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

Abstract

Background: There are growing demands for predicting the prospects of achieving the global elimination of neglected tropical diseases as a result of the institution of large-scale nation-wide intervention programs by the WHO-set target year of 2020. Such predictions will be uncertain due to the impacts that spatial heterogeneity and scaling effects will have on parasite transmission processes, which will introduce significant aggregation errors into any attempt aiming to predict the outcomes of interventions at the broader spatial levels relevant to policy making. We describe a modeling platform that addresses this problem of upscaling from local settings to facilitate predictions at regional levels by the discovery and use of locality-specific transmission models, and we illustrate the utility of using this approach to evaluate the prospects for eliminating the vector-borne disease, lymphatic filariasis (LF), in sub-Saharan Africa by the WHO target year of 2020 using currently applied or newly proposed intervention strategies. METHODS AND RESULTS: We show how a computational platform that couples site-specific data discovery with model fitting and calibration can allow both learning of local LF transmission models and simulations of the impact of interventions that take a fuller account of the fine-scale heterogeneous transmission of this parasitic disease within endemic countries. We highlight how such a spatially hierarchical modeling tool that incorporates actual data regarding the roll-out of national drug treatment programs and spatial variability in infection patterns into the modeling process can produce more realistic predictions of timelines to LF elimination at coarse spatial scales, ranging from district to country to continental levels. Our results show that when locally applicable extinction thresholds are used, only three countries are likely to meet the goal of LF elimination by 2020 using currently applied mass drug treatments, and that switching to more intensive drug regimens, increasing the frequency of treatments, or switching to new triple drug regimens will be required if LF elimination is to be accelerated in Africa. The proportion of countries that would meet the goal of eliminating LF by 2020 may, however, reach up to 24/36 if the WHO 1% microfilaremia prevalence threshold is used and sequential mass drug deliveries are applied in countries.

Conclusions: We have developed and applied a data-driven spatially hierarchical computational platform that uses the discovery of locally applicable transmission models in order to predict the prospects for eliminating the macroparasitic disease, LF, at the coarser country level in sub-Saharan Africa. We show that fine-scale spatial heterogeneity in local parasite transmission and extinction dynamics, as well as the exact nature of intervention roll-outs in countries, will impact the timelines to achieving national LF elimination on this continent.

Keywords: Data discovery; Data-driven parasite transmission modeling; Hierarchical modeling; Lymphatic filariasis; Mass drug administration; Neglected tropical diseases; Parasite elimination programs; Parasite transmission heterogeneity; Scientific computational discovery of knowledge; Spatial scale; Sub-Saharan Africa; Vector control; Vector-borne diseases.

Fig. 1

Mapped inputs for modeling the baseline transmission dynamics and effects of interventions against lymphatic filariasis (LF) in sub-Saharan Africa. a A smooth map of the estimates of LF prevalence is shown. The map was created using a multivariate Bayesian generalized linear spatial model, as described in [44]. b Smooth maps of the annual biting rates (ABRs) of Anopheles and Culex mosquitoes were created by simple kriging of ABR data obtained from literature searches and public databases (e.g., the Malaria Atlas Project (MAP)/Malaria Risk in Africa (MARA) databases). The observed data points are also shown. The Culex distribution is patchier than that of Anopheles, so we consider the Anopheles model to apply wherever Anopheles mosquitoes are implicated in transmission (Table 1). Only in those areas where Anopheles mosquitoes are not implicated at all did we use the Culex model. Note, however, that given the sparseness of the ABR data as shown on the map, we used model-estimated ABR values in the modeling exercise described in the text (see Methods). c Country-level coverages of bed nets (i.e., insecticide-treated nets (ITNs) interpolated from Admin1 data. Smoothed annual maps were developed for 2000–2012; here we show data for years 2000, 2007, and 2012

Fig. 2

Modeling workflow and data inputs for the modeling analysis obtained from literature searches and publicly available databases. a The top schematic depicts the hierarchical modeling steps employed to quantify the expected impacts of MDA-based intervention programs to predict timelines to achieve the elimination of lymphatic filariasis (LF) from an endemic country. b The bottom diagram shows the steps, mainly involving geostatistical mapping, on how the data inputs required to initialize the LF model were obtained from several databases, either publicly available or created for this study from literature searches. Note that although we began examining the development of an ABR smooth map, the sparseness of ABR data precluded their reliable construction for use in this study. We instead used model fits to mf prevalence in a site to estimate ABR input values (see Methods)

