ABSTRACT
Animal-borne or zoonotic human diseases (e.g., SARS, Rabies) represent major health and economic burdens throughout the world, disproportionately impacting poor communities. In 2013-2016, an outbreak of the Ebola virus disease (EVD), a zoonotic disease spread from animal reservoirs caused by the Zaire Ebola virus (EBOV), infected approximately 30,000 people, causing considerable negative social and economic impacts in an unexpected geographical location(Sierra Leone, Guinea, and Liberia). It is not known whether the spatial distribution of this outbreak and unprecedented severity was precipitated by environmental changes and, if so, which areas might be at risk in the future. To better address the major health and economic impacts of zoonotic diseases we develop a system-dynamics approach to capture the impact of future climate, land use and human population change on Ebola (EVD). We create future risk maps for affected areas and predict between a 1.75-3.2 fold increase in EVD outbreaks per year by 2070. While the best case future scenarios we test saw a reduction in the likelihood of epidemics, other future scenarios with high human population growth and low rates of socioeconomic development saw a fourfold increase in the risk of epidemics occurring and almost 50% increase in the risk of catastrophic epidemics. As well as helping to target where health infrastructure might be further developed or vaccines best deployed, our modelling framework can be used to target global interventions and forecast risk for many other zoonotic diseases.
Significance Statement Despite the severe health and economic impacts of outbreaks of diseases like SARS or Zika, there has been surprisingly little progress in predicting where and when human infectious disease outbreaks will occur next. By modelling the impacts of future climate, land use and human population change on one particular disease Ebola, we develop future risk maps for the affected areas and predict 1.7-3.2 times as many human Ebola outbreaks per year by 2070, and a 50% increase in the chance that these outbreaks will become epidemics. As well as helping to target where health infrastructure might be further developed or vaccines deployed, our approach can also be used to target actions and predict risk hotspots for many other infectious diseases.
Introduction
Little is known about how the majority of human infectious diseases will be affected by predicted future global environmental changes (such as climate, land use, human societal and demographic change) (1–5). Importantly, two thirds of human infectious diseases are animal-borne (zoonotic) (6) and these diseases form a major, global health and economic burden, disproportionately impacting poor communities (7, 8). Many zoonotic diseases are poorly understood, and global health responses to them are chronically underfunded (9). The 2013–2016 Ebola outbreak was unprecedented in terms of size, financial cost, and geographical location (10, 11); a stark illustration of our knowledge gaps, and demonstrating that it is imperative we develop quantitative approaches to better forecast zoonotic disease risk.
Ebola virus disease (EVD) was first identified in 1976, and since then there have been approximately 23 recognized outbreaks (12), predominantly within central Africa. EVD is causedby any one of four pathogenic strains of Ebola virus:Zaire (EBOV), Sudan (SUDV), TaïForest (TAFV), and Bundibugyo (BDBV). It presents as a non-specific febrile illness thatcan cause haemorrhagic fever, often with a high case fatality rate in diagnosed patients (13).Some Old World fruit bat species (Family Pteropodidae) have been suggested as reservoir hosts (14), however, while there is limited direct evidence, they are strong candidates to playa key role either as an reservoir or amplifying host (15, 16). In areas with EVD, there are frequent direct and indirect human-bat interactions, e.g., via bush meat hunting andduring fruit harvesting (17), presenting numerous opportunities for bat-to-human pathogen spill-overs to occur. Additionally, a third of known zoonotic spill-overs have been connected to contactwith great apes and duikers, although there is no evidence that these species act as reservoir hosts (10). It is clear, however, that once spill-over occurs human social factors such as movementand healthcare responses greatly influence the cumulative outcome of an outbreak (18). For instance, previous work has highlighted the importance of family interactions (19), funeral practices (20) and differential transmissionrates in hospitalized individuals (18).
