Human case data

Human MPX case information is almost completely based on passive surveillance with heterogeneous case report efforts across the region; as a result, we must assume that some degree of bias will be inherent in our dataset. We used the original human case reports from the WHO surveillance efforts in the DRC in 1970–1986, which were the object of extensive and intensive analyses by our research group [7], [27]. Confirmed human cases were georeferenced at the patient's residence location (representing the place where infection most likely occurred) using digital versions of 1∶250,000 Joint Operational Graphic (JOG) topographic maps for the DRC, and GEOnet Names Server (GNS; http://earth-info.nga.mil/gns.html/), in tandem with detailed case information from the original reports, and following the georeferencing procedures from MaNIS [28]; a detailed analysis of implications of different georeferencing protocols for MPX geography is provided in a separate publication [27]. Given our interest in modeling environmental conditions required for the disease to be transmitted from its wildlife reservoir to humans, cases most likely to have resulted from secondary transmission (i.e., the disease reported as acquired by contact with a sick person in the WHO case records) were eliminated from the study. In all, 100 unique localities from DRC with human cases were used (Figure 1). We assume that all of these cases correspond to the central African clade of MPXV, based on the fact that only that clade has been found in the Congo Basin to date [29], [30], [31], [32].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Human MPX case localities between 1982 and 1986 used for training ecological niche models.

Georeferenced localities in the Democratic Republic of Congo of human MPX cases reported by WHO (n = 100), divided into two groups by median latitude and median longitude (white circles  =  on-diagonal, blue squares  =  off-diagonal). The dashed line delimits the area of interest for the present work.

https://doi.org/10.1371/journal.pone.0074816.g001

Recent human case data were georeferenced to provide an independent test for projections of the ENMs at a more recent time period. We followed the same methodology explained above to obtain geographic coordinates, resulting in a total of 18 unique localities where the disease was reported between 2001 and 2010 in the study area; data were drawn from the literature [33], [34] and the Institut National de Recherche Biomedicale (INRB) in DRC. The majority of localities (present and past) had a spatial uncertainty of ≤4.0 km; only 11 localities presented broader uncertainties, as high as 6.5 km. This known accuracy of georeferenced localities permitted the use of high-resolution satellite imagery (see below) in developing ENM predictions. Monkeypox surveillance in Sankuru District in 2006–2007 [5] could not be associated with precise geographic coordinates and, therefore, was not used for model evaluation; similarly, recent case reports from Central African Republic [35] were not accompanied by specific geographic references, and cases reported recently from Sudan [36], [37] likely represent central African strains exported by human activities [24], [38], such that none of these cases was included in our occurrence data sets. Although we strove for highest geographic precision for human case data, all localities used in this study (present and past) are represented by geographic coordinates for villages and a measure of uncertainty; thus, patient privacy and anonymity is not compromised since cases cannot be linked to individuals.

Satellite imagery

To provide environmental information that covers the same time span as the human case records, we selected a dataset from Global Inventory Modelling and Mapping Studies (GIMMS; http://www.landcover.org/data/gimms) that consists of a series of biweekly maximum-value composites of Normalized Difference Vegetation Indices (NDVI) from 1982–2006 with a nominal spatial resolution of 8 km[39], which is adequate for the level of precision of the MPX dataset. NDVI is calculated based on the reflectance values of the red (λ∼0.6 µm) and near infrared (λ∼0.8 µm) bands of multispectral imagery; reflectivity in these two bands depend on the type of land cover (e.g., barren ground, cities, forest, agriculture fields, etc.) and it has been correlated with more specific characteristics of the vegetation canopy (e.g., vegetation condition, biomass, leaf area index, etc.) at the time of data collection [40], thus it is possible to track the status of the vegetation through time by analyzing the variation of this index throughout the year [41]. For these reasons, this vegetation index has been amply used for a great variety of applications including the production of global land cover classifications [42] and in many previous studies of disease distribution [43], [44] among others.

We calculated average NDVI for each month during between 1982 and 1986 (representing the earliest composites from the GIMMS dataset); then, based on these monthly data, we extracted maximum, minimum, mean, and range values for NDVI across the 12 months of each year. A similar dataset was built for 2002–2006, covering a similar time frame of 5 years, to include the most recent GIMMS composites for projections to the present. Although the environmental data for the early period do not coincide exactly in time with the earliest occurrences, satellite imagery simply does not exist prior to the period that we analyzed; however, the majority of cases come from the period covered by the imagery, which we do not believe will bias results significantly. This study focused on the Congo Basin (Figure 1, dashed-line rectangle), so all environmental layers were trimmed to this area for analysis, the area corresponding to the Atlantic Ocean is not included as part of the modeling process.

Ecological niche models

We used the two algorithms most often applied to disease transmission questions to build ENMs for MPX transmission: a genetic algorithm (GARP) [14], [45] and a maximum entropy algorithm (Maxent) [13], [46]. These algorithms were chosen because our dataset consists of reports of human monkeypox cases that resulted from surveillance efforts that likely were not homogeneous through time across the study area; thus, although this dataset is the most complete representation of the distribution of the disease, we expect some of the bias from disease reports to remain in the dataset. Both ecological niche algorithms have been found to perform well using datasets with varying degrees of completeness [47], [48], [49]; they require presence data only, and aim to find non-random associations of disease case occurrences with the environmental conditions they present. These algorithms have been used widely by biologists and ecologists to predict species' distributions [50], [51], [52], [53]. Model transferability (i.e., prediction of distributional phenomena in other places or at other times) has been identified as an aspect of ENM that varies according to the algorithm used, and that could be the source of misleading interpretation of such models; it has been a topic for ongoing discussion in the field [48], [54]. Performance of these two algorithms (GARP and Maxent) has been tested and compared repeatedly in the last few years, with somewhat variable results, in a variety of scenarios [11], [47], [48], [55], [56], [57], [58]; hence, we used both algorithms in the present study to examine potential discrepancies and agreements of the two methodologies.

