Continental-scale, data-driven predictive assessment of eliminating the vector-borne disease, lymphatic filariasis, in sub-Saharan Africa by 2020

Our proposed approach for addressing the problem of dealing with spatial heterogeneities in LF transmission and extinction dynamics for generating predictions of the effects of interventions at the required aggregated policy levels ranging from district to national scales is essentially hierarchical in nature. In particular, it is based on three coupled components: (1) methods to facilitate assembly/discovery of the relevant model input data at local settings, (2) a data-driven modeling method that can use such data for identifying locality-specific LF models, and (3) an algorithm for combining the heterogeneous simulations from local settings for making upscaled regional predictions. Below, we begin by describing the methods used to execute each of the three components above, and follow this by outlining the scientific workflow we used to implement this hierarchical modeling approach for performing this study.

Developing site-specific parasite transmission models crucially requires input data, viz., exogenous forcing variables and initial state variables, defining a site to facilitate constraining a model’s parameter space so that the effects of local transmission conditions are captured reliably [21, 22, 42, 43]. For modeling impacts of interventions, the type of control carried out in a site together with data on frequency, coverage, and duration are also similarly required. A challenge concerns the availability of all of these data at every site selected in the modeling process; if these data are not available everywhere, then it raises the question of how best to learn about the model inputs for every required location from the available sparse data [33]. These data, when available, are normally also contained within multiple databases maintained by various data providers, with each dataset having unique data access protocols, file formats, and semantics [27, 29, 30]. This heterogeneity and the gaps in source data mean that to derive the input data needed to develop and run site-specific LF models, informatics and analytical tools that can combine empirical data discovery with data integration and estimation protocols will also be required [10, 33, 39]. Below, we describe the combination of data discovery and estimation methods we followed to assemble the site-specific data needed to carry out our landscape-wide modeling exercise under each type/category of input data.

We used a unique LF database containing mf prevalence data for a large number of endemic settings in Africa (a total of 664 data points assembled from the published and gray literatures), which we had previously compiled for use with a Bayesian geostatistical model to construct LF prevalence maps for this continent [44]. These data are mapped in Fig. 1a, and the database includes country and village names, latitude and longitude of the villages, the number of individuals examined and mf positives detected, both stratified by age, as well as key infection-related environmental variables estimated for each site. However, as the existing mf surface map provides only overall mf prevalence values interpolated over a grid rescaled to 1 km × 1 km resolution [44], and as the LF models require age-stratified infection data at every site for parameterization purposes, we needed to estimate the mf age prevalences from the overall community mf prevalence for those sites where such data were unavailable. This was done by converting the community-level overall prevalence into three representative standardized mf age prevalence curves [45], by first constructing an age structure for the community in question using the pertinent national demography profile and then estimating the mf positives in each of the constructed age classes (e.g., 1–10, 11–20, etc.) based on the shape of each of the linear, plateau, and convex mf curves likely or expected to occur in a typical endemic site (see Additional file 1: Figure S1). Note that we consider such derived site-specific age prevalence curves as representing the baseline condition in the year 2000 before large-scale mass drug administration (MDA) programs began to be implemented by the Global Programme to Eliminate Lymphatic Filariasis (GPELF) in Africa. This, however, cannot be expected to be the case for the 11 countries in West Africa (Benin, Burkina Faso, Cote d’Ivoire, Ghana, Guinea Bissau, Guinea, Mali, Niger, Senegal, Sierra Leone, and Togo) that had participated in the Onchocerciasis Control Program (OCP), which began in these countries in 1975 [46, 47]. Apart from vector control (VC), these former OCP countries also had been treated with ivermectin under the community-directed treatment with ivermectin (CDTI) program over a period of 15 years from 1988 to 2002. Such prolonged mass treatment could have significantly reduced the prevalences of LF, as indeed was the case in Sierra Leone, where LF mf prevalences across various sites dropped by rates that ranged between 80 to 100% prior to the delivery of LF MDA under the national LF control program [48]. To account for the impact of CDTI in the above 11 OCP countries, we used this information conservatively and reduced the site-specific map-extracted mf prevalences in all these countries uniformly by 75%, and all baseline model fittings were carried out using these reduced mf prevalences.

