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

Article information

Korean J Nephrol. 2025;.j.krcp.25.004

Publication date (electronic) : 2025 August 11

doi : https://doi.org/10.23876/j.krcp.25.004

Sangwoo Moon¹, Junhyug Noh², Hee Gyung Kang^3,4,5,6, Keum Hwa Lee^7,8, Jae Il Shin^7,8, Ji Yeon Song⁹, Hee Sun Baek¹⁰, Min Hyun Cho¹¹, Jiwon Jung¹², Joo Hoon Lee¹², Heeyeon Cho¹³, Kyoung Hee Han¹⁴, Eun Mi Yang¹⁵, Seong Heon Kim^3,4, Ji Hyun Kim^4,16, Eujin Park^,17,18,*, Yo Han Ahn^,3,4,5,*

¹Department of Computer Science and Engineering, College of Engineering, Seoul National University, Seoul, Republic of Korea

²College of Artificial Intelligence, Ewha Womans University, Seoul, Republic of Korea

³Department of Pediatrics, Seoul National University Children’s Hospital, Seoul, Republic of Korea

⁴Department of Pediatrics, Seoul National University College of Medicine, Seoul, Republic of Korea

⁵Kidney Research Institute, Medical Research Center, Seoul National University College of Medicine, Seoul, Republic of Korea

⁶Wide River Institute of Immunology, Seoul National University, Hongcheon, Republic of Korea

⁷Department of Pediatrics, Yonsei University College of Medicine, Seoul, Republic of Korea

⁸Division of Pediatric Nephrology, Severance Children’s Hospital, Seoul, Republic of Korea

⁹Department of Pediatrics, Pusan National University Children’s Hospital, Pusan National University School of Medicine, Yangsan, Republic of Korea

¹⁰Department of Pediatrics, Yeungnam University Hospital, Yeungnam University College of Medicine, Daegu, Republic of Korea

¹¹Department of Pediatrics, School of Medicine, Kyungpook National University, Daegu, Republic of Korea

¹²Department of Pediatrics, Children’s Hospital, Asan Medical Center, University of Ulsan College of Medicine, Seoul, Republic of Korea

¹³Department of Pediatrics, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Republic of Korea

¹⁴Department of Pediatrics, Jeju National University College of Medicine and Graduate School of Medicine, Jeju, Republic of Korea

¹⁵Department of Pediatrics, Chonnam National University Hospital, Gwangju, Republic of Korea

¹⁶Department of Pediatrics, Seoul National University Bundang Hospital, Seongnam, Republic of Korea

¹⁷Department of Pediatrics, Korea University Guro Hospital, Seoul, Republic of Korea

¹⁸Department of Pediatrics, Korea University College of Medicine, Seoul, Republic of Korea

Correspondence: Yo Han Ahn Department of Pediatrics, Seoul National University Children’s Hospital, 101 Daehak-ro, Jongno-gu, Seoul 03080, Republic of Korea. E-mail: yhahn@snu.ac.kr

Eujin Park Department of Pediatrics, Korea University Guro Hospital, 148 Gurodong-ro, Guro-gu, Seoul 08308, Republic of Korea. E-mail: eujinpark@korea.ac.kr

*Yo Han Ahn and Eujin Park contributed equally to this study as co-corresponding authors.

Received 2025 January 3; Revised 2025 April 19; Accepted 2025 May 12.

Abstract

Introduction

Chronic kidney disease (CKD) is a significant global health concern, characterized by a high prevalence and a range of associated complications, including morbidity and mortality [1,2]. These factors contribute to a substantial economic burden and pose significant challenges for patients and their families [3–5]. Consequently, much effort has been made to predict kidney prognosis in patients with CKD, with the aim of planning adequate treatment and improving outcomes through early intervention [6]. However, the multifactorial nature of CKD progression presents a significant challenge to accurately predicting kidney outcomes in individual patients [7,8].

