Millions of children suffer from food insecurity every year, but a new predictive tool could help governments proactively prevent spikes in malnutrition.
“At a time when humanitarian budgets have been drastically cut around the world, it is more important than ever to direct scarce resources to the places and people where they are needed most and have the greatest rewards,” said Chris Barrett, the Stephen B. and Janice G. Ashley Professor of Applied Economics and Management at the Charles H. Dyson School of Applied Economics and Management.
Researchers at the Cornell S.C. Johnson School of Business and the University of California, Berkeley School of Information are piloting a machine learning system that can predict malnutrition hotspots up to three months before they appear, helping humanitarian organizations and government agencies send preventive aid to people in need. Susana Constenla-Villoslada, M.A. ’19, presented the system on May 12 at the Data Science Thinking Summit to Build Resilience and Improve Humanitarian Response.
“Childhood wasting is a clinical condition with a 45% mortality rate if untreated,” said Dr. Constenla Viroslada. candidate at the University of California, Berkeley, where he studies econometrics and machine learning for international development. “The children who don’t die can have long-term developmental consequences throughout their lives. If we don’t do anything about this problem, we’re putting the next generation at risk, reinforcing the cycle of poverty.”
The research stems from a long-term collaboration with the Kenya National Drought Management Authority (NDMA), the government agency responsible for drought monitoring and response in Kenya’s arid regions. The model was built using NDMA’s consistent drought monitoring data and based on NDMA’s operational needs and local expertise.
“The people who are actually moving the needle are the people on the ground implementing these interventions,” Constenla-Viroslada said. “We support people’s collective efforts to address these issues.”
The research team’s strongest predictive model is derived from multiple public data streams and uses both environmental and socio-economic forecasts to predict the prevalence of malnourished children in a given community. The model primarily uses satellite imagery to generate proxies for rainfall and vegetation health abnormalities that are predictors of the prevalence of child malnutrition.
“Public data streams like these satellite imagery products are very useful to us because of their high frequency,” said Constenla-Villoslada. “We need to protect that public good because we can use them as predictors of important variables such as child wastage.”
The model also utilizes publicly available conflict data (deaths and incidences) and NDMA data on monthly child measurements to generate malnutrition prevalence results.
Many weather and financial forecasting systems trigger alerts based on a single event, but human results are more complex and result from the interplay of multiple factors. The research team found that machine learning expands predictive power by deriving complex relationships from multiple variables.
“Although this is a multi-problem problem, our results show that conflict and drought are the main causes of the rapid rise in child malnutrition,” Constenla Viroslada said. “These spikes take a lot of time to return to normal values, so it’s important to address hotspots before they occur.”
The research team backtested the model using historical data to simulate current predictions and see if the predictions matched what actually happened. One month before a crisis, about half of alerts were in response to an actual crisis, and six months before a crisis, one-third of alerts were in response to an actual crisis.
The accuracy of the model was highly dependent on both the frequency and continuity of the data. From 2006 to 2009, when the data collection sites were stable, the model was able to detect up to 77% of impending hotspots. However, the model’s sensitivity declined sharply over the next two years as Kenya relocated many of its monitoring sites in 2016.
The model also performed well in real-world scenarios. Direct reports from unaffiliated organizations confirmed these situations when their system predicted an increase in malnutrition in Gorbo District, Marsabit County.
The research team has transferred this technology to partner organizations in Kenya and is currently exploring how best to utilize the data.
“The NDMA is using a predictive model retrained based on the latest version of data we have published to begin anticipatory cash transfers to mothers in wards where child malnutrition is on the rise,” Barrett said. “We are working with them to rigorously evaluate their programs.”
The team’s next goal is to develop a large-scale language model within predictive dashboards to explain the impact of the data to audiences with varying levels of technical understanding.
“There are an estimated 10 million malnourished children in this large African country. This is unacceptable at this time,” Constenla-Viroslada said. “The idea is to empower these governments to strengthen democracy in these regions so that the system can more efficiently take care of their citizens.”
The remaining members of the research team were NDMA’s Nelson Mutanda, NDMA’s Clinton Ouma, International Food Policy Research Institute senior fellow Yanyang Liu, and Dr. Linden McBride. ’18.