Many attempts to understand Ebola outbreak dynamics have focused on mechanistic modelling approaches of human-to-human transmission post spill-over from animal hosts (13, 18, 19, 21–24). Mechanistic, or process-based, models are ideal for capturing epidemiological characteristics of diseases and, importantly, testing how disease outbreaks might be impacted by intervention efforts (25). One downside is that mechanistic models rarely incorporate spatially heterogeneous ecological and environmental information (26), such as the known high variance of batabundance and pathogen sero-prevalence across widespread individuals (27). In this context, correlative, or pattern-based, models (e.g.MaxEnt, Boosted-regression trees) have been used to simultaneously capture the spatial risk of both zoonotic spill-over and subsequent human-to-human infection (12). For some spatially-explicit analyses, there have been attempts to incorporate spatial patterns of human populations (28), while other have included air transportation networks (29),but nostudies that we are aware of have considered whole-system analyses for major epidemic zoonoses, such as Ebola. Like other rare or poorly-sampled diseases, Ebola suffers from limited data availability, meaning pattern-finding, correlative analytical techniques are at adisadvantage (30).
In 2014 a spill-over in Gueckedou district, Guinea of Ebola-Zaire virus led to an EVD outbreak approximately 100 times larger than any of the previous 21 known outbreaks (31). Such epidemics have a disproportionate impact on the affected societies. For example, the World Bank estimates a cost of US$2.2 billion to the three most affected countries (32) due to, amongst other drivers, widespread infrastructure breakdown, mass migration, crop abandonment and arisein endemic diseases due to overrun healthcare systems. Recent work has uncovered the human-to-human transmission patterns underlying this outbreak, using case (33) and genomic data (31) to demonstrate that EVD spread can be successfully predicted by a dispersal model that is weighted by both geographic distance and human population density. Attempting to understand zoonotic epidemic risk, however, using a human-only transmission model and without incorporating host ecology would inevitably lead to areas with high human density and connectivity being identified as the regions with the highest risk, despite some areas of these lacking competent hosts. Therefore, to model both the spatial variation in spill-over risk and, concurrently, the likely progression of subsequent outbreaks in human populations, we need to take a system-dynamics modelling approach (1, 34). Key non-linear feedbacks can also be captured, for example, the trade-off between increasing human populations and loss of reservoir host species through anthropogenic land-use conversion, and using this to design the optimal roll-out of vaccinations (35) and other interventions in a changing landscape.
Here, we use a disease system-dynamics approach (Figure 1) to extend a discrete-time, stochastic epidemiological compartmental model incorporating spatial environmental variability (Environmental-Mechanistic Model or EMM, Figure 2) to simulate present day spill-over and subsequent human-to-human transmission of the Zaire Ebola virus (EBOV) (the strain responsible for the 2013-2016 outbreak in West Africa). We model the impact of future anthropogenic changes on the occurrence and spread of this disease in 2070 (36–38) under a variety of possible integrated global change scenarios (39). We use a combination of three Representative Concentration Pathways (RCP) scenarios of increasing greenhouse gas concentrations: RCP4.5, RCP6, and RCP8.5 (40), and three possible socio-economic development scenarios (Shared Socio-economic Pathways or SSP), ordered by increasing human population density and reduced regional socio-economic cooperation:SSP1, SSP2 and SSP3. Finally, we compare the changes to spatial patterns of risk and chancesof outbreaks and epidemics occurring across Africa.
Results
Our EMM simulation for present day EBOV-EVD risk correctly identified areas of observed outbreaks as high risk, such as Democratic Republic of Congo, Gabon and the 2013-2016 outbreak in West Africa, but also identified some areas where EVD has not been reported, such as Nigeria and Ghana (Figure 3A). As a result, our model suggests that the at-risk area for EBOV-EVD is much larger than the areas known to have experienced disease outbreaks thus far. Our risk map also identified areas that are endemic for the other EVD strains,likely due to similar transmission pathways and reservoir host characteristics (Figure 3A). Although the index case risk map (Figure 3B) shows a similar spatial pattern to all cases, high risk spill-over areas are constrained to more distinct hot-spots. Importantly, the locationsof index cases that resulted in epidemics were even more geographically constrained, with G ana, Sierra Leone, Liberia, Kenya Uganda and Cameroon all having medium risk but Nigeria is the focus of the highest potential for epidemic spill-over (Figure 3C). Comparing the mean number of spill-overs per year gave higher results for present day simulations with 2.464 spill-overs per year (95% CI 2.361-2.567) compared to the mean historical number over the last 40 years: 0.75 (95% CI 0.695-0.905). High risk of Ebola case importation using the current network of airline flights was seen in China, Russia, India,the United States as well as many high-income European countries (Figure 4). Especially high importation risk, however, was seen in Italy and Germany.
