Environmental suitability for lymphatic filariasis in Nigeria

17 Apr 2018

This page was adapted from a poster presented at the British Society for Parasitology Spring Meeting 2018 by Obiora Eneanya, Jorge Cano, Ilaria Dorigatti, Rachel Pullan, Tini Garske, Christl Donnelly

Background

  • Lymphatic filariasis (LF) is a mosquito-borne Neglected Tropical Disease (NTD), which in its advanced stages can manifest as severe lymphedema, and/or hydrocele (Figure 1).
  • Prevention and treatment are by the use of bed nets and treating entire endemic communities with ivermectin + albendazole + diethylcarbamizine (DEC).
  • In this work we:

    1 - Describe and map the ecological niche of LF in Nigeria.
    2 - Estimate population living in areas that are environmentally suitable for disease transmission 

The data

• Data were collected during mapping surveys conducted by the Federal Ministry of Health Nigeria from 2000-2013.
• We had 1378 site-level data points covering all 36 States and the Federal Capital Territory (Figure 2).
• For analysis, ‘Presence’ was considered when at least 1 LF case was recorded, while ‘Absence’ was considered if no LF cases were recorded in survey locations.
• All survey sites tested for the presence of filarial antigenemia using immunochromatographic card test.
• Environmental covariates with known biological plausibility for LF occurrence were considered as predictors (Figure 3).

The Model

• Presence/absence data were fit to 7 different ecological niche modelling algorithms.
• Data was partitioned into 2, with a random sample of 30% retained for evaluation and 70% used to calibrate the model.
• An iteration of 100 model runs was performed for each of 7 model algorithms and the evaluation values for each run was stored and then averaged.
• The area under the receiver operator curve (ROC) and True Skills Statistics (TSS) were used as measures for model performance.
• An ensemble of two best performing modelling algorithms was used for final projection (here, machine-learning algorithms, generalised boosted models + random forest).

Results

A continuous risk map of environmental suitability of LF was projected on a geographic space at a spatial resolution of 1km x 1km (Figure 4). Median ROC for the ensemble model was 0.972 (95%CI: 0.798-0.984), and median TSS was 0.815 (95%CI: 0.812- 0.824).

A suitability threshold of 0.711 with sensitivity of 95% and specificity of 96.2% provided the best discrimination between presence and absence values. This was then used to generate binary maps, delineating areas into ‘suitable’ or ‘unsuitable’ (Figure 5).

Estimating population living in LF risk areas

By overlaying a gridded map of population density for Nigeria, we used the binary maps generated to estimate that 110 (95%CI: 106- 127) million people live in areas that are environmental suitable for LF transmission. This was calculated on a 100m x 100m scale.

Conclusions

• Machine-learning and ensemble modelling are powerful tools to map disease risk and are known to yield more accurate predictive models.
• The resulting maps provides a geographical framework to target control efforts and assess its potential impacts.

References

Funding

OE thanks the Commonwealth Scholarship Commission, United Kingdom for doctoral studentship funding. We also thank MRC for Centre funding.