Interpretable spatial machine learning framework for evaluating malaria intervention effectiveness in Northern Uganda

Abstract

Malaria prevalence among children under five in Northern Uganda remained high in 2024-25 (37.4%), similar to 2018-19 (35.6%), despite a 13.2 percentage point rise in insecticide-treated net (ITN) coverage. This study examined why prevalence stagnated while coverage increased. It used Malaria Indicator Survey (MIS) data on 3,627 children in 110 clusters in Northern Uganda. The methods were wave-interaction logistic regression, Random Forest classification with SHAP value decomposition, and Local Indicators of Spatial Association (LISA). The association between ITN use and malaria reversed between waves (OR: 0.72 to 1.56; interaction OR = 2.38, p< 0.001). This reversal is better explained by confounding by indication, pyrethroid resistance and net degradation than by a real loss of protective effect. There was no significant wave interaction for indoor residual spraying (IRS) (OR = 1.03, p = 0.903). SHAP analysis identified a non-linear coverage threshold: the largest protective differences were associated with ITN coverage of 12% to 15%, below which population-level protection did not occur. Older children within the under-five group were the most affected, with the odds of malaria rising with each month of age in both waves. Prevalence differed sharply between sub-regions, from 14.9% in West Nile to 56.8% in Karamoja, and malaria was spatially clustered in both waves (Moran’s I = 0.28 and 0.31, p = 0.001). Hotspots persisted even though their ITN coverage was higher than in cold spots, showing that coverage alone has not reduced the burden where it is greatest. Successful malaria control in Northern Uganda requires a shift from aggregate coverage targets to spatially stratified, hotspot-based ITN distribution. PBO net scale-up should be accelerated in pyrethroid-resistant areas. IRS structural coverage should reach at least 85 percent to achieve community-level impact. Spatial surveillance should be built into national health information systems. These measures should be implemented as a package rather than as separate vertical programmes, because coverage gains are not enough when resistance and degradation reduce net efficacy. NMCP distribution should be coordinated using evidence from repeated MIS waves, so that programmatic decisions are based on sub-regional burden data.