Background
Malaria, alongside HIV and tuberculosis, is considered one of the “big three” infectious diseases of humans. The global response to malaria transmission has been significant, with
Plasmodium falciparum malaria eliminated from 79 countries from 1979 to 2010 [
1]. Modelling suggests that 70% of the reduction in malaria cases in sub-Saharan Africa (SSA) between 2000 and 2015 was attributable to the implementation of intervention strategies [
2]. Key interventions included insecticide-treated bed nets (ITNs), artemisinin-based combination therapy (ACT) and indoor residual spraying (IRS). Often, field data are used as evidence for the efficacy and cost-effectiveness of selected interventions; however, these methods can be resource intensive, or have prohibitive ethical barriers. In such situations, mathematical simulation is increasingly used to provide further insights.
Infectious disease modelling of malaria has existed for over a century [
3], with the dominant paradigm being the Ross–Macdonald models used by the Global Malaria Eradication Programme (GMEP) from 1955 to 1969 [
4,
5]. These are examples of compartmental transmission dynamic models, in which the simulated human population consists of groups of individuals in disease states such as “susceptible”, “exposed”, “infectious” and “recovered”. More recent compartmental models of malaria provide insights into risk-stratification of populations, multiple mosquito populations, and waning immunity; however, the majority of models still closely resemble the Ross–Macdonald configuration [
6]. Comparing a variety of modelling approaches can provide robustness of results, and highlight areas for development of modelling techniques [
7]. As such, comparing alternative model frameworks may accelerate learning about disease transmission and control.
One approach that has significant potential is the use of agent-based models (ABMs). There appears to be no universal definition of ABMs; this review includes any model that explicitly models individual actions and responses, and associates with each individual respective state variables and parameters. In the following review, all published ABMs are stochastic in nature.
As ABMs focus on the individual, they afford flexibility in modelling factors, such as spatial heterogeneity (e.g. host movement, heterogeneous implementation of interventions) and stochasticity (e.g. inter-patient variability in time of infection, time to recovery, and location of infection). Compartmental models of malaria transmission do exist that incorporate either stochasticity of individual infections [
8] or spatial heterogeneity [
9]. However, in areas of increasing spatial variation, compartmental models may face convergence issues, or provide no more insight than alternative model structures. In low-transmission environments, where patient variability is more pronounced, ABMs can better represent the stochasticity of disease progression and transmission than compartmental models, where (to some resolution) people are grouped together and treated as interchangeable, so individual agent behaviour cannot be determined. Models that accommodate patient individuality and spatial variation can help fill knowledge gaps [
7] about transmission heterogeneities important in malaria elimination strategies.
The flexibility of agent-based approaches also allows models to be constructed to address practical questions relating to malaria control and elimination in specific local contexts [
7,
10]. This is advantageous because identifying optimal local intervention strategies can provide a strong evidence-based framework for National Malaria Control Programmes (NMCPs). ABMs can be constructed to resemble such specific settings closely, due to their flexibility in altering model attributes to reflect local individual characteristics and geographical factors.
As more is learned about malaria transmission, the complexity of the questions asked increases, which in turn calls for more nuanced models. The role of mathematical models continues to grow as both technical expertise and computing power increase. With the increasing capacity for modelling to assist in malaria elimination programmes, a review of the published literature for ABMs of malaria transmission was performed. Analysis included characterization of the structure of existing models, the factors influencing malaria transmission modelled, and the methods of data use and output analysis. The approaches used were highlighted and ideas for advancing the field proposed.
Discussion
Mathematical modelling plays an important role in malaria elimination, and agent-based approaches make a major contribution to these efforts. The extension of compartmental models to their early ABM equivalents arose from the need to understand malaria transmission at the individual level. The result is a rich array of model families and simulation techniques, adapted to a range of key issues in transmission and control.
In general, three core themes emerged regarding justification of ABM use. First, the greater importance of stochasticity in low-transmission settings, particularly settings approaching elimination, requires an alternative approach to traditional compartmental methods. Second, attempts to eliminate local transmission require discrete population simulations to incorporate spatially explicit environments at increasingly fine resolutions. Third, heterogeneities in disease progression and severity on the individual patient level result in varying efficacy of drug and vaccine interventions, which may be difficult to capture within a compartmental framework. These three arguments stem from a common point: a compartmental structure, based on averaging over a population, has limitations when that average does not adequately represent the individuals.