Similar to historic data (Figure S3), the distribution of the final size of the simulated outbreaks was multimodal with distinct peaks at very low numbers (less than 3 cases) and medium outbreaks (approximately 3-1500 cases) (Figure S3).Through extensive simulations we were ableto explore the lower probability areas of the distribution effectively and, unique to the simulation data, there is a third peak of outbreaks(here we term ‘epidemics’) with high, to very high, numbers of cases (1500-100,000,000 cases). This threshold ofassigning an outbreak with greater than 1500 cases as an epidemic also corresponds to the top 1 percentile of a log-normal distribution approximating the variation in pre-2016 observed outbreak sizes (~1538 cases per year). Of the ~2500 simulation runs for present day conditions, epidemics (>1500) occurred approximately in 5.8% of the early simulations, with catastrophic epidemics (>2,000,000) occurring in around 2.3%of simulations, or once every 43.5 years given current conditions. From the sensitivity testing, the key parameters that affected outbreak size were illness length and R0, which positively increased case numbers (Figure S4a), where as the annual spill-over rate (Figure S4a) was most impacted by the spill-over rate constant (strongly positive), shape of the poverty/spill-over curve (weakly positive), and by host movement distance (weakly negative).
Our future EMM simulations estimate an annual increase in maximum area impacted by the disease from 3.45 million km2 to 3.8 million km2 under the scenario by 2070, with increases inmaximum area seen under all future scenarios (Figure 5A, Figure 5D, Figure 5G). The maximum areas where spill-overs could occur, however, increased by just 1% under the RCP4.5 SSP1 (Fig. 5B: 2.01 million km2), when compared to present day (Figure 3B :1.99 million km2), but increased by 14.7% under the RCP8.5 SSP3 (Figure 5H: 2.29 million km2) scenario. Conversely, the total area where epidemics could start decreased under the RCP4.5 SSP1 by 47% (Figure 5C: 0.444 million km2), when compared to present day (Figure 3C: 0.836 million km2), but again increases under RCP6 SSP2 this time by 20.5%, and by 34% under the RCP8.5 SSP3 scenario (Figure 5F, Figure 5I).
The increases seen in the area affected is mirrored by greater total numbers of spill-overs experienced in future scenarios, with the greatest increase seen under the RCP8.5 SSP3 scenario at 7.92 (CI 7.62-8.19) spill-overs per year. Spill-over numbers increased with greenhouse gas concentrations (represented here by the RCP value) with a mean 0.257 spill over a year increase between the RCP4.5 SSP2 and RCP6 SSP2 scenarios, and a mean 0.343 spillover a year increase between the RCP6 SSP3 and RCP8.5 SSP3 scenarios. Greater increases wereseen, however, with SSP change, with a mean 1.297 spill over a year increase between RCP4.5 SSP1 and RCP4.5 SSP2 and a mean 1.475 spill over a year increase between RCP6 SSP2 and RCP6-SSP3. In general, the probability of the index cases resulting in small outbreaks reduced infuture environments, whereas the chance of epidemics increased (Figure 6). For instance, the proportion of epidemics per year (>1500 cases)decreased in the RCP4.5 SSP1 to 3.43% (from 5.8% in present day) but increasedin all others, with RCP6 SSP3 gaining the greatest number, with epidemics in 9.5% ofall simulations. The number of catastrophic epidemics (>2,000,000), generally increased with bothRCP and SSP values up to 3.43% and 3.54% for the RCP6 and RCP8.5SSP3 scenarios respectively, but again saw a decrease from the present day level (2.3 %) to 1.19% for just the single ‘bestcase’ future scenario (RCP4.5 SSP1).
