High-frequency monitoring enables machine learning-based forecasting of acute child malnutrition for early warning.

The number of acutely food insecure people worldwide has doubled since 2017, increasing demand for early warning systems (EWS) that can predict food emergencies. Advances in computational methods, and the growing availability of near-real time remote sensing data, suggest that big data approaches might help meet this need. But such models have thus far exhibited low predictive skill with respect to subpopulation-level acute malnutrition indicators. We explore whether updating training data with high frequency monitoring of the predictand can help improve machine learning models' predictive performance with respect to child acute malnutrition by directly learning the dynamic determinants of rapidly evolving acute malnutrition crises. We combine supervised machine learning methods and remotely sensed feature sets with time series child anthropometric data from EWS' sentinel sites to generate accurate forecasts of acute malnutrition at operationally meaningful time horizons. These advances can enhance intertemporal and geographic targeting of humanitarian response to impending food emergencies that otherwise have unacceptably high case fatality rates.