Hiding in plain sight: finding undiagnosed chronic kidney disease without kidney function data through a machine learning model
This study demonstrates that a frozen, kidney-function-free machine learning model originally developed in Latin America can effectively prioritize testing for undiagnosed chronic kidney disease in a diverse US population by maintaining strong risk discrimination, though it requires recalibration to ensure accurate probability estimates across different demographic subgroups.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Millions of people around the world live with a silent threat to their health: chronic kidney disease. This condition slowly damages the organs that filter waste from the blood, yet it often produces no symptoms until it is very advanced. Because the disease is so quiet, most people who have it do not know it exists. This is a critical problem because modern medicine now offers effective, low-cost treatments that can slow or stop the damage, but these treatments only work if doctors can find the patients early. The challenge is that checking everyone's kidney function is expensive and inefficient. Doctors need a way to figure out which people in a crowded waiting room are most likely to have the hidden disease so they can prioritize who gets tested first. This is where computer models come in, acting as a filter to sort through vast amounts of routine health information to spot the people who need help.
A team of researchers set out to test whether a specific computer model, originally built using data from Latin America, could successfully find these hidden cases in a completely different population: adults in the United States. The model was designed to look at a person's basic health details—such as their age, blood pressure, weight, and whether they have diabetes or high blood pressure—without ever needing to look at a kidney function test result. The goal was to see if the model could rank people by risk, putting those most likely to have the disease at the top of the list, even if the model had never been trained on American patients. The researchers applied this tool to a massive, diverse group of over 228,000 American adults who had no prior diagnosis of kidney disease but had a routine blood test on record. They wanted to see if the model could identify the small fraction of these people who actually had moderately to severely reduced kidney function that had gone unnoticed.
The results showed that the model worked remarkably well at its primary job: sorting people by risk. When the researchers looked at the top 5% of people the model flagged as highest risk, it captured nearly a quarter of all the hidden kidney disease cases in the group. If they looked at the top 10%, it found over 40% of the cases. In the top 20%, the model identified more than 60% of all the undiagnosed cases. This means the model successfully concentrated the hidden patients into a small group, making it much easier for doctors to focus their limited testing resources where they would do the most good. The model performed this task with a high degree of accuracy, distinguishing between those with the disease and those without it far better than random chance. It was even more effective in the high-risk group of people who already had diabetes or high blood pressure, which is the exact population where screening is most needed.
However, the study also found a specific limitation that is common when moving medical tools from one country to another. While the model was excellent at ranking people from highest to lowest risk, it was not perfect at predicting the exact probability of disease. In the American group, the model tended to guess that the risk was higher than it actually was. For example, if the model predicted a certain number of people would have the disease, the actual number was lower. This is not a failure of the model's logic, but rather a difference in how common the disease is between the two populations. The researchers found that this error could be completely fixed by a simple mathematical adjustment, known as recalibration, which tuned the model's predictions to match the local reality without changing its ability to rank patients. This adjustment was crucial because it meant the model could be used in the United States with the same high performance it had in Latin America, provided the local numbers were updated.
The study also looked closely at whether the model treated different groups of people fairly. They examined if the model worked equally well for men and women, for people of different ages, and for people from different racial and ethnic backgrounds. The ranking ability remained strong across all these groups, meaning the model did not unfairly ignore any specific demographic. However, the accuracy of the risk predictions varied between groups. For instance, the model's predictions were very accurate for Black or African American participants but overestimated the risk for White and Hispanic/Latino participants. This variation happened because the actual rate of undiagnosed kidney disease was different in these groups. The researchers concluded that to use this tool fairly and effectively across a diverse population, doctors would need to apply specific adjustments for different groups rather than using a single, one-size-fits-all setting.
Ultimately, this research demonstrates that a tool built in one part of the world can be successfully adapted to find hidden disease in another, provided it is tuned to local conditions. The model offers a practical, low-cost way to identify people who need a confirmatory kidney test, especially in settings where resources are scarce or where many people have never been screened. By using routine health data that is already collected during regular doctor visits, the model can direct medical attention to the people who need it most, potentially preventing thousands of cases of kidney failure from going undetected. The study confirms that while the exact numbers might need local adjustment, the core ability to find the needle in the haystack travels well across borders, offering a new path to closing the gap in kidney disease care.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.