Information on the presence of LF vector mosquito species in countries can be obtained from the published literature, including from WHO [49]. Since there are very sparse corresponding spatial data available on ABR (Fig. 1b), we developed a reverse engineering approach to produce model-generated data to overcome this gap. This was done by estimating sets of plausible ABR values from the ensembles of fitted parameter vectors such that the model-generated mf prevalences matched either the observed or geostatistically extracted mf prevalence in a site (see details of the Bayesian melding-based ensemble modeling approach given in Additional file 1). Comparison of such model-generated ABRs for sites where measured ABRs were available showed that we were able to reasonably recover these values using this approach (Additional file 1: Figure S2).

The WHO preventive chemotherapy and transmission control (PCT) databank (http://​www.​who.​int/​neglected_​diseases/​preventive_​chemotherapy/​lf/​en/​) provides the following information about the transmission and control status of LF world-wide: (1) LF endemic countries given by regions; (2) the mapping status of LF prevalence in the countries of these regions provided as completed, in progress, or not started; (3) information on drug regimens applied during the years in which MDA was administered; (4) years of annual MDA with the numbers of implementation units (IUs) covered in individual years at the country level; and lastly (5) information on national program drug coverages achieved during the period 1999–2013 for all countries undergoing annual LF MDA. In Africa, as of 2013 these data show that 23 countries have completed mapping status for LF; these countries have either already implemented several rounds of annual MDA or are yet to start MDA (Additional file 1: Table S2).

An examination of the MDA data available from the PCT databank (Additional file 1: Table S2) showed that not all endemic IUs received MDA since the start of mass treatment in a given country. An analysis of the data showed a clear occurrence of three distinct phases of MDA implementation in the period 2000–2013 in terms of the numbers of IUs starting annual mass treatments in each country (Table 1). These three phases can be termed as the initial phase covering the first 1–3 years of mass treatments, the expansion phase (lasting for 4–5 years after the initial phase), and the later/saturation phase in which almost all endemic IUs in a country started implementing annual MDAs. In addition, we also found a fourth group of IUs present in all LF endemic African countries except Togo, where MDA apparently had not begun or no MDA information was available by 2014–2015. We presumed that these IUs would begin annual MDAs from 2016 in all the simulations described below.

Information on VC methods for African countries was assembled from the country-specific Demographic and Health Surveys (DHS) sites (www.​dhsprogram.​com) as well as from WHO [49]. In the majority of countries, household usage of bed nets, mainly insecticide-treated nets (ITNs), has been the mainstay of the approach to control mosquitoes primarily under the Roll Back Malaria program (www.​rollbackmalaria.​org). We therefore used the ITN coverage data available for the period 2000–2012 for countries in Africa from the MAP database (Malaria Atlas Project, Oxford: www.​map.​ox.​ac.​uk) in order to derive estimates of VC efforts in this study. These data are available at the administrative level 1 (i.e., region/province level, Admin1 level hereafter), but tend to exhibit missing information for a few administrative units in one or more years, which complicated the calculations of annual VC coverage rates. In order to overcome this problem, we first created smooth maps of VC coverage for 2000–2012 using the available Admin1 data via simple kriging carried out using ArcGIS. Then, the national-level annual VC coverage values were extracted by simple averaging of the smoothed values occurring within each country boundary. These extracted annual coverages are shown in Fig. 1c and Additional file 1: Figure S3. These data showed the occurrence of high temporal variability (i.e., they describe sudden jumps from a low to very high value or vice versa in the coverage values) between years within a country. We smoothed this temporal variability in the country-level data by modeling trends in the annual values observed in each country. As shown in Additional file 1: Figure S3, this predictive model was able to capture the general trend in the VC coverage levels in the studied countries for the period 2000–2012. Thus, in the present analysis, we used the smoothed annual coverage values for 2000–2012 and the predicted values for the future time period in all the LF intervention scenarios where VC was applicable.

The technical details of the LF transmission models used and the Bayesian melding approach employed to calibrate these models to local data, as well as specifics of how LF MDA and VC interventions are simulated, have been described extensively previously [21, 22, 50, 51] and are outlined in Additional file 1. Here, our focus is on the coupling of this data-driven modeling framework to input data assembled at local settings (here at the village level) as a means for better capturing the effects of local spatial heterogeneity in LF transmission dynamics when making predictions of the impacts of applied interventions at higher spatial scales. We describe the algorithm that used this coupling for facilitating modeling of the outcomes of the various currently applied or proposed LF intervention strategies within each study country below.

