A Comparative Machine Learning Analysis of Neural Networks and XGBoost for Predicting Maternal Mortality in Maiduguri, Nigeria

In many parts of the world, the journey of pregnancy is fraught with danger, where a single complication can turn a moment of anticipation into a tragedy. In low-resource settings, and particularly in areas affected by conflict, the ability to predict who is at risk of dying during childbirth is often limited by a lack of data and the sheer complexity of human biology. For decades, doctors have relied on standard statistical methods to weigh factors like a mother's age, her history of previous births, and her access to care. These methods are useful, but they often struggle to see the intricate, non-linear ways these factors interact. In recent years, a new generation of tools has emerged from the field of computer science, capable of finding hidden patterns in vast amounts of information. Two of the most powerful of these tools are artificial neural networks, which mimic the way the human brain connects ideas, and gradient boosting machines, which build predictions by learning from a series of small, sequential mistakes. The question facing public health officials is whether these advanced digital tools can actually help save lives in the most difficult environments, or if they are simply too complex for the data available.

In the conflict-affected city of Maiduguri, Nigeria, where insecurity and a fragile health system make maternal care a daily struggle, two researchers set out to test these digital tools against the reality of local medical records. They gathered information from 158 women who had received care at selected hospitals in the city. The goal was straightforward: to see if a computer could look at a woman's health history and correctly predict whether she would survive childbirth or not. The team compared two different approaches. One approach used an artificial neural network, a type of program that learns by adjusting connections between layers of digital units, much like a student refining their understanding of a subject through repeated practice. The other approach used a method called XGBoost, which works by building a forest of decision trees, where each new tree corrects the errors of the ones before it, gradually honing in on the most accurate prediction.

The researchers began with a dataset that included details such as the mother's age, education, income, the number of times she had been pregnant, and whether she had experienced complications like severe bleeding or high blood pressure. However, as they prepared the data for the computer to learn from, they discovered a critical flaw that would have completely misled their results. One variable in their records, labeled "Severe Morbidity," turned out to be a perfect mirror of the outcome itself. In every single case where a woman died, this variable was marked as severe, and in every case where she survived, it was marked as not severe. If the computer had been allowed to use this information, it would not have been predicting the future; it would simply have been reading the answer key. The researchers made the difficult but necessary decision to remove this variable entirely. They also removed another piece of data that was the same for every single person in the study, as it offered no useful information for distinguishing between survivors and those who did not survive.

With the data cleaned and the misleading information removed, the team ran their models using a rigorous testing method. Instead of splitting the data once into a training set and a test set, they repeatedly shuffled the 158 records into different groups, training the models on some and testing them on others, a process they repeated one hundred times to ensure the results were stable and not just a lucky break. The results were striking. Both the artificial neural network and the XGBoost model performed with exceptional accuracy, correctly classifying the vast majority of cases. The neural network was slightly more accurate overall, getting the right answer in nearly every instance, while the XGBoost model was also highly effective, missing only a few cases. Both of these advanced tools significantly outperformed a traditional statistical method, which, while still good, was less precise.

Perhaps more important than the raw scores was what the computers told the researchers about the causes of death. When the misleading variable was removed, the models identified a new set of factors that drove the risk of maternal mortality. The most powerful predictor was eclampsia, a dangerous condition involving seizures during pregnancy. Following that were the method used to deliver the baby, the number of previous births, the total number of pregnancies, and the household income of the mother. These findings aligned perfectly with what medical experts already knew about the dangers of pregnancy, giving the researchers confidence that their models were learning real medical truths rather than just memorizing data quirks. The models successfully captured the complex ways these factors worked together, showing that a woman's risk is not just about one thing, but a combination of her medical history, her delivery circumstances, and her social circumstances.

Despite these impressive results, the researchers were careful not to declare the problem solved. They noted that the dataset was quite small, containing only 158 records, and that the high accuracy might be partly because the computer had learned the specific patterns of this small group of women too well. It is possible that if the models were tested on a completely different group of women from another hospital or a different city, their performance might drop. The study serves as a strong proof of concept, showing that these digital tools can find meaningful patterns in local data, but it is not yet a final solution ready for immediate use in every clinic. The researchers emphasized that before these tools can be trusted to guide life-or-death decisions in a hospital, they must be tested on much larger groups of people and validated in different settings.

The study concludes that while artificial intelligence holds great promise for helping doctors in resource-limited and conflict-affected areas identify high-risk pregnancies early, it must be used with caution and humility. The technology is powerful, but it is only as good as the data it is fed and the care taken to ensure that data is honest and complete. For the women of Maiduguri, the path forward involves not just adopting new technology, but also improving how medical records are kept and ensuring that the information recorded is truly independent of the outcome. With better data and further testing, tools like neural networks and gradient boosting could become vital partners for healthcare workers, helping them allocate scarce resources to the mothers who need them most, potentially saving lives in some of the world's most challenging environments.