Title : Machine learning for predicting malaria by accounting for human and mosquito movement
Abstract:
Background: Malaria remains a major public health challenge across sub-Saharan Africa, where transmission is shaped by complex interactions between environmental conditions, human settlement patterns, and vector ecology. Existing malaria risk-mapping studies typically characterise exposure using environmental conditions at individuals’ home locations, implicitly assuming static exposure. However, people routinely move within their local environments, where they encounter varying ecological and socioeconomic conditions that may influence malaria risk. Ignoring this local movement can therefore lead to mischaracterisation of exposure and reduced predictive accuracy.
Aims: This study aimed to develop and evaluate machine learning models that explicitly account for local environmental heterogeneity surrounding malaria survey locations.
Methods: Using georeferenced malaria prevalence data and high-resolution environmental and socioeconomic covariates across six African countries, we compared three exposure representations: (i) conventional point-based raster extraction, (ii) pixel-level environmental data within10 km buffers, and (iii) buffer-level summary statistics capturing both mean conditions and local variability. Elastic Net regression and Random Forest models were trained and evaluated using cross-validation, with predictive performance assessed using RMSE and R2.
Results: We find that models incorporating local environmental context consistently outperform baseline point-based approaches. In particular, Random Forest models using buffer-level summary statistics achieve the highest predictive accuracy. Notably they improve the identification of high-risk areas while remaining computationally feasible. Feature attribution analysis highlights precipitation, temperature, temporal trends, and settlement related indicators as key drivers of malaria risk.
Conclusions: These results demonstrate that incorporating local movement-relevant exposure information can meaningfully improve malaria risk prediction and mapping. The results show that malaria modelling frameworks can benefit from moving beyond static home-location assumptions and incorporating neighbourhood-based exposure. Using real-world human mobility data and movement models may improve risk prediction and guide more targeted malaria control efforts.
