Community health worker examining a young child
Healthcare

AI Predicts Child Malnutrition in Kenya Six Months Before It Hits

7 min read|Updated November 2026
Share

At a glance

Researchers at USC, working with Amref Health Africa, Kenya's Ministry of Health and the Microsoft AI for Good Lab, built a machine learning model that combines routine data from more than 17,000 Kenyan health facilities with satellite measurements of crop health to forecast acute child malnutrition at sub county level up to six months in advance. It reaches 89 percent accuracy one month out and 86 percent accuracy over six months.

  • About 350,000 children under five in Kenya, roughly 5 percent, suffer from acute malnutrition, and rates reach 25 percent in some regions
  • The model forecasts malnutrition up to six months ahead, giving health agencies time to move food and supplies before a crisis peaks
  • It combines clinical records from more than 17,000 health facilities with satellite data on crop health and productivity
  • Accuracy is 89 percent at one month and 86 percent at six months, beating baseline models that only follow historical trends
  • The underlying health data system, DHIS2, is used in more than 125 countries, so the approach can be adapted far beyond Kenya

Malnutrition weakens a child's immune system and turns common illnesses like diarrhea and malaria into life threatening events. In Kenya, 5 percent of children under five, an estimated 350,000 individuals, suffer from acute malnutrition. In some regions the rate climbs to 25 percent. Globally, undernutrition is linked to nearly half of all deaths in children under five. Until now, the response has mostly been reactive: aid arrives after children are already sick. A team of researchers has built an AI model that flips that timeline, predicting where acute malnutrition will surge up to six months before it happens.

The model is the work of a multidisciplinary team from the USC School of Advanced Computing and the Keck School of Medicine, working with experts from the Microsoft AI for Good Lab, Amref Health Africa and Kenya's Ministry of Health. Published in PLOS One, it forecasts acute childhood malnutrition across Kenya at the sub county level with enough lead time for governments and humanitarian organizations to deliver food, health care and supplies to at risk areas before the crisis peaks.

Why Forecasting Malnutrition Is Hard

Current forecasting in Kenya relies largely on expert judgment and historical knowledge. These methods work reasonably well where patterns repeat, but they struggle to anticipate new hotspots or rapid shifts driven by drought, flooding and displacement. When the rains fail in one county and not its neighbor, last year's map of need is a poor guide to this year's.

The new model takes a different approach. It draws on routine health data collected monthly through the District Health Information System 2, known as DHIS2, from more than 17,000 Kenyan health facilities. It then layers on satellite derived indicators of crop health and productivity, because harvest failure is one of the strongest signals that children will soon go hungry. By learning the relationship between these variables, the model captures complexity that simple trend following misses.

Accuracy Where It Matters Most

The results are striking. The model achieves 89 percent accuracy when forecasting one month ahead and maintains 86 percent accuracy over a six month horizon. That is a significant improvement over simpler baseline models that rely only on recent historical malnutrition prevalence trends.

Crucially, the model is especially effective in the hardest places to plan for: regions where prevalence fluctuates and surges are difficult to anticipate. These are exactly the communities that expert judgment tends to miss, and exactly the ones where an early warning buys the most lives saved.

"This model is a game changer," said Bistra Dilkina, associate professor of computer science at USC and co director of the USC Center for Artificial Intelligence in Society. "By using data driven AI models, you can capture more complex relationships between multiple variables that work together to help us predict malnutrition prevalence more accurately."

From Model to Dashboard to Action

A forecast only matters if someone acts on it. The team has built a prototype dashboard that visualizes regional malnutrition risk, allowing officials to filter data, zoom in on impacted locations and toggle between immediate and long term forecasting timeframes. It is designed to be usable without technical training, so a district nutrition officer can open it and understand what is coming.

Amref Health Africa, the largest health focused organization of its kind in Kenya, is working to put the tool in the hands of partners and policymakers. Amref's community health promoters already visit households across the country, teaching prevention and nutrition and referring sick children to care. The forecasting model gives that network something it never had: a map of where to go before the emergency, not after.

"Trends tell us a story," said Murage S.M. Kiongo of Kenya's Ministry of Health. "Multifaceted data sources, coupled with machine learning, offer an opportunity to improve programming on nutrition and health issues."

Built to Travel Beyond Kenya

The most important design decision may be what the model does not require. It uses no special surveys, no new data collection and no expensive equipment. It runs entirely on health records that countries already collect and satellite data that is publicly available. More than 125 countries use DHIS2, including about 80 low and middle income countries, which means the framework could be adapted to fight malnutrition across much of the world.

"If we can do this for Kenya, we can do it for other countries," Dilkina said. "The sky's the limit when there is a genuine commitment to work in partnerships."

The team is now working with the Kenyan Ministry of Health and Amref to integrate the model and dashboard into government systems and decision making, with the goal of creating a sustainable, regularly updated public resource. Girmaw Abebe Tadesse, principal scientist at the Microsoft AI for Good Lab in Nairobi, believes the dashboard will change how partners intervene, enabling evidence based and timely action instead of reaction.

Prevention Is the Cheapest Medicine

Laura Ferguson, director of research at USC's Institute on Inequalities in Global Health, calls malnutrition a public health emergency in Kenya. "Children are sick unnecessarily. Children are dying unnecessarily," she said. The cost of waiting is measured in hospitalizations, permanent stunting and lives lost. The cost of acting early is a truck of therapeutic food sent to the right county, at the right time.

That is the quiet promise of this work. Not a robot, not a diagnostic machine, but a forecast. An AI model reading ten years of clinic records alongside images of crops from space, telling health workers where hunger will strike next, months before the first child loses weight.

Common Questions

What does the Kenya malnutrition AI model do?

It forecasts acute child malnutrition at sub county level across Kenya up to six months in advance, so government and humanitarian organizations can pre position food, health care and supplies in at risk areas before children get sick.

How accurate is the malnutrition prediction model?

The model achieves 89 percent accuracy forecasting one month ahead and 86 percent accuracy over six months, a significant improvement over baseline models that rely only on recent historical malnutrition trends. It performs best where it matters most: regions where prevalence fluctuates and surges are hard to anticipate.

Who built the AI malnutrition forecasting tool?

Researchers from the USC School of Advanced Computing and Keck School of Medicine worked with the Microsoft AI for Good Lab, Amref Health Africa and Kenya's Ministry of Health. The findings were published in PLOS One in a study titled Forecasting acute childhood malnutrition in Kenya using machine learning and diverse sets of indicators.

What data does the model use?

It uses routine clinical data from more than 17,000 Kenyan health facilities collected through the District Health Information System 2, combined with satellite derived indicators of crop health and productivity. It requires no new data collection, which is why it can be adapted to other countries.

Can this approach be used outside Kenya?

Yes. The health data system it relies on, DHIS2, is used in more than 125 countries, including about 80 low and middle income countries. The researchers say the same framework, built from existing health and satellite data, could be adapted to fight malnutrition elsewhere.

Sources: USC Viterbi School of Engineering; PLOS One study Forecasting acute childhood malnutrition in Kenya using machine learning and diverse sets of indicators; Microsoft AI for Good Lab; Amref Health Africa; Kenya Ministry of Health; Microsoft Customer Stories.