Our hierarchical algorithm for discovering and using local LF models for simulating the effects of interventions within each country comprised the following steps:

In addition to assessing the achievement of elimination under the current strategy, several remedial strategies were also modeled. All remedial strategies were simulated beginning in 2016 to assess when elimination can be achieved under these improved strategies (via increased coverage, frequency of drug treatment, or by switching to a new combination drug regimen (Table 2)). Analysis of remedial strategies was not undertaken for those countries that were predicted to have achieved LF elimination under their current approaches. Table 2 outlines all intervention strategies, current and remedial with corresponding coverages, used in this work.

We implemented the scientific workflow depicted schematically in Fig. 2 to perform the spatially hierarchical modeling exercises reported here. Figure 2a describes the sequence of steps involved in calibrating the LF model to local conditions, simulating the impacts of interventions, and assessing prospects for elimination. Figure 2b defines the data sources and pipelines we used for assembly, integration, and transformation of data into the information required by the LF transmission models as inputs. Briefly, we began the modeling process by determining whether key model input data (e.g., baseline ABR: the number of bites, on average, a person receives per year) and/or LF infection (mf age profile) were directly available from existing databases or needed to be derived for a site. Once these input data were either collated or estimated, the relevant LF model was parameterized for individual localities using the Bayesian melding algorithm, and the quantities of interest for modeling the prospects of parasite elimination, viz., the local infection and vector biting thresholds, were calculated via the fitted models. Intervention modeling was then undertaken using the available or estimated data on MDA coverage and duration and on supplemental VC, for predicting the timelines to crossing the model-derived elimination thresholds in each site. Note that if intervention data are not available, simulations can also be done using hypothetical intervention scenarios at this stage. This exercise allowed us to determine if implemented and/or hypothesized interventions will lead to parasite elimination by a set target date. In cases where the modeling analysis indicated that interventions (actual or hypothesized) are unlikely to achieve elimination by a given target year for a site (e.g., by 2020), simulations for estimating the required remedial measures (e.g., increasing MDA coverage and frequency, switching to annual ivermectin-diethylcarbamazine-albendazole (IDA), including or increasing supplementary background VC using long-lasting insecticidal nets (LLINs)) to meet the goal of elimination for various target end dates were undertaken, and the outputs evaluated.

Figure 3a shows an example of the fits of the anopheline LF model (broken bars) to the three characteristic age profiles (boxes depicting mean prevalences predicted in each respective age class with the vertical lines showing the 95% confidence intervals of these means) of expected mf prevalences derived for a site in Senegal depicting an overall mf prevalence of 32.1%, as obtained using our site-specific data-driven modeling algorithm. The corresponding predictions of overall mf prevalences and ABRs from the model fits to each age profile of infection are shown in Fig. 3b and c, respectively; these results highlight not only how our modeling approach can allow estimation of LF infection patterns and ABR inputs for a site, but also how we may use the model outputs to learn about which specific type of model may capture the dynamics of parasite transmission in a particular locality. This ability is demonstrated by the results shown in Fig. 3b, which compares the overall mf prevalences predicted by the ensembles of fits obtained for each type of age curve (histogram bars) with the mf prevalence extracted for this site (solid line = 32.1%). The median predicted prevalences of the fitted models are given by the dashed vertical lines, and the distance between these values and the derived mean mf prevalence can be used to assess which of the three modeled age profiles may best give rise to the overall mf prevalence extracted for a site. The results show, for example, that while the predictions of the plateau and convex models are close to the overall mf prevalence, the median of the linear model predictions is too distant from the mean mf prevalence value, suggesting that it is a less reasonable contributor to overall infection prevalence in this Senegalese site. Figure 3d is a cluster plot showing the relative contributions of each age-infection profile to all of the site-specific overall mf prevalence data we used to discover LF models for the country of Senegal, and shows that as a whole, the plateau-type age-infection model contributed most to infection for the sampled Senegalese sites, followed in importance by the convex and the linear age prevalence models. This cluster analysis of the ensemble of age-stratified model fits to overall mf prevalences obtained for a spatially representative sample of sites in each country was used to discover the corresponding LF models in each African endemic country investigated in this study.