Discussion
We show that changes in future expected disease incidence are likely to be related to therate of global environmental change. According to our study, EVD mitigation attempts would be best placed in efforts to reduce both population growth, increase socioeconomic development and ameliorate climate change, such that global change most closely tracks the RCP4.5 SSP1scenario. Global binding commitments to reducing climate change may act to slow the effects,but evidence (41) suggests a wholesale change is difficult. Expected decreases in povertyand a concomitant increase in healthcare resources, therefore, would appear to be the most realistic approaches to reduce the future EVD disease burden. While vaccinations may be effective, the sporadic nature of spill-over events mean it is unclear where vaccination should be targeted and whether it would be cost-effective at this time (35). More generally, increasing health care provision and poverty reduction efforts in West Africa would not only reducethe potential effects of EVD but also other diseases, including those that have yet to emerge in earnest, suchas Marburg virus disease (42), Lassa fever (43), and Nipah/Hendra virus infection (44).This, in turn, could limit disease emergence to local outbreaks, preventing nosocomial infections and acting to prevent subsequent epidemics.
Changes to SSP scenarios, which control levels of poverty and human population size in our models, had a greater impact than changing the climate and land-use chang (here mediated via RCP scenario). This is not surprising as poverty reduction increases the presumed EVD-EBOV healthcare response in our simulations, and many of the countries in the endemic region are expected to have substantial reductions in poverty levels by 2070 (37). Similarly, contact rates in our simulation (both between humans and between humans and wildlife) depend linearly on human population growth, whereas climate change increases EVD-EBOV cases through more complex interactions. Species distribution models indicate that the presumed wildlife hosts prefer warm and wet conditions (Figure S1-2), which are expected to increase in these regions according to the HADGem3-AO climate model (38)(Fig. S5). This expansion of the optimal conditionsfor presumed the wildlife host species effectively increases the at-risk human population byincluding more of the northern, eastern and southern areas of Africa (Figure 3A). Predicted future anthropogenic land-use changes, however, reduces the optimal wildlife host habitat, thereby reducing human-wildlife interactions.
We identify Nigeria as, not only a key area for epidemics to be initiated, but also an area with potential for many small outbreaks. This might indicate that our model has not correctly balanced the impact of healthcare infrastructure on disease spread, regional behavioural barriers to infection or regional differences in contact rates between both humans and hosts. Until these additional factors are explicitly tested, the high human density and known presence of putative wildlife hosts mean that this area should be consider at high risk of initiating epidemics.
There is a pressing need to better understand the spatial variation in other key disease transmission parameters. For instance, bush-meat hunting is an important process by which human populations come into contact with large bats resource (45) and the spatial variation inbush-meat extraction is likely a component of spill-over variation. Little is known, however, about bush-meat hunting outside a few specific studies but there is potential to use spatial interpolation techniques to make reasonable predictions in un-sampled areas. Our model does not incorporate this data or test its impact and, similarly due to lack of data resources, we do not use information about local differences in funeral practices. Hospital compartments are thought to be useful to understand quarantineand super-spreading events but there is very limited data on the quality and geographic reach of small health clinics. Some other important behavioural trends are not captured in our model, such as the post-outbreak behavioural reactions of human populations e.g. mass migration away from affected regions. Recent findings regarding the persistence of Ebola virus in semen of convalescent men may also help explain the intermittent spatiotemporal patterns of infections in endemic areas (46, 47). Future work incorporating such data, may further improve the spatial resolution and accuracy of risk estimates.
Our approach demonstrates not only an important frame work understanding Ebola but also for other diseases. Analysing diseases singly cannot be an effective approach for policy making at a large geopolitical scale, particularly in regions with multi-disease burden and limited healthcare resources. Net disease risk patterns, when summed across a wide variety of zoonoses, will be an emergent property of the distribution of very different wildlife host species and their respective responses to increasing anthropogenic land-use conversion and climate change. Any lack of data in the short-term does not reduce the obvious importance of understanding future disease trends. Attempts, such as ours, establish a first heuristic step on a pathway to building intervention measures aimed at reducing overall future disease burden.
