Predicting Under-Five Child Stunting Using Machine Learning and Probability Calibration: A Nationally Representative 2024/25 EDHS Study from Ethiopia

In many parts of the world, a child's height tells a story that goes far beyond their physical size. When a young child is significantly shorter than expected for their age, it is often a sign of chronic malnutrition, a condition known as stunting. This is not merely a matter of being small; it is a marker of long-term deprivation that can affect a child's brain development, their ability to learn, and their health for the rest of their life. For decades, public health workers have tried to identify which children are at risk before this damage becomes permanent. Traditionally, they have relied on statistical tools that look for straight-line relationships between factors like a mother's education or a family's income and a child's growth. However, the reality of human health is rarely a straight line. It is a complex web where a mother's weight, the cleanliness of the water a family drinks, and the age of the child interact in unpredictable ways. Understanding these tangled connections is essential if communities want to stop stunting before it starts.

A team of researchers in Ethiopia has taken a fresh approach to this challenge, moving beyond traditional statistics to use a type of computer learning known as machine learning. They analyzed data from over ten thousand children across the country, gathered during a major national health survey. Instead of forcing the data into simple, rigid formulas, they trained six different computer algorithms to find the hidden patterns that predict which children are likely to become stunted. These algorithms acted like different types of detectives, each looking for clues in a unique way. Some looked for simple rules, while others built complex decision trees that could weigh dozens of factors at once. The researchers were careful to ensure the computers only used information available before a child's growth was measured, such as the mother's health and the family's living conditions, to avoid using future knowledge to predict the past.

The results showed that the more complex computer models were far better at spotting at-risk children than the traditional methods. The most successful tool was a sophisticated algorithm called Gradient Boosting, which learned by making small mistakes and correcting them over and over again. However, the researchers noticed that this powerful tool had a flaw: it was often too confident in its predictions, assigning extreme risk scores that did not match reality. To fix this, they applied a calibration process, a method that gently adjusted the computer's confidence levels so that a predicted risk of eighty percent actually meant an eighty percent chance of stunting. Once this adjustment was made, the model became a highly reliable guide. It correctly identified more than eighty-three percent of the children who were stunted while also correctly ruling out more than eighty-six percent of those who were not.

The study revealed that the risk of stunting is not spread evenly but follows clear patterns. The computer learned that the risk increases significantly as a child gets older, particularly between the ages of one and five years. It also found that children born to mothers who are underweight, those living in the poorest households, and those without access to clean drinking water face the highest dangers. Recent illnesses, such as diarrhea, also served as a strong warning sign. Crucially, the researchers tested whether using this computer model would actually help health workers in the field. They found that by using the model to decide which families to visit first, health workers could save resources and help more children than if they simply visited everyone or no one at all.

This work suggests a new way forward for public health in Ethiopia and similar regions. The country already employs thousands of community health workers who visit homes regularly. The researchers propose that these workers could use a simple mobile application powered by this calibrated computer model. During a routine visit, a worker could enter a few basic facts about the family and the child, and the app would instantly generate a risk score. This would allow the health worker to focus their time and resources on the families most in need, offering nutritional advice or support long before a child's growth stops. By turning complex data into a clear, actionable signal, this approach offers a practical way to intervene early, potentially preventing the irreversible effects of stunting and giving more children the chance to grow up healthy and strong.