Malaria is one of the leading causes of morbidity in Farafangana, an urban area in the southeastern coast of Madagascar with around 35,000 inhabitants. Predictive models of incidence are needed to strengthen prevention measures in case of an epidemic risk. In a context of large scale vector-control interventions, with changing mosquito density and behaviour, entomological surveillance data could be useful to better forecast malaria incidence. Our primary objective was to build a predictive model of malaria incidence (up to two months in advance) in the primary healthcare center (PHC) of Farafangana, including past incidence (at least two months back), climatic, environmental, vector-control and entomological covariates. Our secondary objective was to quantify how entomological covariates might have improved the model fit.

Diagnosed malaria incidence data at the PHC of Farafangana came from the Fever Sentinel Surveillance Network of the Institut Pasteur de Madagascar. Climatic data (temperature and precipitations) and environmental data (Normalized Difference Vegetation Index - NDVI) were extracted from the International Research Institute for Climate and Society (missing values were imputed by exponential smoothing). Vector-control covariates, insecticide treated nets from mass distributions and indoor residual spraying, were binary (1 value means effective). From January 2014 until March 2017, human landing mosquito collections were performed every two months for two consecutive nights (from 6 pm to 6 am), inside and outside five houses spread over one central district, close to the PHC. As entomological data were not measured continuously, we made the hypothesis that they could be repeated until the next capture session. We focused our work on the three most abundant vectors: Anopheles gambiae (n=209), An. coustani (n=215) and An. funestus (n=19). Aggressiveness was calculated for each vector species and capture time [number of bites per human and per evening (from 6 to 10 pm), night (from 10 pm to 2 am) or morning (from 2 to 6 am)]. Exophagy percentage was calculated for each species as the number of captures outside houses on the total number of captures (except for An. funestus due to a small number of captures). These entomological data were transformed on a weekly-basis, resulting in 167 weeks of observations. Lags between incidence and non-entomological covariates were determined by cross-correlation maps for quantitative variables and by univariate analyses for binary variables. Accordingly, lagged covariates were built. Only non-correlated lagged covariates were kept (|Pearson's correlation coefficient|<0.70) to avoid colinearity. Negative binomial regression models were built with or without entomological covariates. Better models were chosen according to Akaike's information criteria (AIC) and validated by leave-one-out cross-validation. Model fit was measured by the determination coefficient (R²) and the root-mean-square error (RMSE).

Lagged NDVI was excluded because of a correlation with lagged temperature and a negative correlation with incidence. Entomological covariates in the reduced model included the evening aggressiveness of An. gambiae, the night and morning aggressiveness of An. funestus and the exophagy percentage of An. gambiae. The model with entomological covariates (R²=0.619 and RMSE=11.607 malaria cases per week) had a better fit than the model without (R²=0.406 and RMSE=14.654 malaria cases per week).

Entomological surveillance data may improve the prediction of malaria incidence, even in a context of large scale vector-control and when mosquitos are only captured every two months and from one town district. More frequent captures may generate a better predictive model. However, field Anopheles surveillance remains time-and-money consuming. More studies in various transmission contexts are needed to confirm these results and assess benefits in terms of malaria control and prevention.