In addressing these issues, the benefits of agent-based techniques in this space are evident. Many papers in this review explicitly aimed to fill the knowledge gap regarding intervention use in low transmission environments. Most projects provided outputs robust at multiple transmission intensities, highlighting the flexibility of ABMs in low-prevalence settings. The HYDREMATS framework was used in multiple locations, incorporating environmental factors such as temperature and rainfall at different times, [
25,
38,
39]. The OpenMalaria models progressed from assessing the force of infection of malaria transmission [
32], to estimating cost-effectiveness of a vaccination programme [
27]. Given the similarities of compartmental models of malaria to the original Ross–McDonald framework [
6], it is clear that the depth and flexibility of agent-based methods are allowing new insights into malaria transmission and prevention.
The variation in the models described above highlights the difficulties in developing a standardized style of ABM for use in malaria epidemiology. However, this is arguably a major advantage, with the abundance of techniques allowing for the flexibility desired when transitioning from solely using compartmental models. Instead of suggesting a “gold standard” approach, it may be preferable to ensure the model style used is appropriate for the question at hand. For example, OpenMalaria’s early modelling of gametocyte densities did not use vector agents [
18], but successfully provided insights into risks of fever, morbidity and mortality of patients [
24]. The EMOD models initially described host-vector interactions without spatial consideration [
19], but added this capability when required to assess interventions [
62,
85]. Therefore, while not every model incorporated every aspect of malaria epidemiology, each was tailored to the research question at hand.
Conversely, if modelling groups are considering extending their model frameworks, particularly to influence policy, there is potential to draw from the features of one another. For example, HYDREMATS currently includes human and mosquito agents, while the characteristics of human infection are more detailed in the OpenMalaria simulations. Therefore, the time variability of individual gametocyte density, probabilities of fever, morbidity and mortality, and the infectivity of hosts to vectors used in the OpenMalaria framework could be adapted into HYDREMATS to more realistically replicate disease transmission. However, in neither of these simulations do humans move, whereas this process is explicitly simulated in Zhu et al. [
31] and Pizzitutti et al. [
37] to better represent vector-host feeding patterns. Pizzitutti et al. [
30] and EMOD [
19] include human behavioural reactions to biting rates (i.e. time-dependent intervention use) and the probabilities of successful blood meals, respectively. These five models have components that simulate vector, egg and human populations, effects of climate on larval habitats, anthropophily, ITN, IRS, larval habitat removal, vaccination, anti-malarial use, attractive toxic sugar baits, and rates of human disease. As each framework provides insights into key components of malaria transmission, all of which are important in guiding elimination strategies.
To some extent, combining model structures across research teams can be considered an extension of the use of “submodules” already undertaken by larger modelling frameworks. The HYDREMATS team have successfully integrated detailed larval habitat and entomologic models [
55], and OpenMalaria now includes upwards of seven modules of human disease states and interventions [
100]. Modular projects such as HYDREMATS, EMOD [
85] and OpenMalaria have provided insights into transmission dynamics, vector populations, disease severity, and the contributors to these factors. Given the importance of comprehensive modelling to guide policy decisions, the potential for combining the strengths of validated models to enhance decision-making capabilities of ABMs could be explored.
A key target for modelling low-transmission settings is a focus on spatial representation and heterogeneity. ABMs can shift spatial heterogeneity from a typically “patch-based” compartmental framework into a continuous space, by having explicit locations for environmental objects, dwellings and agents. These detailed descriptions of the landscape are coupled with local knowledge of physical characteristics (such as host/vector movement patterns) to simulate malaria transmission, ecology, and the impact of interventions based on their location. These insights include the distances between larval habitats and houses to effectively reduce malaria transmission [
53], and the impact on systematic versus random location of attractive toxic sugar baits (ATSB) on mosquito abundance [
21]. This style of intervention inherently requires spatial modelling, although interventions such as ITNs and IRS have been modelled in both spatial and non-spatial simulations. Examining these interventions in physical environments allows the impact of factors such as vector movement and the proximity of unprotected individuals to be measured.
Potential extensions to spatial models include varying elimination strategies across a landscape, and increasing the size of the geographical area modelled. For example, consider a small community near a local water source, and a nearby larger population with better access to healthcare. A model could implement larval source management at the local water sources, whilst increasing access to vaccines and anti-malarials in the healthcare centre. Human movement dynamics [
121] could be incorporated to assess the relative effectiveness of each interventions across both populations. This style of modelling may more accurately represent the manner in which interventions are implemented at the local level. Regarding model areas, most simulations of real-world environments only covered the area of a specific village, with sizes ranging from 600 m × 600 m to 3000 m × 3000 m. These spatial ranges have been limited by computational power, but this limitation will continue to decrease over time. There also may be a lack of access to consistent geospatial data over larger areas; however, many models only included spatial data on physical habitats, which would be collected in a similar manner over larger areas. If it is deemed useful to model malaria over a wider area, techniques from other fields may be used, such as probability modelling of invasive species, which has been performed for an area of over 35,000 km
2 [
122].
