Developing a predictive model for an emerging epidemic on cassava in sub-Saharan Africa

Abstract

The agricultural productivity of smallholder farmers in sub-Saharan Africa (SSA) is severely constrained by pests and pathogens, impacting economic stability and food security. An epidemic of cassava brown streak disease, causing significant yield loss, is spreading rapidly from Uganda into surrounding countries. Based on sparse surveillance data, the epidemic front is reported to be as far west as central DRC, the world's highest per capita consumer, and as far south as Zambia. Future spread threatens production in West Africa including Nigeria, the world's largest producer of cassava. Using innovative methods we develop, parameterise and validate a landscape-scale, stochastic epidemic model capturing the spread of the disease throughout Uganda. The model incorporates real-world management interventions and can be readily extended to make predictions for all 32 major cassava producing countries of SSA, with relevant data, and lays the foundations for a tool capable of informing policy decisions at a national and regional scale.

Figure 1

Maps representing the rasterised data-driven model layers: (A) the model host landscape, representing the number of cassava fields at a 1 km² resolution derived from CassavaMap. (B) The vector abundance layer represents the relative abundance of B. tabaci at a 5 km resolution derived from the Ugandan CBSD field surveys.

Figure 2

Overview of the three summary statistics used for ABC parameter estimation. Both ${\mathrm{S}}_{\mathrm{cen}}$ and ${\mathrm{S}}_{\mathrm{nat}}$ statistics were derived by calculating the proportion of survey points in a given year that were reported as positive within a given region: (A) represents the two non-overlapping areas of Uganda covered by ${\mathrm{S}}_{\mathrm{cen}}$ and ${\mathrm{S}}_{\mathrm{nat}}$ and (B) summarises the values derived when applying ${\mathrm{S}}_{\mathrm{cen}}$ and ${\mathrm{S}}_{\mathrm{nat}}$ to the Ugandan national survey data covering the period 2004 to 2017. The dotted black line indicates the divide between the fitting data, ${\mathrm{d}}_{\mathrm{real}}^{\mathrm{fit}}$, and the validation data, ${\mathrm{d}}_{\mathrm{real}}^{\mathrm{val}}$. (C) Overview of the summary statistic ${\mathrm{S}}_{\mathrm{grid}}$ highlighting the survey year, up to and including 2010, in which a CBSD infected field was first detected in a given quadrat. If no positive surveys were reported prior to 2010, as in the case of quadrat 10, the quadrat is shaded green. If no surveys were carried out prior to 2010, the quadrats have been excluded from the plot. Quadrat indices are shown in the top left corner of each quadrat.

Figure 3

Posterior distribution of the three parameters estimated using the fitting data from 2004 to 2010, ${\mathrm{d}}_{\mathrm{real}}^{\mathrm{fit}}$. The parameters are the transmission rate, β, the kernel exponent, α, and the proportion of dispersed inoculum that remains in the source cell, p. The posterior is composed of 1440 fitting simulations that met the fitting criteria out of a trial of 233,600 fitting simulations. Five outliers, indicated by black crosses, were excluded from sparsely sampled parameter space (Supplementary Fig. 7).

Figure 4

Overview of the districts where clean seed dissemination programmes were primarily carried out: (A) shows the location of the districts in Uganda as well as the region covered by ${\mathrm{S}}_{\mathrm{cen}}$ (dotted line) and (B) summarises the proportion of field surveys that reported CBSD in a given year from each district.

Figure 5

Time series probability distributions for the statistics (A) ${\mathrm{S}}_{\mathrm{nat}}$ and (B) ${\mathrm{S}}_{\mathrm{cen}}$ for the subset of validation simulations that pass within the tolerances of the fitting and validation criteria. (C,D) The same statistics but without applying any tolerances to illustrate the unconstrained behaviour of the parameterised model. The red line indicates the target value of each statistic derived from surveillance data. Tolerances are indicated by green arrows. The central blue band is the median ± 10% and each gradation beyond is a further ± 10% from the median. The dotted black line indicates divide between ${\mathrm{d}}_{\mathrm{real}}^{\mathrm{fit}}$ and ${\mathrm{d}}_{\mathrm{real}}^{\mathrm{val}}$.

Figure 6

Comparison of a single validation simulation that passes the fitting and validation criteria with the real-world surveillance data from 2005 to 2017. We define the fitting criteria as the values selected for the three tolerances for parameter estimation using Ugandan survey data from 2004 to 2010: ${\mathrm{\epsilon }}_{\mathrm{cen}}^{\mathrm{fit}}=0.25$, ${\mathrm{\epsilon }}_{\mathrm{nat}}^{\mathrm{fit}}=0.25$, ${\mathrm{\epsilon }}_{\mathrm{grid}}^{\mathrm{fit}}=0.48$ and the validation criteria as the comparison of the simulated surveillance data to the Ugandan survey data covering the validation period from 2011 to 2017, ${\mathrm{d}}_{\mathrm{real}}^{\mathrm{val}}$, using the two ${\mathrm{S}}_{\inf }$ statistics with tolerances ${\mathrm{\epsilon }}_{\inf }^{\mathrm{val}}=0.25$. Red crosses indicate an observation of CBSD at the field-level. Green crosses indicate no CBSD observed.