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

Estimated glomerular filtration rate (eGFR) is preferred, especially in more vulnerable populations, since direct GFR measurement involves an invasive and time-consuming procedure. Routinely, various existing equations are used to calculate eGFR, but there is room for improvement, motivating interest in using machine learning (ML) models for eGFR prediction. Evidence has shown that ML models can outperform equations in younger populations, though performing comparably in general populations. We included seven ML algorithms and three equations and conducted intra-comparisons and inter-comparisons across three variants: serum creatinine- (SCr), cystatin C- (CysC) and SCr/CysC-based models for predicting eGFR in children, adolescents, and young adults. Hyperparameter optimization was performed to support model selection. We compared the selected ML models built either with data from the full age range or between 2 and 25 years. Finally, we investigated including equation-derived estimates as additional features in ML models. Performance was evaluated using Lin’s concordance correlation coefficient (Lin’s CCC), P10, P30, and Passing-Bablok regression. the SCr/CysC-based random forest (RF) appeared as one of the most robust and strong ML models, and showed improved agreement with mGFR when compared to evaluated equations and single biomarker machine learning model and equation variants, with Lin’s CCC values of 0.80 (0.78, 0.82), P10 0.41 (0.38, 0.43) and P30 0.89 (0.87, 0.91). However, no performance difference was detected between RF models trained on different age cohorts. Finally, integrating equation-derived estimates as additional features into the SCr/CysC-based RF model improved the accuracy to P10 0.45 (0.42,0.48). When focusing on children, adolescents, and young adults (2–25 years), RF, irrespective of the training cohort, is one of the most robust ML models and had an improved or comparable performance with the evaluated equations. Using equation-derived estimates as features significantly improved RF performance. The best-performing model combined biomarkers, demographic information, and equation-derived estimates, indicating a complementary role of ML and domain knowledge captured by the evaluated GFR equations.