A substantial body of research has been conducted to predict the kidney prognosis in patients with CKD using traditional statistical models [9–11]. Although these methods have the advantage of identifying risk factors that are highly associated with kidney prognosis, their ability to reflect the complex interactions among multiple clinical variables is limited. Therefore, recent advances in machine learning have been applied to identify the complex interactions among a wide range of clinical variables, thereby improving the accuracy of predicting kidney function decline in CKD [12,13].

However, despite extensive research on prognosis prediction using machine learning in adult CKD, models for predicting prognosis in pediatric CKD have yet to be optimized. This is largely attributable to a lack of clinical data, given that the prevalence of pediatric CKD is significantly lower than that in adults. Moreover, it is difficult to apply models that have been validated in adults to children because of their distinctive clinical characteristics. These include the fact that the primary cause of CKD in children is congenital anomalies of the kidneys and urinary tract (CAKUT), that children’s kidneys continue to grow and develop, and that children with CKD generally experience slower disease progression compared to adults [14–16].

The objective of this study is to develop a novel data-driven approach for predicting kidney function decline in pediatric CKD by applying various machine learning techniques that can effectively utilize data from the KoreaN cohort study for Outcomes in patients With Pediatric Chronic Kidney Disease (KNOW-PedCKD).

Methods

Variables and data processing

The overall data preprocessing and model training processes are shown in Fig. 1. Participant records, including baseline characteristics, annual visit data with medical information, and laboratory test results, were converted into person-visit pair data for analysis. For example, if 5 years of follow-up data are available for a specific patient, the data can be organized as follows: (variable at 0-year, target at 1-year) to (variable at 4-year, target at 5-year) constitute the 1-year prediction dataset; (variable at 0-year, target at 2-year) to (variable at 3-year, target at 5-year) constitute the 2-year prediction dataset; and (variable at 0-year, target at 3-year) to (variable at 2-year, target at 5-year) constitute the 3-year prediction dataset. A total of 2,344 cases (person-visit pairs) were included in the 1-year prediction analysis, 1,911 cases in the 2-year prediction analysis, and 1,550 cases in the 3-year prediction analysis. The cases were divided into training (80.0%) and test (20.0%) sets. The clinical variables included in the model were classified into three categories according to their relevance to kidney function: high, moderate, and low [8]. The specific variables in each category, along with their null ratios, are presented in Supplementary Tables 1 and 2 (available online). The proportions of composite kidney outcomes by eGFR decline and time period are summarized in Supplementary Table 3 (available online).

Figure 1.

Overview of the model development.

Overview of the data collection, preprocessing, model training, and evaluation processes.

AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; CKD, chronic kidney disease; DT, decision tree; eGFR, estimated glomerular filtration rate; KNN, k-nearest neighbor; KRT, kidney replacement therapy; MICE, Multiple Imputation by Chained Equations; MLP, multilayer perceptron; RF, random forest; ROC, receiver operating characteristic curve; SHAP, Shapley Additive ExPlanations; SMOTE, Synthetic Minority Oversampling Technique; SVM, support vector machine; XGB, XGBoost.

To efficiently utilize the collected data, the Multiple Imputation by Chained Equations (MICE) technique was employed to handle missing data and maximize the use of available information [21]. MICE is a multiple imputation method that iteratively predicts and replaces missing values based on the relationships between variables. Furthermore, given the imbalance observed in the target variables, the Borderline-Synthetic Minority Oversampling Technique (Borderline-SMOTE) was applied to mitigate this issue [22]. Borderline-SMOTE is an oversampling approach that generates synthetic samples near the decision boundary of the minority class to address the imbalance. We will discuss the ablation study of MICE and Borderline-SMOTE in the Discussion section.