Materials and Methods
Environmental Mechanistic Model (EMM) EBOV
Using our discrete-time, stochastic epidemiological compartmental model incorporating spatial environmental variability (30), we extended the approach to not just simulate pathogen spill-over but also subsequent human-to-human transmission, focusing on the Zaire Ebola virus (EBOV) (Figure 2). Within grid cells (0.0416°) covering continental Africa, we used a Susceptible, Exposed, Infectious, Funeral and Removed (SEIFR) EVD-EBOV disease compartmental model (following13, 19, 23) to estimate the number of individuals per compartment, in each time stept, for present day bioclimatic, land use and demographic conditions. Although some previous compartmental models for EBOV have included a Hospital compartment (48), adding this complexity was not feasible over large and poorly known geographical areas. Without knowing more about the spatial variation in health seeking behaviour, exactly which grid cells contain clinics, and the variation of healthcare resources in these clinics, adding in this compartment would not likely significantly improve our model's ability to predict the progression of outbreaks. Further more, hospital interventions had the least impact controlling EVD outbreaks in a recent meta-analysis (24). All analyses were carried out in R v.3.2.2 (49). Each stage of the EMM simulation is discussed in more detail below:
Stage 1: SEIFR compartmental model within grid cells
We used starting EBOV transmission characteristics of incubation time = 7 days, onset of symptoms to resolution = 9.6 days, case fatality rate(CFR) σ = 0.78, and burial time = 2 days (23) to parameterize the SEIFR compartmental model to determine transition rates α(between Exposed to Infectious compartments), γσ (Infectious to Funeral),γ1−σ (Infectious to Removed), and γF(Funeral toRemoved) (Figure 2). Toincorporate sensitivity around these transmission parameters, we allowed values to vary for each simulation run by sampling from a Gaussian distribution where themean was their initialvalue and standard deviation was fifth of the mean, to give a reasonable spread of values. For each time stept, the number of individuals moving between all compartments was estimated by drawingrandomly from a binomial distribution (Section S1 equation (1)), parameterized using the respective compartmental transition rates. Transition rates for compartments were assumed to be the same in all grid cells except for the transition between Susceptible to Exposed. The per grid cell Susceptible to Exposed transition rates were determined by the force of zoonotic infection λz, and the force of infection λ (Figure 2) and these were calculated as follows:
(a) Force of Zoonotic Infection, λz
The force of infection for zoonotic transmission λz, per time step t, was estimated as the product of the probability of host presence H, and spill-over rate k (Section S1 equation (2)). Without any evidence to the contrary (15, 50), we parameterized H by calculating the spatial probability of the presence of the mostlikely EBOV reservoir host species based on available data (Old World fruit bat species Epomophorus gambianus gambianus, Epomops franqueti, Hypsignathus monstrosus, and Rousettus aegyptiacus see Table S1) within each grid cell (0.0416°) across the African continent using species distribution models (SDMs) (51) and assuming constant pathogen prevalence. We also calculated the spatial probability of thepresence of other species which are known to provide an alternative route of infection, butlikely do not act as reservoirs (Gorilla spp., Pan spp., and Cephalophus spp.) (12). SDMs for each species were inferredusing boosted regression trees (BRT) using distribution data from the Global Biodiversity InformationFacility (GBIF) (52) and 11 presentday bioclimatic and landuse variables (Table S2). Data with coarse scale GBIF spatial coordinates (decimal degree coordinates with less than four decimal places)were filtered out of the analysis. To reduce spatial autocorrelationand duplicate records, any records that co-occurred in the same grid cell were removed. Lastly, GBIF records older than 1990 were discardedto ensure samples more closely matched the current landscapes. BRT tree complexity was set at5 reflecting the suggested value and the learning rate was adjusted until >1000 trees were selected (53). A total of 25 models were estimated for each species using four fifths of the distribution data asa training dataset andone fifth as a testing dataset, chosen randomly for each model. Thosewith the highest predictiveability (high area under operating curve, AUC and true-skill statistic, TSS values) were selected as the best model for each species (Figure S1). The most important spatial variables determining distributions acrossthe different reservoir host species were BIO7 Temperature Annual Range, BIO13 Precipitationof Wettest Month, BIO2 Mean Diurnal Temperature Range and LandUse-Land Cover (Figure S2). The outputs from all putative reservoir (bat) specieswere combined into a single value representing the probability of any reservoir species being present and a similar approach was taken for the non-reservoirhost species. The reservoir and non-reservoir host layers were then combined, but since onlya third of index cases were attributed to non-reservoir host spill-overs (10), wedown-weighted the probability of the non-reservoir occurrence by two thirds and reservoir occurrence by one third when combining the layers. The final resulting probability was bounded by zero and one. Additionally, as EBOV presence in non-reservoir host species is impossible without the presence of reservoir hosts, cells with a reservoir host probability of zero were given a value of zero irrespective of the non-reservoir host score. For computational simplicity, we assume that all human individuals have equal chance of exposure to infected host species. The initial value used for spill-over rate k, per time step t, was estimated fromthe number of historic outbreaks O (defined here as distinct clusters of cases) (taken from empirical EBOV outbreak data 12), and the number of historically susceptible individuals Sh (inferred from human population estimates from 1976 to 2015 from 37) (see Section S1 equation (3)). During each simulation run, κ was allowed to vary using the same method as the compartmental transmission parameters above.
