Development of a prediction tool for kidney function decline in children with chronic kidney disease

A paucity of literature exists on the development of predictive tools for the decline of kidney function in pediatric chronic kidney disease (CKD). The objective of this study is to develop and internally validate a tool for the short-term prediction of a kidney function decline in pediatric patients with CKD. A total of 539 patients participating in the KNOW-PedCKD (KoreaN cohort study for Outcomes in patients With Pediatric Chronic Kidney Disease) were evaluated for 48 variables related to sociodemographic characteristics, laboratory data, and treatment use. These variables were assessed as potential predictors of a kidney function decline in pediatric patients with CKD using a range of machine learning algorithms. The models demonstrated strong predictive performances in identifying kidney function decline, defined as an estimated glomerular filtration rate (eGFR) decline of ≥20%, which includes progression to kidney replacement therapy or death. The random forest and XGBoost models demonstrated the best performance in predicting eGFR outcomes at 1 year compared with 2 and 3 years, respectively. The spot urine protein-to-creatinine ratio was the most influential variable in the prediction model, followed by baseline eGFR and serum albumin, chloride, and hemoglobin levels. A tool for predicting kidney function decline in children with CKD over a short period of time was developed using potential predictors and machine learning methods in a large Korean pediatric CKD cohort.