We used the subset of 10% best-fitting parameter vectors selected via the cluster analysis of model fits to each selected spatially representative study site (Fig. 3d) in order to make inferences regarding the expected distribution of mf elimination thresholds within a country. Figure 4 portrays the values of mf breakpoints estimated for a selection of study countries using these best-fitting site-specific parameter vectors following the numerical method outlined in our previous work [21, 22, 50, 53]. Two features of the results depicted in Fig. 4 are immediately apparent. First, the results indicate that considerable variability in LF mf breakpoint values may occur within each country, and second, that both the distribution and mean values of these breakpoints will vary markedly between countries. Table 3 provides the mean and the 5th and 95th percentile values of the mf breakpoints obtained for each study country at both the threshold biting rate (TBR) (important as prevalence targets for assessing the impact of supplementary VC in breaking LF transmission [22, 50]) and at the prevailing ABR values in a site (important for serving as mf threshold targets for evaluating the impact of MDA [22]). The data support the impression gained from Fig. 4, and the conclusions made in our previous studies [20‐22], that mf breakpoints will vary significantly between LF endemic regions, including in the present case at the country level, will be higher at TBR, and will invariably also be much lower (with 5^th percentile values lower by between twofold to fourfold) than the WHO-suggested threshold of 1% mf prevalence (Table 3).

The principal focus of this study was to use models in conjunction with intervention data estimated or assembled at spatially representative sites in order to evaluate if currently applied interventions in LF endemic African counties will meet the goal set by WHO for achieving LF elimination on the continent by the year 2020. As described in Methods, we used a hierarchal modeling framework to undertake this analysis in order to take a fuller account of the impacts of spatial heterogeneity in transmission conditions and ultimately with respect to interventions applied across a country. Our approach was thus to first assess the prospects of eliminating LF transmission at local sites and then to aggregate this information to the IU level (this is commonly the district), and following this to the country level. In this approach, we therefore considered an IU to have accomplished the goal of LF elimination when all sites sampled within its boundary were predicted to have their infection levels reduced from baseline to below their own 95% mf prevalence elimination threshold (see Methods [22, 52]) as a result of applied interventions, and a country to have accomplished elimination at the time point when all corresponding IUs are estimated to have met the goal of elimination. Meeting the year 2020 LF elimination target was primarily assessed by inspecting if timelines predicted by local models crossed site-specific mf prevalence thresholds by 2020. Note that while we are able to estimate site-specific models and infection-related data (baseline age-stratified mf prevalences, ABRs, and mf breakpoints) to carry out the modeling analyses, intervention data (MDA start, duration and coverage, VC type and coverage) were only available at the national or IU levels. Thus, these data at a higher hierarchical level were applied uniformly to all local sites when modeling interventions in this work.

Altogether, a set of 12 intervention scenarios were investigated (Table 2). These intervention scenarios include variants of annual and biannual MDAs, and annual IDA — a triple-drug regimen comprised of standard doses of ivermectin-diethylcarbamazine-albendazole (IVM-DEC-ALB) that were recently reported to be highly efficacious against LF infection in non-onchocerciasis LF endemic areas [54]. The base intervention investigated for assessing if current MDA programs are able to bring about LF elimination by 2020 pertains to that labeled as MDA1 in Table 2, which denotes annual MDA with supplemental VC applied at the country-level coverages reported in the WHO LF PCT databank for each IU cohort in the case of MDA (Table 1), and at a coverage of 58.78% for the usage of ITNs (see Additional file 1: Figure S3). MDA2 stands for annual MDA at the most recent reported coverage in the WHO LF PCT databank, with VC coverage raised to 80% — the optimal VC coverage thought to be ideally achievable under the WHO malaria control program; MDA3 for annual MDA at 80% coverage with VC coverage at 58.78%; and finally MDA4 where annual MDA and VC both have coverage values of 80%. All four variants of MDA strategies were also considered in the case of biannual and annual IDA remedial mass treatment plans as outlined in Table 2. All remedial plans were simulated from the year 2016.

Figure 5 depicts the patterns of timelines to LF elimination predicted both within a country (Fig. 5a) and across a selection of representative African countries (Fig. 5b) (results shown for the MDA1 scenario). Note that site-specific mf prevalence values representing a 95% elimination probability at the TBR [22] were used as targets for signifying elimination in this exercise because of the involvement of VC in all the MDA variants investigated (Table 2). The results portrayed in Fig. 5a show firstly how the staggered delivery of annual MDA may affect the timelines to eventual LF elimination in an individual country. The simulations are based on the expected declines modeled in four cohorts of early to late phase IUs in Kenya (Table 1), which are assumed to have been treated sequentially during the phased roll-out of the MDA program in that country. The results show that, depending on the start year of MDA and base mf prevalences in these different cohorts, times to LF elimination will vary between treated IUs in a country, with delayed treatments (e.g., the late phase IU cohort beginning treatment only in year 2016) lengthening the time to LF elimination for the entire country. This result highlights the crucial importance of considering both the start years as well as baseline mf prevalences of IUs when modeling the feasibility of current MDA programs for achieving LF elimination in a country.

