Forecasting acute childhood malnutrition in Kenya using machine learning and diverse sets of indicators

Abstract

Objectives: Malnutrition is a leading cause of morbidity and mortality for children under-5 globally. Low- and middle-income countries, such as Kenya, bear the greatest burden of malnutrition. The Kenyan government has been collecting clinical indicators, including on malnutrition, using District Health Information Software-2 (DHIS2) for over a decade. We aim to address the existing gap in decision-makers’ ability to develop and utilize malnu trition forecasting capabilities for timely interventions. Specifically, our objectives include: develop a spatio-temporal machine learning model to forecast acute malnutrition among children in Kenya using DHIS2 data, enhance forecasting capability by integrating exter nal complementary indicators, such as publicly available satellite imagery-driven signals, and forecast acute malnutrition at various stages and time horizons, including moderate, severe, and aggregated cases. Methods: We propose a framework to forecast malnutrition risk for each sub-county in Kenya based on clinical indicators and remote sensory data. To achieve this, we first aggregate clinical indicators and remotely sensed satellite data, specifically gross pri mary productivity measurements, to the sub-county level. We then label the rate of chil dren diagnosed with acute malnutrition at the sub-county level using the standard Inte grated Food Security Phase Classification for Acute Malnutrition. We then apply and compare several methods for forecasting malnutrition risk in Kenya using data collected from January 2019 to February 2024. As a baseline, we used a Window Average model, which captures the current practice at the Kenyan Ministry of Health. We also trained machine learning models, such as Logistic Regression and Gradient Boosting, to forecast acute malnutrition risk based on observed indicators from prior months. Different metrics, mainly Area Under Receiver Operating Characteristic Curve (AUC), were used to evaluate the forecasting performance by comparing their forecast values to known values on a hold-out test set. Results: We found that machine learning based models consistently outperform the Win dow Average baselines on forecasting sub-county malnutrition rates in Kenya. For exam ple, the Gradient Boosting model achieves a mean AUC of 0.86 when forecasting with a 6-month time horizon, compared to an AUC of 0.73 achieved by the Window Average model. The Window Average method particularly fails to correctly forecast malnutrition in parts of West and Central Kenya where the acute malnutrition rate is variable over time and typically less than 15%. We further found that machine learning models with satellite based features alone also outperform Window Averaging baselines, while not needing clinical data at inference time. Finally, we found that recently observed outcomes and the remotely sensed data are key indicators. Our results demonstrate the ability of machine learning models to accurately forecast malnutrition in Kenya at a sub-county level from a variety of indicators. Conclusions: To the best of the authors’ knowledge, this work is the first to use clini cal indicators collected via DHIS2 to forecast acute malnutrition in childhood at the sub county level in Kenya. This work represents a foundational step in developing a broader childhood malnutrition forecasting framework, capable of monitoring malnutrition trends and identifying impending malnutrition peaks across more than 80 low- and middleincome countries collecting similar DHIS2 datasets.