Model development and performance

We applied machine learning algorithms, such as decision tree (DT), k-nearest neighbor (KNN), support vector machine (SVM), random forest (RF), XGBoost (XGB), and multilayer perceptron (MLP), and compared their performance. DT is a model that solves classification or regression problems by recursively splitting data into subsets [23]. It uses a tree structure to make the predictions. KNN is a simple algorithm that classifies or predicts a value based on the k-nearest data points [24]. It relies on measuring the distance between points in the feature space. SVM is an algorithm that determines the optimal hyperplane to separate data in a high-dimensional space [25]. It was used for classification and regression tasks. RF is an ensemble learning method that combines many DTs to perform classification or regression [26]. Each tree was trained on a random subset of the data. XGB is an optimized version of the Gradient Boosting algorithm known for its speed and high performance in many machine learning tasks [27]. An MLP is a type of neural network with multiple layers of neurons that can model complex nonlinear relationships in data [28]. It is commonly used in deep learning tasks. The hyperparameter search space used to determine the best model for each model type is presented in Table 1.

Table 1.

Hyperparameter search range for each model

We primarily evaluated model performance using the area under the receiver operating characteristic curve (AUROC) as the primary evaluation metric. Additionally, we recorded supplementary metrics including accuracy, precision, recall, the F1 score, and the area under the precision-recall curve (AUPRC). To analyze the importance of the features in the trained model, we utilized the Shapley Additive ExPlanations (SHAP) method [29].

Results

Characteristics of the cohort

We included 539 pediatric patients with CKD in our analysis. The mean age of the patients was 9.09 ± 5.17 years, with a mean eGFR of 58.34 ± 29.35 mL/min/1.73 m². Of the patients, 68.1% were male and 46.0% had CAKUT as the primary underlying cause of CKD. The mean height Z-score was –0.88 ± 1.55, and the mean body mass index Z-score was –0.37 ± 1.40. The mean spot protein-to-creatinine ratio (UPCR) was 1.64 ± 5.63 mg/mg. The mean follow-up duration was 5.37 years. A comparison of the key characteristics of the training and test sets revealed no statistically significant differences in demographics, clinical findings, or laboratory results (Table 2).

Table 2.

Demographics and clinical characteristics of the training and test sets

Model performance

The performance of the machine learning algorithms in predicting composite kidney outcomes based on the three definitions of kidney function decline is presented in Figs. 2 to 4. Models predicting a decline of more than 20% in eGFR, progression to KRT, or death demonstrated superior AUROCs compared to models using other definitions of kidney function decline. Based on these results, we selected this definition for further feature importance analysis.

Figure 2.

Performance comparison of prognostic prediction models trained with various feature Sets for 1- to 3-year outcomes (any eGFR decline or progression to kidney replacement therapy or death).

The experimental results were categorized based on the degree of eGFR decline. The area under the receiver operating characteristic curve (AUC) performance of each machine learning model was then compared across different prediction years and the features utilized. H features, highly relevant 22 features; H + M features, 42 highly and moderately relevant features; all features, 48 features.

DT, decision tree; eGFR, estimated glomerular filtration rate; KNN, k-nearest neighbor; MLP, multilayer perceptron; RF, random forest; SVM, support vector machine; XGB, XGBoost.

Figure 3.

Performance comparison of prognostic prediction models trained with various feature sets for 1- to 3-year outcomes (eGFR decline greater than 5% or progression to kidney replacement therapy or death).

The experimental results were categorized based on the degree of eGFR decline. The area under the receiver operating characteristic curve (AUC) performance of each machine learning model was then compared across different prediction years and the features utilized. H features, highly relevant 22 features; H + M features, 42 highly and moderately relevant features; all features, 48 features.

DT, decision tree; eGFR, estimated glomerular filtration rate; KNN, k-nearest neighbor; MLP, multilayer perceptron; RF, random forest; SVM, support vector machine; XGB, XGBoost.

Figure 4.

Performance comparison of prognostic prediction models trained with various feature sets for 1- to 3-year outcomes (eGFR decline greater than 20% or progression to kidney replacement therapy or death).

The experimental results were categorized based on the degree of eGFR decline. The area under the receiver operating characteristic curve (AUC) performance of each machine learning model was then compared across different prediction years and the features utilized. H features, highly relevant 22 features; H + M features, 42 highly and moderately relevant features; all features, 48 features.

DT, decision tree; eGFR, estimated glomerular filtration rate; KNN, k-nearest neighbor; MLP, multilayer perceptron; RF, random forest; SVM, support vector machine; XGB, XGBoost.

