Increasing the resolution of malaria early warning systems for use by local health actors

Background: The increasing availability of electronic health system data and remotely-sensed environmental variables has led to the emergence of statistical models capable of producing malaria forecasts. Many of these models have been operationalized into malaria early warning systems (MEWSs), which provide predictions of malaria dynamics several months in advance at national and regional levels. However, MEWSs do not generally produce predictions at the village-level, the operational scale of community health systems and the first point of contact for the majority of rural populations in malaria-endemic countries. Methods: We developed a hyper-local MEWS for use within a health-system strengthening intervention in rural Madagascar. It combined bias-corrected, village-level case notification data with remotely sensed environmental variables at spatial scales as fine as a 10m resolution. A spatio-temporal hierarchical generalized linear regression model was trained on monthly malaria case data from 195 communities from 2017-2020 and evaluated via cross-validation. The model was then integrated into an automated workflow with environmental data updated monthly to create a continuously updating MEWS capable of predicting malaria cases up to three months in advance at the village-level. Predictions were transformed into indicators relevant to health system actors by estimating the quantities of medical supplies required at each health clinic and the number of cases remaining untreated at the community level. Results: The statistical model was able to accurately reproduce village-level case data, performing nearly five times as well as a null model during cross-validation. The dynamic environmental variables, particularly those associated with standing water and rice field dynamics, were strongly associated with malaria incidence, allowing the model to accurately predict future incidence rates. When compared to historical stock data, the MEWS predicted stock requirements within 50 units of reported stock requirements 68% of the time. Conclusion: We demonstrate the feasibility of developing an automatic, hyper-local MEWS leveraging remotely-sensed environmental data at fine spatial scales. As health system data become increasingly digitized, this method can be easily applied to other regions and be updated with near real-time health data to further increase performance.