Spatio-temporal modelling of weekly malaria incidence in children under 5 for early epidemic detection in Mozambique

Abstract

Malaria is a major cause of morbidity and mortality in Mozambique. We present a malaria early warning system (MEWS) for Mozambique informed by seven years of weekly case reports of malaria in children under 5 years of age from 142 districts. A spatio-temporal model was developed based on explanatory climatic variables to map exceedance probabilities, defined as the predictive probability that the relative risk of malaria incidence in a given district for a particular week will exceed a predefined threshold. Unlike most spatially discrete models, our approach accounts for the geographical extent of each district in the derivation of the spatial covariance structure to allow for changes in administrative boundaries over time. The MEWS can thus be used to predict areas that may experience increases in malaria transmission beyond expected levels, early enough so that prevention and response measures can be implemented prior to the onset of outbreaks. The framework we present is also applicable to other climate-sensitive diseases.

Figure 1

Statistics comparing model performance. Root-mean-square-error (RMSE) and 95% coverage probabilities (CP) for: the model that only uses the climate data (M1); the model without the climate data and purely based on the spatio-temporal random effects; the model that combines both the climate data and spatio-temporal random effects (M3). The horizontal dashed line in the lower panel corresponds to 0.95 which is the nominal level of coverage for the prediction intervals.

Figure 2

Map of Mozambique’s administrative districts. Map of the administrative districts of Mozambique with four highlighted districts as indicated by the legend.

Figure 3

Prediction results for the districts of Machaze, Massangena, Chibabava, and Sanga over the following 26 weeks after the 35^th week in 2016. The left panels show the posterior mean (solid line) and 95% credible intervals (dashed lines) for malaria incidence. The right panels show the exceedance probability for a relative risk threshold of 2. Dots show observed incidence. The results for the 9th week in 2017 of Massangena district are not reported because the incidence data were not available.

Figure 4

Prediction results for the 1^st week of 2017. The left panel shows the posterior mean of malaria incidence for each of the 142 districts. The right panel shows the exceedance probability for a relative risk threshold of 2.

Figure 5

Prediction results for the districts of Machaze, Massangena, Chibabava, and Sanga over the following 8 weeks after the 1^st week in 2017. The left panels show the posterior mean (solid line) and 95% credible intervals (dashed lines) for malaria incidence. The right panels show the exceedance probability for a relative risk threshold of 2. Dots show observed incidence. The results for the 26th week (the 8th week of the hold-out sample) of Massangena district are not reported because the incidence data were not available.