(b) Force of Infection, γ
The force of infection for human-to-human transmission λ per time step t, was estimated as the product of the effective contact rate β, and the number of individuals that cantransmit the disease in each relevant compartment (Infectious and Funeral) per grid cell (0.0416°) (Section S1 equation (4)). We assumed that β for the Infectious and Funeral compartments was equivalent, due to the contact rates of moving individuals in the Infectious compartment being off set by large aggregations of individual sat funerals. We estimated the effective contact rate β, as the basic reproduction number R0 divided by the product of the total number of individuals N, and infectious duration D (the sum of Infectious and Funeral compartment time, 11 days takenfrom 23). As a starting value for R0 we used avalue of 1.7 (54) and this was allowed to vary per simulation run using the same method as the compartmental transmission parameters above. As per previous research(30), we incorporated spatial variance in contact rates among grid cells usingaweighting factor m,where by the effective contact rate in grid cells with greater than expected contact rates wasincreased and decreased where fewer contacts were predicted(Section S1 equation (5)). We estimated m by creating an ideal free gas model of human movement within each grid cell and approximated collision frequency per person per day, using the following: the total individuals in each grid cell (estimated from Gridded Population of the World v3 55), an individual interaction sphere of radius 0.5 m, and using per person, daily walking distances in meters vΔt, where v is walking velocity, andΔt equals time period (Section S1 equation (6)). To capture geographic variation in human movement patterns, each grid cell was assigned a value for per person daily walking distance, based on the empirical relationship between daily walking distances and per person percountry Gross Domestic Product (measured as Purchasing Power Parity from 37) (Table S3). Asthe availability of mass transit as alternative to walking tends to be centrally controlled, we assumed that grid cells in each country had the same value.
Under real conditions, the effective reproduction number Redecays over time as both efforts are made to control disease spread and as the pool of susceptible reduces, which results in R0 being equal to Re only when time step tis zero. Therefore, to calculate effective contact rate β, we allowed Re to decay per time step t (Section S1 equation (7), equation (8) and equation (9)). However, countries thatcan invest more in health infrastructure (e.g., barrier nursing, surveillance) should see amore rapid reduction in Re over time compared to countries that do not have such infrastructure and also a concomitantly, a decrease in CFR. Therefore we derived an empirical estimate of the relationship between wealth (measured using GDP-PPP per capita) and both the relative rate of decay of Re over time (Section S1 equation 10) and CFR (Section S1 equation (11)), and using a spatially disaggregated poverty data layer (56) we weighted the per grid cell per time step Re reduction and CFR accordingly to the values in each gridcell. While we found the relationship between wealth and both Re and CFR reduction over time to be best described using curves with exponents of −0.08 and −0.02, respectively, thiswas inferred using relatively few data points (Table S4). In our simulation runs, therefore,we allowed these exponents to vary similarly to the parameters above, to allow either morelinear declines or deeper curves to best estimate the true impact ofthis relationship.
