simpler “rules of thumb”. This helps managers and researchers understand where the deterministic models are appropriate and where optimisations need to use different methods to account for the randomness of real epidemics.