GARP builds sets of rules that describe environmental conditions associated with localities at which disease transmission events have been recorded [14], [45]. The model is developed through an iterative process of creation, evaluation, modification, and inclusion/exclusion of rules of four basic forms (bioclimatic, atomic, negated, and logistic regression rules); the algorithm stops when the optimization parameter changes by less than 1% from one iteration to the next, or when the maximum number of iterations is reached (1000). We specified that 50% of occurrence points be used by GARP for training models and the rest for internal model evaluation; we ran 100 replicate models relating human MPX cases to NDVI composites for 1982–1986. Of the initial 100 models, we selected those with the best internal performance (i.e., evaluated during the model building process) based on the ‘best subset’ consensus approach: we retained the 20% of the distribution showing lowest omission error and then the central 50% of the remaining models in terms of commission error [59]. This process resulted in selection of 10 models that were summed to identify areas of highest agreement between predictions of different replicate models.

The Maxent algorithm fits a probability distribution based on the environmental conditions at the locations where disease transmission has been recorded, in comparison with the broader set of conditions across the landscape. This algorithm estimates a model in the form of a probability distribution that is constrained to the environmental conditions present at the given locations; thus, the mean values for the environmental variables in the model will be very similar to the averages from the empirical data, the degree of similarity between the model and the data is determined by the regularization parameter (β). The maximum entropy estimate of the probability distribution is considered to be the least biased of all distributions that meet these constrains, making the model more robust regarding missing information [13], [46]. We used default settings in Maxent 3.2.1 (i.e., regularization multiplier  = 1.0, 1500 maximum iterations, 10,000 background points, convergence limit  = 10⁻⁵), but with a random selection of 50% of points for testing and refining the model.

First, we subjected both algorithms to the challenge of predicting human MPX in areas from which no occurrences were available for model training (spatial challenge). We divided the georeferenced WHO human cases spatially into four quadrants based on the median longitude and latitude of all cases [58] to produce two subsets (on-diagonal  =  NW + SE quadrats and off-diagonal  =  NE + SW quadrats) to break up patterns of spatial autocorrelation that could emerge from environmental variables (i.e., closer localities have similar environmental conditions). To test the ability of ENMs to capture the environmental characteristics common to MPX localities and predict other MPX localities in a different geographic area, we trained models based on each of these subsets; representing a more rigorous test than using random subsets in which localities of the testing dataset are allowed to be geographically close to a training locality [58].

We, then, performed a second challenge in which the ability of the models to predict cases in a different period of time was tested. For this second challenge, we built ENMs using all localities from 1970–1986 and environmental conditions in 1980–1986; these models were projected onto environmental conditions for 2002–2006 to be compared with the distribution of recent MPX cases.

For all models, we converted continuous model outputs into binary predictions using a modification of the Least Training Presence Threshold approach of Pearson et al. [57] which consists on the use of the lowest suitability value (logistic output in Maxent and model agreement in GARP) at any of the training points as the threshold value to discriminate between suitable (greater than or equal to threshold value) and unsuitable (lower than threshold value). Considering that some occurrence data may include errors since the exact locality of exposure to MPXV is not known and it could be outside of the village of residence of the reported case [58] (here estimated as E = 10% in light of considerable human movements from infection sites to residences), we chose as a criterion for prediction of presence the suitability value that included (100−E)% of the training data (i.e., we allowed the model to omit up to 10% of the training localities).

Model evaluation

We explored the area under the curve (AUC) of the Receiver Operating Characteristic (ROC) plot as used by Elith et al. [11], which has been accepted broadly as the standard method to assess performance of such models. However, because the standard ROC AUC method has several weaknesses when it is applied to predictions of potential distributions [60], we also performed a variation of the original ROC method (called “modified ROC” from now on) proposed by Peterson et al. [58] that addresses most of these problems. The modified ROC approach has the advantage of prioritizing omission error considerations over commission error [58] and consists of selection of the domain within which the AUC is calculated via specification of an admissible omission error value (E again) that meets the researcher's expectations for predictive ability of the model; this step allows the method to focus on those sections of the ROC curve that are relevant to predictive performance. The AUC is calculated via the trapezoid method, and divided by the AUC value of the null expectation (i.e., the area under the line connecting 0,0 and 1,1; which represents the expected value from a random guess) over the same interval; thus, the modified ROC value departs upward from unity as the model has better predictive ability with respect to random expectations; statistical significance of this approach is calculated via bootstrapping [58], in which 1000 replicate subsamples of 50% of testing data are developed. For the spatial challenge, the ability of ENMs to predict MPX cases in a different geographic region was evaluated by determining how well a model predicts an independent set of cases that were not used to train the model; specifically, models built with the on-diagonal subset of cases were evaluated with the off-diagonal subset and vice versa. Evaluation of the predictive potential of the models for ‘future’ disease transmission events was achieved by assessing whether the model built using environmental conditions and reported cases from the 1980s could anticipate the spatial position of monkeypox cases reported in 2001–2010. One possibility is that environmental conditions and their dynamics have no effect on monkeypox transmission, resulting in no change in its geographic distribution over time; thus, we compared the predictive ability of both the 1980s model per se regarding 2001–2010 cases against its projection onto environmental conditions from the 2000s for prediction of the human cases in 2001–2010. We performed a Mann-Whitney U test based on the AUC ratio values of the respective modified ROC bootstrap analyses [58] to assess which projection predicts the distribution of recent cases better–this latter step allows us to assess whether environmental changes between the two time periods have significant explanatory power as regards monkeypox distributional patterns.