Developing a Predictive Geospatial Model for Malaria Hotspot Identification in Nigeria Using Machine Learning

High-resolution predictive spatial modelling of malaria transmission hotspots can fundamentally transform resource allocation efficiency in large endemic countries like Nigeria, where heterogeneous transmission requires sub-national targeting rather than uniform national strategies. This doctoral study developed and validated a machine learning-based geospatial predictive model for malaria hotspot identification across Nigeria. A national dataset was constructed integrating malaria case data from 2,480 health facility sentinel sites, satellite-derived environmental covariates (land surface temperature, normalised difference vegetation index, rainfall, and elevation), intervention coverage indicators, and socioeconomic indices at ward level. Random forest, gradient boosting, and deep learning convolutional neural network models were trained on 70 percent of the dataset and validated on a 30 percent hold-out sample. Model performance was assessed using area under the receiver operating characteristic curve (AUC-ROC), spatial accuracy, and external validation against independent prevalence survey data from 2023 to 2024. The gradient boosting model demonstrated the highest predictive performance (AUC-ROC = 0.91). Rainfall seasonality, proximity to water bodies, insecticide-treated net coverage gaps, and under-five population density were the most important predictive features. The validated model correctly identified 87.4 percent of known hotspot clusters in external validation data. The resulting national hotspot prediction map is freely available to the National Malaria Elimination Programme and State Ministry of Health planners. Keywords: malaria hotspots, machine learning, geospatial modelling, Nigeria, predictive epidemiology.