Augmented analytics: statistical models that users can question in plain language
Augmented analytics pairs rigorous statistical models with an AI assistant, so that decision makers, researchers and programme staff can question the results directly. Our first application, LUMINA-AI, shows where and in which season children are most at risk of acute malnutrition, and why.
- First applied in South Sudan, with FAO's Office of Emergencies and Resilience (OER) and the Land and Water Division (NSL)
- Estimates by county and on a finer grid, with seasonal peaks and their drivers
- An AI assistant that answers with model results and scientific evidence
The Data Lab is seeking partners to apply this approach to other countries and outcomes. Contact us.
The challenge
Many outcomes that matter for food security and nutrition have interacting causes spread across sectors. Statistical models can combine them, but turning their outputs into operational advice often depends on a few experts. Acute malnutrition is a clear example: food insecurity, prices, conflict, climate shocks and disease all play a part.
How it works
The statistical model used in LUMINA-AI links data from several levels, such as individuals, households and areas, with survey, market, conflict and Earth observation sources. Bayesian methods estimate how an outcome varies across places and months, together with the uncertainty of each estimate.
The AI assistant answers questions in plain language. It gives the statistical findings first, then evidence from a curated library of scientific literature, and adapts its explanations to each user's level of expertise. Every figure it reports comes from the models.
First application: LUMINA-AI
- Risk maps. The estimated probability of child wasting by county, and on a finer grid where survey data allow.
- Seasonal patterns. The months when risk peaks, nationally and by county, from monthly data since 2018.
- Drivers. The factors associated with wasting, from child illness and feeding practices to household food insecurity, prices, conflict, displacement, rainfall, vegetation and floods.
Knowing where and when risk peaks lets humanitarian actors plan anticipatory action before the high-risk season.
Pilot and next steps
The South Sudan pilot is complete. It produced county- and grid-level risk maps, seasonal profiles and a driver analysis, all accessible through the assistant. The same approach can extend to other countries and to outcomes.
LUMINA-AI stands for Livelihood, Food Insecurity and Undernutrition Modelling for Anticipatory Action using AI. The pilot was funded through ELEVATE, FAO's internal innovation incubator.