A problem with modeling this potential IU cohort effect on timelines to LF elimination is that the WHO PCT database currently does not provide details of the IUs in a country that underwent MDA at various start dates; only numbers of IUs that began annual MDA at different time periods are provided (Table 1). As identity of IUs is required to estimate their likely baseline mf prevalence and ABRs from our maps to serve as model inputs, this lack of information means that we currently can only attempt to quantify the likely cohort effect in a country via considering possible best- and worst-case scenarios that may apply. Here, we indicate that the best case is to consider that IUs are recruited sequentially into national programs based on their endemicity level (i.e., IUs with higher baseline mf prevalence are recruited earlier than IUs with lower endemicity). This scenario can then be contrasted with the null and worst-case scenario in which IUs are recruited into annual MDA programs randomly (i.e., the baseline endemicity of IUs does not matter in determining cohort membership).

Figure 5b compares the timelines to LF elimination by IUs under these two scenarios graphically for 14 of the 36 sub-Saharan African LF countries, while Table 4 displays the actual years by which 100% of IUs are expected to achieve LF elimination in all 36 countries under each scenario. The results depicted in Fig. 5 and given in Tables 4 and 5 highlight as expected that the numbers of rounds of annual MDA required for breaking LF transmission will be lower (by between 1 to 5 years) in the case of the sequential scenario compared to the scenario in which IUs are randomly recruited into annual MDA programs. Nonetheless, a major finding is that for both scenarios the majority of African countries, except for Egypt and Togo under randomly assigned MDAs, and Egypt, Ghana, and Togo in the case of sequential treatments, will not meet the goal of achieving LF elimination by 2020 when local extinction thresholds are used in the modeling exercise. By contrast, if the WHO threshold of 1% mf prevalence is used as a target, while 12/36 of these countries are found to be able to achieve LF elimination under the current MDA program in the random IU treatment scenario by this time point, a corresponding 24/36 or 66.6% of countries are predicted to be able to achieve LF elimination under the same MDA program in the sequential treatment scenario. These results indicate that the way in which IUs are recruited into MDA programs, their original endemicity levels, treatment coverage, and the threshold value set as targets for signifying transmission breakage will all combine to govern the expected timelines to achieving LF elimination in any given country.

We evaluated next the effect that switching from annual MDA plus VC at current coverages from year 2016 to various remedial measures could have in accelerating the progress to LF elimination. This exercise was carried out under the sequential program expansion scenario, which we believe is a more reasonable model followed in reality compared to the random phased IU recruitment scenario, and noting that Togo, Ghana, and Egypt are expected to meet the goal of LF elimination by 2020 under the current sequential annual MDA program (Table 4). The predicted LF elimination years as a result of switching to the various currently proposed remedial drug plus VC strategies in comparison to the base annual MDA plus VC at current coverages strategy (Table 2) are summarized for all 36 countries in the form of boxplots in Fig. 6. These results show that while switching from any of the annual MDA strategies to biannual MDA and annual IDA variants will decrease the expected year of achievable LF elimination, none of the strategies will be able to eliminate LF transmission across all endemic countries on the continent by year 2020 (see also Additional file 2: Movie S1, Additional file 3: Movie S2, and Additional file 4: Movie S3). Figure 6 shows that increasing the coverages of MDA and VC from current levels to 80% will have only little impact within each MDA variant. The results also show that switching to annual IDA compared to biannual two-drug regimens will have only a small positive effect in reducing the mean year of LF elimination in Africa even in the case of the most optimal subvariant strategy, viz., providing drugs and VC at a coverage of 80%, respectively (Fig. 6).

Figure 7 presents the predicted elimination years due to the application of two best and worst performing variants in each mass treatment plus VC plans — the worst (the left panel plots showing the impacts of MDA1, bi-MDA1, and IDA1) and the best (the right panel plots depicting effects of MDA4, bi-MDA4, and IDA 4, respectively) by order in which individual countries in Africa may be able to eliminate LF (with the elimination year for each country under all remedial strategies shown in Additional file 1: Table S3). As shown in Fig. 7b, the results highlight that achieving LF elimination is not possible in all 36 endemic countries of sub-Saharan Africa even by 2025 using the best variant of a remedial strategy implementing annual MDA. However, if MDA delivery is switched to biannual (Fig. 7c and d), achieving LF elimination by 2025 is possible in the majority of these countries. All annual IDA-based remedial strategies (Fig. 7e and f) clearly predict more optimistic timelines to LF elimination across the continent (at the latest by 2023).