Any decline in the estimated glomerular filtration rate

A machine learning algorithm was employed to predict any decline in eGFR, progress to KRT, or death at 1, 2, and 3 years. The model exhibited superior predictive performance in forecasting any eGFR decline in the 2- and 3-year periods compared with the 1-year period. This was achieved not only when using highly relevant clinical variables but also when using moderately and highly relevant clinical variables, as well as when using all clinical variables, such as XGB (AUROCs of 0.67, 0.71, and 0.70, respectively) and RF (AUROCs of 0.68, 0.71, and 0.68, respectively) (Fig. 2). In contrast, the predictive models based on DT, KNN, SVM, and MLP exhibited inferior performance compared to RF and XGB in all eGFR decline prediction models.

Decline in estimated glomerular filtration rate greater than 5%

As with previous analyses, the optimal predictive performance for a decline in eGFR of ≥5%, or progression to KRT or death, was observed at 3 years. Moreover, RF and XGB exhibited high performance for all highly relevant clinical variables, moderately relevant and highly relevant clinical variables, and all clinical variables such as XGB (AUROCs of 0.76, 0.76, and 0.75, respectively) and RF (AUROCs of 0.74, 0.75, and 0.75, respectively) (Fig. 3). The predictive models based on DT, KNN, SVM, and MLP exhibited suboptimal performance.

Decline in estimated glomerular filtration rate greater than 20%

The model for predicting a decline in eGFR of 20% or greater, or progression to KRT or death, demonstrated consistently good predictive performance across the 3-year observation period, in contrast to the results of previous models. Notably, the model demonstrated the best performance in predicting a kidney function decline at the 1-year mark in comparison with the 2- and 3-year periods. This was achieved when using moderately and highly relevant clinical variables compared to all clinical variables and when using only highly relevant clinical variables, such as XGB (AUROCs of 0.88, 0.89, and 0.87, respectively) and RF (AUROCs of 0.86, 0.88, and 0.89, respectively) (Fig. 4). The predictive models based on DT, KNN, SVM, and MLP exhibited suboptimal performance.

Feature importance

In this section, we performed a feature importance analysis using RF, which generally showed high performance alongside XGB. Feature importance analysis for predicting the composite kidney outcome over 1 year identified UPCR as the most influential variable in the prediction models. This was followed by baseline eGFR and hemoglobin, serum levels of albumin, and total bilirubin. An increased risk of composite kidney outcomes was associated with elevated UPCR, lower hemoglobin levels, lower serum albumin levels, and higher total bilirubin levels. Additional predictors included age, total blood iron-binding capacity, neutrophil count, systolic blood pressure, reticulocyte count, serum calcium, height Z-score, serum iron, serum uric acid, serum magnesium, urine pH, serum phosphorus, and history of individual medication use, such as calcium-based phosphorus binders. Fig. 5 shows these variables and highlights their relative contributions to the prediction model. The importance of the features of the 2- and 3-year prediction models is shown in Supplementary Figs. 1 and 2 (available online).

Figure 5.

Feature importance analysis for predicting a 1-year eGFR decline.

This is Shapley Additive ExPlanations (SHAP) analysis for predicting a 1-year eGFR decline of ≥20%, kidney replacement therapy, or death. The analysis was conducted based on the random forest model that showed the highest performance. Higher SHAP values, based on absolute values, indicate greater impact on the model’s predictions. The analysis results of the prediction model for 2 to 3 years later are provided in Supplementary Fig. 1 and 2 (available online).

BP, blood pressure; ESA, erythropoietin-stimulating agent; eGFR, estimated glomerular filtration rate; H, highly relevant feature; L, low relevant feature; M, moderately relevant feature; RBC, red blood cell; TIBC, blood total iron-binding capacity; UPCR, urine protein-to-creatinine ratio; WBC, white blood cell.

Discussion

In this study, we developed a practical prediction tool for assessing short-term kidney function decline in children with CKD using various machine learning techniques. Our findings revealed that the decline in kidney function is influenced by multiple factors, including established risk factors such as baseline eGFR and UPCR, in addition to the clinical variables identified in our analysis.

