Authors - Lydie Simone Tapsoba, Salifou Ouoba, Athanase Sawadogo Abstract - Malaria is the leading cause of child mortality in Burkina Faso and accounts for over 40% of public health expenditure. This study develops and validates a hybrid machine learning model combining Random Forest (40%) and XGBoost (60%) to predict the spatio-temporal dynamics of malaria across the country’s 13 health regions over the period 2010–2024. The model incorporates satellite-derived climatic variables (CHIRPS precipitation, MODIS NDVI, temperature) and demographic variables, enhanced by advanced feature engineering: 3- and 6-month moving averages of precipitation, temporal lags and circular encoding of seasonality. Validation combines a prospective temporal split for 2024 and spatial GroupKFold cross-validation. On the independent 2024 test set, the hybrid model achieves exceptional performance, substantially outperforming the best previously published approaches for this context. Interpretability analysis reveals surprising determinants of transmission, highlighting the equal role of environmental and anthropogenic factors in Burkina Faso. The SHAP analysis identifies the 3-month moving average of rainfall as the dominant variable, ahead of population density, highlighting the equal role of environmental and anthropogenic factors. The priority areas are Hauts-Bassins, the East and the South-West. The risk maps and 6-month predictions serve as operational tools for the National Malaria Control Programme.
Open Zoom