Stage 2: SEIFR compartmental model between grid cells
We allowed those individuals that had contracted EBOV to travel between grid cells, specifically individuals in Exposed and Infectious (but not Funeral) compartments (Figure 2), but assumed for simplicity that the overall net movementof susceptible individuals between cellswas zero. As previously supported with empirical data, we employed a movement model that was weighted by both geographic distance and human density (31, 33) and was also geographically constrained toknown transportation routes. The transmission rate ϵ, of individuals between target compartments of different grid cells was estimated by two different methods: between grid cells along road networks ϵr, andalong flight routes ϵF. We sampled randomly, froma binomial distribution, the number of travellers per grid celland time step t (Section S1 equation (1)) with the probability of travel by road per day ϵr, being proportional to the distance to the nearest road using the Global Roads Open Access Data Set (Global Roads Open Access Data Set from 57). Global roads data set contains intotal 585413 routes from tracks to multi-lane highways and has been extensively validated for Africa (58). We allowed travellers tomove freely (agnostic to any particular transportation method or country boundary) across the continent upto 10 road junctions in any direction from the centroid of the starting cell along the road network (Global Roads Open Access Data Set from 57), giving a potential of upto 500 km of linear travel per time step. Each proposed travel end point was given an individual probability from the daily distance travelled probability curve from (Figure 2(f) of 59), which is derived from transport data and validated against mobile phone data. For air travel, we set the potential pool of travellers as the individuals in grid cells containing airports across the world (from Open Flights Airport Database 60) plus all the Exposed individuals in the 8 grid cells surrounding each airport grid cell. We sampled randomly from a binomial distribution the number of travellers per grid celland time stept (SectionS1 equation (1)) with the probability of travel by air per dayϵf, being proportional to the total number of flights per day divided by the population ofthat country (37). We allowed travellers to move up to 2 edges on the current airline routes from airport origin using the (from Open Flights Airport Database 60). Thisapproximates a traveller taking either a one or two-legged journey. Final destinations weresampled at random, based on all potential air routes having equal priority, but in most cases potential destinations were located nearby which by default meant that more distance travel was less likely than travel to a nearby location. For both road and air travellers, individuals were then added to the correct compartment of their final destination in the new grid cell and removed from the same compartment from the original source grid cell.
Stage 3: Impact of future anthropogenic change
(a) Future force of zoonotic infectionλz
We recalculated values of the force of zoonotic infection λz, by estimating the probability of EBOV host presence, H2070 under several different future integrated scenarios that incorporate projections of bioclimatic and land use variables (Table S2). Estimates of bioclimatic variablesfor 2070 were based on the HADGem3-AO climate model (61) under three Representative Concentration Pathways:RCP4.5, RCP6, and RCP8.5 (RCP45, RCP60 and RCP85 40). To estimate host presence probability in the future we needed to predict fine-scale future habitat data under theRCPscenarios. As only coarse categorisations are currently available (62), we therefore separately empirically estimated future land use-land cover (LULC) change (using MODIS data 36).Foreach grid cell we calculated the probability of each possible LULC change within the 2001-2012 MODIS dataset within a surrounding 5x5 cell grid using satellite data from 20. Based on these probabilities we simulated yearly LULC change across the region of interest for each grid cell from 2012 until 2070,and ran this simulation 100 times to create a bank of future possible landscapes, which werethen summarized into three consensus landscapes representing low (with anthropogenic changesrejected where possible),medium by choosing the majority consensus across all 100 runs) andhigh anthropogenic change, (anthropogenic changes were chosen if available across the landscape) and we aligned these threescenarios to SSP1, SSP2 and SPP3 respectively.
(b) Future force of infection λ
Using predicted human demographicvariables and poverty levels for 2070, we recalculated values for the force of infection λ, by estimating the number of individuals per grid cell, n and effective reproduction number, Re. We inferred human population estimates per grid cell for 2070 by using the Gridded Population of the World v4(55) for present day and multiplying each cell by the expected future proportional change over that time period predicted by three Shared Socio-economic Pathways: SSP1, SSP2 and SSP3. Future poverty estimates per country were similarly inferred using a spatially-disaggregated GDP layer(63)multipliedby the expected change in per country GDP over the time period as predicted by the SSP integrated scenario. We note that as our travel probability is defined per person, increasing future populations will see aproportion increase in the amount of both road and air travel.