Fig. 3

Learning ensembles of local LF transmission models. a An example of baseline fits to three age profiles (boxes depicting mean prevalences predicted in each respective age class with the vertical lines showing the 95% confidence intervals of these means) derived for a site in the country of Senegal with overall mf prevalence of 32.1%. b The corresponding predictions of overall mf prevalence are shown in a histogram where the solid vertical line represents the known overall mf value and the dashed vertical line represents the median of the predicted distribution. c The predicted distribution of baseline annual biting rates (ABRs) from the fits to each age-infection profile is shown with the median indicated by a vertical dashed line. The inset plot zooms in on the distributions in the lower ABR region. d A cluster plot showing the relative contributions of each age-infection profile to the pooled fits across all modeled sites in Senegal. For this country, the plateau-type age-infection model contributed most to infection for the sampled sites, followed in importance by the convex and the linear age models

Fig. 4

Histograms of breakpoints and threshold biting rates aggregated across modeled sites in a country. Solid vertical lines represent the mean value of the distribution, and dashed vertical lines represent the 5^th and 95^th percentile values, respectively. The histograms show that there is variation in transmission thresholds both within and between countries

Fig. 5

Patterns of elimination timelines of lymphatic filariasis (LF) by calendar years in sub-Saharan Africa. a An example of the effects of staggered annual MDA implementation in cohorts of implementation units (IUs) across a country on LF elimination trajectories, using data and model simulations for Kenya. The modeled decline in the overall mean microfilariae (mf) prevalence as a result of LF intervention is shown, where each cohort had a different MDA start year: 2001, 2003, 2011, and 2016, respectively. The elimination year for each cohort is shown by an open circle on the x-axis of the respective subplot. b Patterns in timelines to LF elimination based on annual MDAs provided to IUs either randomly or in a phased sequential manner starting with provision of treatments to the highest prevalence IUs first. The vertical bars show fractions of IUs in a country achieving LF elimination by calendar years. The results for sequential coverage of IUs are shown in orange (i.e., IUs with higher baseline endemicity receiving MDA earlier than lower endemicity IUs), and those for random coverage of IUs are shown in blue. The results for the random selection of IUs are more pessimistic than those for the sequential approach. Data on the years and number of IUs implementing annual MDA and drug coverages are from the WHO LF PCT databank. Future LF interventions were simulated for the current MDA and VC coverages for a given country

Fig. 6

Timelines to the elimination of lymphatic filariasis (LF) from sub-Saharan Africa under various intervention strategies. Distributions of calendar years when LF elimination is predicted to be achievable in sub-Saharan Africa under the 12 considered intervention scenarios (see Table 2). These values were calculated by pooling together model-predicted country-specific LF elimination years. The country-specific LF elimination years were calculated as the year by which community-level mf prevalences in all selected spatially representative sites from a given country are predicted to be reduced below their respective 95% elimination probability thresholds (see Table 3). The horizontal line indicates the year 2020 — the target year set for global LF elimination. The error bars show the 2.5^th and 97.5^th percentile values

Fig. 7

Variability in country-specific timelines to lymphatic filariasis (LF) elimination in sub-Saharan Africa. Distributions of the calendar years of LF elimination are shown for 6 out of the 12 considered remedial intervention scenarios (see Table 2). Countries are depicted in the graphs ranked by the year of elimination. Note that some countries are able to meet elimination by 2020 under their current strategy, and so would not need consideration of alternative strategies (Egypt, Ghana, and Togo). The results are shown for MDA1 and MDA4 (a and b), Bi-MDA1 and Bi-MDA4 (c and d), and IDA1 and IDA4 (e and f) (see Table 2 for descriptions). The left panel plots are for the current coverages of MDA and VC, while the right panel ones are for optimal 80% coverages for both. The widths of the boxplots denote the 25^th and 75^th percentile values of the calendar years shown on the x-axis. The vertical lines indicate the year 2020 — the target year set for global LF elimination (red) and the continental-wide model-predicted median elimination year for each strategy (blue). The whiskers show the 90% confidence interval computed using the 5^th and 95^th percentile values