Regarding the locations of malaria modelling, there is an understandable focus on SSA, which was responsible for 88% of the global malaria burden in 2015 [
2]. However, there has also been a recent increase in attention on South-East Asia (SEA), which is responsible for 10% of global cases and has emerging issues with drug resistance [
2]. Moreover, challenges such as artemisinin resistance and insecticide resistance are more prominent in these areas. Despite this, approximately half of all malaria cases outside Africa in 2015 were due to
P. vivax [
2], while
Plasmodium knowlesi malaria transmission is increasing in locations, such as Malaysia [
123]. The methods of ABM construction used in SSA and SEA, and for
P. falciparum malaria, suggest transferability to other regions and
Plasmodium species, which will be important as data availability from these areas improves and attention turns to global elimination.
Many ABMs reviewed here had an emphasis on informing policy and explicitly aimed to understand specific programmatic questions (e.g. [
16,
28,
78,
83]). To reliably inform public health decisions, there must be confidence in the assumptions guiding model creation, in particular regarding choice of parameter values. The methods for estimating key parameters varied greatly across the literature. Parameter justification was not always clear [
13,
22,
51,
65]; when explained, models generally calibrated a range of parameters to existing data, or provided references for their choice of fixed values. Further, no previous models have calibrated all parameters to data. Importantly, ensemble modelling of the OpenMalaria variants calibrated 14 variants to the same dataset, but parameter values varied significantly across between models [
107]. Given that uncertainty remains even after calibration, it is important to apply a systematic and comprehensive approach to parameter estimation before using models for predicting parameter impact or forecasting.
Whilst field data exists for a range of parameters, researchers must be pragmatic about the possibility of adequately calibrating complex ABMs, particularly when data is required in resource-poor settings. The fixing of well-established values can reduce the parameter space to be searched using calibration methods. Alternatively, techniques such as Markov chain Monte Carlo (MCMC) can search the entire parameter space, or more precisely that part of parameter space that has non-negligible posterior probability. MCMC has already been used in malaria ensemble modelling [
108,
116,
118] and parameter estimation by Griffin et al. [
14], as well as in modelling of other infectious diseases [
124,
125]. Approaches such as MCMC and approximate Bayesian computation are increasing in popularity as including uncertainty in model parameters becomes more common [
126].
An arguably more pressing area of need for development is optimization, with methods for agent-based models still in their infancy. In this review, studies that reported optimization usually simulated a suite of different interventions, or the same interventions at different levels of coverage or timing. Cost-effectiveness analysis was typically approached in the same manner. Whilst conclusions were provided as to the most effective simulation approach, true optimization was rare [
69,
97], using formal techniques to identify parameter values that optimize one or more objective functions. Given the role of ABMs in modelling interventions in low-transmission settings, formal optimization techniques are important for enhancing the ability of models to guide policy.
Deterministic models have already been used as the basis of optimization of interventions for various infection diseases [
125,
127,
128]. Strategies for the optimization of interventions within ABMs appear less common, possibly due to the high computational burden of finding consistent minima in the presence of stochasticity. It is difficult to define how to best approach optimization from an agent-based standpoint. A systematic review of ABMs for optimization problems [
129] highlighted techniques used for disciplines such as scheduling, supply chain management, energy systems planning and transportation and logistics. As is likely the case regarding spatial methods, optimization of malaria transmission modelling (and infectious disease simulation more broadly) may benefit from adapting approaches outside the field to a new context.
Increased clarity in model reporting would be of great benefit to both the creators of ABMs and their audience. While many papers included detailed supplementary materials for additional results, project descriptions and calibration, validation, sensitivity analysis and optimization techniques, the intricacies of these techniques often unclear. A protocol exists for the description of ABMs [
130], and models that used it [
21,
31,
76] were simple to understand and appeared easily replicable by external groups. Further transparency includes sharing of the mathematics and code [
76] of models. These small steps in documentation would allow for increased verification and validation of models, as well as increasing opportunities for collaboration between modelling groups.
Beyond individual models, ensemble modelling is an important tool for generating robust conclusions about malaria transmission. A review of ebola models advocated for an ensemble modelling approach that adequately compares state-of-the-art models, but also allows for model diversity [
131]. For malaria models built for similar purposes (for example, to estimate certain parameters of interest, or to predict the value of an intervention), both inter- and intra-model comparison has been conducted. In some cases, two or three interventions have been simultaneously assessed using this approach [
110,
114,
115]. The next steps for ensemble and consensus modelling may include more interventions, spatial modelling, and developing techniques to determine which models are most appropriate when ensemble members differ in parameter estimates or outputs.