Advanced Spatio-Temporal Neural Networks for Malaria Prevalence Forecasting in Sub-Saharan Africa

Accurate malaria prevalence forecasting is critical for timely interventions in Sub-Saharan Africa, where the disease remains a major health burden. Traditional regression models often struggle to capture the complex interactions between the epidemiological, environmental, and spatial factors that drive malaria transmission. This study develops and compares two distinct deep learning approaches for predicting the Plasmodium falciparum parasite rate in children (PfPR ₂₋₁₀ ), using historical prevalence data (1900–2015) and climate variables—temperature and precipitation—sourced from NASA’s POWER project. The first framework is a Spatiotemporal Transformer developed to capture intricate, local dependencies between covariates. This is contrasted with a second approach based on the Patch-based Time Series Transformer (PatchTST), which includes a standard implementation for point forecasting. Both models were further extended with uncertainty estimation using Monte Carlo Dropout, providing probabilistic forecasts in addition to point predictions. Our findings highlight a key architectural trade-off: for this dataset, characterized by short time-series and strong cross-feature interactions, the Spatiotemporal Transformer yielded higher predictive accuracy. Conversely, while demonstrating lower point-forecast accuracy in this context, both models successfully provided valuable uncertainty estimates for their predictions. This comparative analysis provides a nuanced perspective, offering decision-makers a choice between a highly accurate model for direct forecasting and probabilistic models for risk assessment and strategic planning.