Predicting Cervical Cancer Screening Utilization Among Women of Reproductive Age in Ethiopia Using Supervised Machine Learning: A Nationally Representative 2024/25 EDHS Study

Cervical cancer is a disease that can be stopped before it ever becomes fatal, yet it remains a leading cause of death for women in many parts of the world. The solution is simple in concept but difficult in practice: regular screening allows doctors to find and treat early signs of the disease long before symptoms appear. In wealthy nations, organized programs have made this screening routine, drastically reducing deaths. However, in countries like Ethiopia, where resources are stretched and communities are vast, getting women to these screening services is a massive challenge. Current efforts often rely on broad, one-size-fits-all approaches that miss the specific women who need help the most. To solve this, researchers are turning to a different kind of tool: computer models that learn from data to predict human behavior. These models do not guess; they analyze thousands of past records to find hidden patterns in what makes a woman likely to seek care or stay away from it.

A team of researchers in Ethiopia recently applied this approach to a critical question: who among the millions of women of reproductive age in the country is most likely to skip cervical cancer screening, and why? Using the most recent national health survey, which gathered information from over 15,000 women across the entire nation, they built a sophisticated computer system designed to learn the complex reasons behind screening decisions. Instead of relying on simple rules, the team trained six different types of learning algorithms, teaching them to recognize the subtle combinations of factors—such as education, wealth, and access to information—that drive a woman to get screened. The goal was not just to understand the past, but to create a reliable tool that could help health workers in the future identify the women who have never been screened and target them with the right support.

The study began with a massive dataset representing the Ethiopian population, capturing details on everything from a woman's age and where she lives to her knowledge of the disease and her ability to make her own health decisions. The researchers found that the national screening rate was alarmingly low, with only about 4 percent of women having ever been screened. This meant that for every 100 women in the study, 96 had never received this life-saving checkup. The data revealed stark divides: women in cities were far more likely to be screened than those in rural villages, and women with higher education or more wealth were much more likely to access services than those without. Crucially, the researchers discovered that knowing about cervical cancer was the single strongest predictor of whether a woman would get screened, yet the vast majority of women in the study had never even heard of the disease.

To make sense of these complex patterns, the researchers tested several different computer learning methods. They found that a specific type of advanced algorithm, known as extreme gradient boosting, was far superior to the traditional statistical methods usually used in public health. While standard methods could identify general trends, this advanced model could see the intricate web of factors that influence a woman's decision. For instance, it learned that education alone was not enough; the benefit of education was multiplied when a woman also lived in an urban area. Similarly, knowing about the disease only led to screening when a woman also had the power to make her own health decisions without needing permission from a partner. The model was so accurate that it could distinguish between women who had been screened and those who had not with a level of precision that far exceeded previous attempts.

One of the most important steps the team took was ensuring the computer's predictions were honest and reliable. Complex computer models often become overconfident, predicting probabilities that are too high or too low. The researchers used a technique called calibration to adjust the model's output, ensuring that if the model predicted a 10 percent chance of a woman being screened, that probability matched reality. After this adjustment, the model became a trustworthy guide. It showed that by focusing on specific, non-invasive questions—like whether a woman has heard of the disease, if she has visited a clinic recently, and how much say she has in her own healthcare—health workers could accurately identify women who were at high risk of never getting screened.

The study also used a method to explain exactly how the computer reached its conclusions, breaking down the "black box" of the algorithm to show the human logic behind the numbers. This analysis confirmed that the most powerful drivers for screening were awareness of the disease, having a secondary or higher education, and the freedom to make one's own health choices. It also highlighted that recent contact with the health system, such as visiting a clinic for a child's vaccination or prenatal care, was a strong signal that a woman was ready to be screened for cervical cancer. These findings suggest that the barriers to screening are not just about distance or cost, but deeply rooted in knowledge, autonomy, and the structure of daily life.

The researchers propose that this calibrated computer model could be integrated directly into the mobile tablets used by Ethiopia's thousands of community health workers. As these workers visit homes, they could input simple answers to a few questions, and the system would instantly flag women who have never been screened and explain why. Instead of guessing which households to prioritize, health workers could focus their limited time and resources on the women who need them most, offering tailored advice and arranging referrals. This approach moves public health from a broad, scattered effort to a precise, targeted strategy. By using data to understand the specific barriers facing each woman, the country can begin to close the gap between the women who are protected and those who remain vulnerable, turning a complex statistical model into a practical tool for saving lives.