Machine Learning Model Improves Kidney Function Estimates in Young People

A computer model that uses machine learning has shown better results when estimating how well young people's kidneys work. The model uses two substances found in blood, along with basic information about a person, such as their age and sex. According to a study published in BMC Pediatrics, its estimates were closer to actual kidney function measurements than those made using standard formulas. The study focused on children, teenagers and young adults. It was led by TingTing Shao, Felipe Kenji Nakano and other researchers, including Pierre Delanaye and Celine Vens. They compared different machine learning methods with traditional formulas for estimating the glomerular filtration rate (GFR). GFR is the standard way to measure how well the kidneys filter waste from the blood.

Doctors usually estimate GFR instead of measuring it directly, especially for patients who may be more vulnerable. This is because directly measuring GFR takes time and involves a procedure that goes inside the body. Doctors normally use different mathematical formulas to estimate GFR. However, the researchers say these formulas could be improved. This has led to growing interest in machine learning, which allows computers to learn patterns from data. Earlier research found that machine learning models can sometimes give better estimates than traditional formulas in younger people. However, in the general population, the two methods usually perform about equally well.

The researchers tested seven machine learning methods and three traditional formulas. They compared three ways of estimating kidney function: using only a blood substance called serum creatinine, using only another substance called cystatin C, and using both substances together. These substances are known as biomarkers because they can give clues about what is happening inside the body. The researchers also adjusted the settings of the machine learning models to find which worked best. They compared models that used the same biomarkers and models that used different combinations. They also tested whether models trained using people of all available ages performed differently from models trained only using people aged 2 to 25 years. Finally, they checked whether giving the machine learning models estimates from traditional formulas as extra information would improve their results.

To measure how well the methods worked, the researchers used three types of statistical tests: Lin's concordance correlation coefficient, P10 and P30 measurements, and Passing-Bablok regression. These tests helped them see how closely the estimates matched actual kidney function measurements. One of the strongest and most reliable machine learning methods was a random forest model that used both serum creatinine and cystatin C. A random forest combines the results of many decision-making models to make a prediction. It achieved a Lin's concordance correlation coefficient of 0.80 (0.78, 0.82), which showed how closely its estimates agreed with actual measurements. For P10, 41 percent of its estimates (0.38, 0.43) were within 10 percent of the measured values. For P30, 89 percent (0.87, 0.91) were within 30 percent of the measured values.

These results showed that the model using both blood biomarkers gave estimates that were closer to the actual GFR measurements than the traditional formulas tested. It also performed better than machine learning models and formulas that used only one biomarker. In other words, using serum creatinine and cystatin C together worked better than using either one alone, whether the researchers used a traditional formula or machine learning.

The researchers also found something useful about the data used to train the models. They could not find a difference in performance between random forest models trained using different age groups. Training a model only with data from people aged 2 to 25 years did not make it more accurate than training it with data from the full available age range.

The biggest improvement came when the researchers combined machine learning with traditional formulas. They gave the random forest model, which already used serum creatinine and cystatin C, extra information from the GFR estimates produced by the traditional formulas. This made the model more accurate. The percentage of estimates that were within 10 percent of the actual measurements increased to 45 percent (0.42, 0.48). According to the researchers, the best model used blood biomarkers, basic personal information and estimates from traditional formulas together. This suggests that machine learning and traditional medical knowledge can work well together. Instead of completely replacing traditional formulas, machine learning can use the information they provide to make better predictions.

The study was published in BMC Pediatrics on 23 October 2026. It was an open-access study, meaning anyone could read it, and it had been reviewed by other researchers. The published article was an accepted version shared early with a permanent digital object identifier (DOI). It could still receive further edits before the final official version was published. Based on the seven machine learning methods, three formulas and people aged 2 to 25 years examined in the study, the researchers found that a random forest model using both blood biomarkers was one of the strongest and most reliable ways to estimate kidney function in this age group. It worked especially well when it also used estimates from traditional formulas.

Subject of Research: Machine learning models that estimate how well the kidneys filter blood in children, teenagers and young adults

Article Title: Machine learning models to estimate the glomerular filtration rate in children, adolescents and young adults

Article References: Shao, T., Nakano, F. K., Van Kleunen, L., Ilídio, P., Björk, J., Derain-Dubourg, L., Hansson, M., Littmann, K., Bökenkamp, A., Berg, U., Åsling-Monemi, K., Sjöström, P., Nyman, U., Lanot, A., Pottel, H., Delanaye, P., & Vens, C. (2026). Machine learning models to estimate the glomerular filtration rate in children, adolescents and young adults. BMC Pediatrics. [https://doi.org/10.1186/s12887-026-07877-1](https://doi.org/10.1186/s12887-026-07877-1)

Image Credits: Created using artificial intelligence (AI)

DOI: 10.1186/s12887-026-07877-1

Keywords: glomerular filtration rate, machine learning, random forest, serum creatinine, cystatin C, eGFR, children's medicine, kidney function, biomarkers, BMC Pediatrics, Machine, learning