One of three ‘research accelerator projects‘ starting in late 2024, the University of California-led project will improve early warning of food insecurity by integrating child malnutrition forecasts based on machine learning and anthropometric data into Kenya’s National Drought Monitoring Authority services.
The challenge
Accurate forecasting of sub-national anthropometric-based food insecurity has so far proven an elusive task. Early warning (EW) systems that alert of hunger crises among vulnerable population subgroups have therefore traditionally relied on heuristic evaluation of historical data.
Kenya’s National Drought Monitoring Authority (NDMA) is a governmental organization that monitors and coordinates relief action for drought-driven hunger crises in the country’s arid and semiarid regions. Part of their monitoring activity entails the monthly collection of anthropometric data (Middle Upper Arm Circumference) from children under five, from which the prevalence of global acute malnutrition (GAM) can be derived.
Despite these monitoring efforts, the NDMA does not currently employ any GAM prevalence forecasting system, which limits their ability to trigger timely and well-dimensioned humanitarian response to avoidable acute malnutrition and death.
Our recent work has established that these high frequency anthropometric data, combined with secondary remote sensing data and machine learning methods, can forecast GAM with sufficient accuracy to be operationalized.
In this project, we will co-develop and launch an early warning tool that can be incorporated into NDMA’s operations and adapted to other settings with similar hunger crises problematic where sentinel site data exist.
Outcomes
The focus of this project will be to help NDMA and its partners leverage improved forecasts of GAM prevalence. In close collaboration with the agency, we will develop an operationally useful display of our model’s results that will be relevant to NDMA decision making. We expect that this new addition will increase trust in the system’s capability of detecting impending hunger events, aiding with the timely triggering of humanitarian response.
The Kenya Drought Early Warning System (KDEWS) is currently a systematic collection of data that aims to facilitate the timely detection and monitoring of drought indicators that define the onset, continuation, termination, and effects of drought conditions. The EWS is an integral part of drought planning and preparedness, whose aim is to mitigate and respond to drought crises.
The Kenya Food Security Meeting (KFSM) is the main coordinating body that brings together food security actors in a forum where information is exchanged, options debated and decisions on activities formulated for referral to the Government of Kenya and donors. Bulletins produced from the analysis of early warning data are presented to the during the monthly meetings and shared with various agencies. It is an open forum of high-level representation of a broad grouping of organizations at the national level. An important sub-component of the KFSM is the Kenya Food Security Steering Group (KFSSG), which is the principal food security and drought management institution in the country. The members of KFSSG include the Office of the President (Ministry of State for Special Programmes), Ministry of State for Development ASALS, National Drought Management Authority (NDMA), Ministry of Agriculture, Ministry of Medical Services, Ministry of Public Health and Sanitation, UNICEF, WFP, DFID, USAID, and OXFAM.
Despite this current data gathering, processing and dissemination efforts, Kenya’s drought EWS does not currently include any routinely forecasting method to inform planning for humanitarian response in the event of an impending food security emergency. All current needs assessments are based on the analysis of historic data. The incorporation of our model to the KDEWS will be therefore one of the first examples of the use of an anthropometric-based, sub-population specific forecasting method for food security EW in empirical settings.
Activities and outputs
We have been prototyping a prediction model that produces GAM forecasts at 1-, 3-, and 6-months in advance, using supervised machine learning methods. The model is trained on sliding windows of 36 months of data and includes an autoregressive component —3-month lags of the wards’ wasting prevalence — and a set of secondary geospatial variables hypothesized to be underlying predictors of U5 children’s nutritional status in this setting (e.g., proxies for land productivity, weather variables, price information, malaria incidence, and conflict). We determine out-of-sample performance in identification accuracy of impending hunger crises via sensitivity, precision, and specificity.
We follow the WHO’s “wasting prevalence for public health significance” cutoffs and define a wasting prevalence over 15% as the threshold to determine whether a ward is experiencing a malnutrition crisis in a particular month. This threshold, or the proximity to it, will be the one triggering a humanitarian response in real-world settings. Based on that value, we calculate the sensitivity, precision, and specificity of our model for the generated predictions in each month of interest.
The predictions coming from this model will be incorporated into a co-developed open-source dashboard that can be routinely accessed in the monthly KFSSG meetings. We will also carry out technology transfer of the model itself, so that the NDMA can manage the production of these results moving forward.
In situ meetings will be carried out with the KFSSG during the first and second quarters of 2025, with the research team traveling to Kenya to gather the necessary feedback to define the final design of the EW tool and carry out the necessary training for its implementation. We will co-develop a set of open-source tools and dashboards that make these prediction algorithms accessible and actionable to governments and humanitarian organizations at large.
In partnership with NDMA, we will also develop plans for an impact evaluation to rigorously document how improved forecasts impact NDMA’s decision making processes. While the actual impact evaluation will require additional resources beyond the scope of this project, the meetings in Kenya will be instrumental in fleshing out the design of this evaluation.
Pathways to impact
Through this project, we hope to set an example of collaboration between academia and international development stakeholders to ultimately create widespread adoption of modern forecasting early warning (EW) systems. In turn, we believe that empowering government and humanitarian organizations can have a material impact on the lives of tens of millions of people who are vulnerable to climate shocks and food insecurity. We will measure impact by (i) adoption and use of the tools; (ii) changes in decision-making processes and interventions by implementing organizations; and (iii) downstream impacts on household welfare.
The research team will develop the impact evaluation design based on a rollout implementation of the tool. In partnership with the NDMA and under their approval, we implement the analysis. The results of this impact evaluation will serve as a rigorous measure of impact on the use of the co-developed EW forecasting tool on the reduction of hunger prevalence among U5 children. The research team will disseminate the results with the NDMA and present the study for peer reviewed publication. With this exercise, we ultimately aim to raise awareness and adoption of anthropometric-based, sub-population specific EW forecasting systems as an effective way to combat the detrimental effects of severe food insecurity among vulnerable population subgroups with high-rate case fatality.
Partners
More
Download a short presentation of the project
Contact: Susana Constenla-Villoslada, University of California at Berkeley (susana_constenla@berkeley.edu)
The research described here is funded by the Jameel Observatory for Food Security Early Action. The original work to develop the underlying predictive model was funded by the United States Agency for International Development (USAID) under cooperative agreement number 7200AA18CA00014, “Innovations in Feed the Future Monitoring and Evaluation – Harnessing Big Data and Machine Learning to Feed the Future”. This work was also undertaken as part of the Food Security Portal project led by the International Food Policy Research Institute (IFPRI), with funding support provided by the European Commission (EC).
Image: A child’s arm is measured at a malnutrition clinic in Lekwasimyen in Northern Kenya’s Turkana province. Photo: Lisa Murray / Concern Worldwide.