(c) Comparison of simulation runs
We reran the EMM simulations under 5 plausible combinations of 2070 future environmental-socioeconomic scenarios of global change and greenhouse gas concentrations: RCP4.5/SSP1, RCP4.5/SSP2, RCP6/SSP2, RCP6/SSP3, RCP8.5/SSP3 (64). These different input data options were,specifically: (i) RCP4.5-stabilization scenario in which total radiative forcing is stabilized shortly after 2100, (ii) RCP 6 - stabilization scenario in which total radiative forcing is stabilized short lyafter 2100, without overshoot, by the application of a range of technologies and strategies for reducing greenhouse gas emissions (iii) RCP 8 – worsening scenarios with increasing greenhouse gas emissions over time,leading to high greenhouse gas concentration levels, (iv) SSP1 – high regional cooperation, low population growthdue high education and high GDP growth, (v) SSP2 – a ‘processes as usual’ scenario with ongoing levels of population growth and wealth,with medium estimates for both these by 2070, and (vi)SSP3 – regional antagonism, high population growth, unsustainable resource extractionand low economic growth. For each ofthe six scenarios we aimedfor 2500 runs of 365 days, each day measuring the number of spill-overs, the number of secondary cases associated with each spill-over, and the geographical areas affected. This allowed us to measure likelihood of spill-overs leading to small, mediumand very large outbreaks, and also to determine the geographical areas with the highest riskof experiencing cases. We also noted the destination of any flights out of Africa that contained infected people.
Acknowledgements
This work, Dynamic Drivers of Disease in Africa Consortium, NERC project no. NE-J001570-1was funded with support from the Ecosystem Services for Poverty Alleviation Programme (ESPA). The ESPA programme is funded by the Department for International Development (DFID), the Economic and Social Research Council (ESRC) and the Natural Environment Research Council (NERC). AAC is additionally supported by a Royal Society Wolfson Research Merit Award. We thank Prabu Sivasubramaniam for technical assistance, and A., Jones, M. Wilson, G. Mace, M. Leach,and C. Watts for comments on previous versions of the manuscript. All simulation data are available on figshare (figshare.com/s/c41c50a0675311e5b6b306ec4bbcf141) and the supplementary materials contain all other data. D.W.R. and K.E.J. developed the overall study design.D.W.R. carried out the modelling and data processing with assistance from K.E.J. All authors contributed to writing the manuscript. The authors declare no competing financial interests.
References
- 1.↵
- 2.
- 1.
- 4.
- 5.↵
- 6.↵
- 7.↵
- 8.↵
- 9.↵
- 10.↵
- 11.↵
- 12.↵
- 13.↵
- 14.↵
- 15.↵
- 16.↵
- 17.↵
- 18.↵
- 19.↵
- 20.↵
- 21.↵
- 22.↵
- 23.↵
- 24.↵
- 25.↵
- 26.↵
- 27.↵
- 28.↵
- 29.↵
- 30.↵
- 31.↵
- 32.↵
- 33.↵
- 34.↵
- 35.↵
- 36.↵
- 37.↵
- 38.↵
- 39.↵
- 40.↵
- 41.↵
- 42.↵
- 43.↵
- 44.↵
- 45.↵
- 46.↵
- 47.↵
- 48.↵
- 49.↵
- 50.↵
- 51.↵
- 52.↵
- 53.↵
- 54.↵
- 55.↵
- 56.↵
- 57.↵
- 58.↵
- 59.↵
- 60.↵
- 61.↵
- 62.↵
- 63.↵
- 64.↵
- 65.↵
- 66.↵
- 67.↵
- 68.
- 69.
- 70.
- 71.↵
- 72.
- 73.↵
- 74.↵
- 75.
- 76.
- 77.
- 78.
- 79.↵
- 80.
- 81.
- 82.
- 83.↵
- 84.↵
- 85.
- 86.↵