A new doctoral study by Clarisse Kagoyire is offering Rwanda and other countries facing similar challenges a more precise way to identify and address childhood stunting.
Conducted as part of her PhD at Lund University under the UR-Sweden Programme, the research introduces advanced spatial modelling techniques that help pinpoint where stunting is most severe and what factors are driving it at a local level.
Childhood stunting, defined by the World Health Organization as impaired growth and development due to chronic malnutrition and adverse living conditions, remains a major public health concern. While Rwanda has made notable national progress in reducing stunting, disparities persist across districts and even within communities.
These inequalities highlight the need for more targeted, data-driven interventions.
Kagoyire’s research responds to this need by combining two rich datasets. The first comes from a detailed household survey conducted in Rwanda’s Northern Province, focusing on children aged 1 to 36 months. This dataset captures growth measurements alongside household conditions, maternal and child health, hygiene practices, and geocoded locations.
The second dataset provides nationwide coverage through geospatial estimates derived from Rwanda’s Demographic and Health Survey. Together, these datasets allow for both fine-scale local analysis and broader national insights.
Across four interconnected studies, the research moves from traditional spatial analysis to more advanced machine learning approaches. Findings from the Northern Province show that about 27% of children in the study area are stunted, with clear geographic clustering higher rates in Musanze and lower rates in Rulindo.
Crucially, the study reveals that the factors associated with stunting are not uniform across locations. Access to basic services such as handwashing facilities and electricity consistently reduces the likelihood of stunting, but their impact varies from place to place.
The research goes further to show that these relationships are often complex, operating at different scales and in non-linear ways.
This means that national averages or one-size-fits-all interventions may overlook important local realities. To address this, Kagoyire applies innovative models that “learn” local patterns directly from the data, improving the accuracy of identifying high-risk areas while maintaining interpretability for policymakers.
At the national level, the study uses graph-based deep learning to account for connections between neighboring regions, recognizing that nearby communities often share similar risks and services. It also incorporates explainable machine learning techniques to identify key drivers of stunting in different locations and to simulate “what-if” scenarios. These scenarios suggest that the most significant reductions in stunting could be achieved through combined, multisector interventions; particularly improvements in water and sanitation, maternal and child healthcare, and malaria prevention. Notably, the greatest potential gains are concentrated in areas already experiencing the highest burden.
Rather than producing a single model or map, the thesis contributes a scalable and transferable workflow that links detailed local data with national-level decision-making.
The outputs; presented as practical maps and decision-support tools; are designed to help policymakers prioritize interventions, allocate resources more effectively, and sequence integrated actions.
While the study relies on observational data and does not establish direct causation, it provides strong predictive insights that can guide policy and program design. By emphasizing place-specific and scale-sensitive analysis, Kagoyire’s work demonstrates how modern data science can move beyond describing problems to actively supporting solutions; ensuring that interventions reach the children and communities that need them most.