The importance of risk prediction models in predicting kidney failure has been emphasized in the recently updated KDIGO guidelines, which advocate their application to individuals in real-world clinical settings [19]. In adults, validated risk equations for predicting kidney failure have been developed and are widely used to guide planning and care decisions. However, pediatric CKD progresses slower than adult CKD, necessitating the development of prediction tools specifically tailored for children. To date, the most extensive work on predicting kidney outcomes in pediatric patients with CKD has been performed in a cohort of Chronic Kidney Disease in Children (CKiD) studies that developed a pediatric end-stage kidney disease (ESKD) risk prediction equation using strong predictors of progression to ESKD, including underlying CKD etiology, initial GFR stages, and proteinuria [10]. Recently, CKiD investigators proposed a prediction tool that uses statistical learning methods to predict the time to KRT initiation in pediatric patients with CKD using multiple clinical variables [12]. Although this web-based prediction tool provides valuable insights into the probability and time frame for pediatric patients with CKD to reach kidney failure, which is advantageous for clinical decision-making and patient care planning, it does not fully meet the needs of clinicians. Clinicians typically seek to understand the risk of a significant decline in eGFR over a shorter period, such as after 2 or 3 years [30,31]. Furthermore, the prediction tool was developed and validated using data from pediatric CKD cohorts in the United States and Europe; therefore, it may not adequately reflect the characteristics of patients in centers that primarily treat Asian pediatric patients with CKD.

In this study, we developed models for predicting kidney function decline after a short period in Korean pediatric patients with CKD. Our results showed that the majority of algorithms performed better when predicting an eGFR decline of >20%, progression to KRT, or death. The XGB model performed best at predicting an eGFR decline of more than 20% or progression to KRT or death over 1 year by applying moderately and highly relevant clinical variables, and the RF model performed best at predicting an eGFR decline of more than 20% or progression to KRT or death over 1 year by applying all clinical variables. Interestingly, tree-based models, such as XGB and RF, have shown better performance than the typical MLP-based deep learning model in previous studies [32,33]. The studies have shown that tree-based models (such as XGB and RF) often outperform deep learning models such as MLP, despite having significantly lower computational costs. This is particularly evident when working with small datasets approximately 10,000 samples, where tree-based models generally yield better performance. In a study by Grinsztajn et al. [32], analysis of 45 diverse benchmarks identified the following contributing factors: 1) neural networks tend to be biased toward overly smooth solutions, 2) uninformative features affect MLP-based neural networks more severely, and 3) data are non-invariant by rotation, which should be considered in the learning procedure. As a prospective medical dataset with a relatively small sample size approximately 2,000, The KNOW-PedCKD aligns with the scale at which tree-based models typically show superior performance, as suggested by previous research.

In addition, classification tasks in the medical domain, including this study, often involve class imbalance between normal and abnormal classes; Therefore, we used AUROC as a reliable metric for model comparison, in line with previous studies [34,35]. AUROC assesses the balance between how well a model detects cases of declined kidney function (true positive rate) and how well it excludes other cases (false positive rate) over a range of thresholds. The closer the AUROC value is to 1, i.e., the wider the AUROC, the better the performance of the model. Especially when dealing with the problem of class imbalance, AUROC is a valuable metric to evaluate a model’s ability to accurately identify minority classes.

Notably, using only highly relevant clinical variables did not consistently result in the highest predictive performance. Our findings indicated that a decline in kidney function was highly predictive, even when accounting for clinical variables within the cohort that were not previously considered to be strongly correlated. This finding suggests that complex clinical variables may exert an influence on kidney function decline beyond the traditional strong predictors of kidney function decline in children with CKD. Furthermore, UPCR, eGFR, hemoglobin, albumin, chloride, and age were identified as significant influencing factors in this study, and these findings align with those of CKiD researchers, who identified UPCR and eGFR as the most important predictors of KRT, and other blood tests such as albumin, phosphate, and chloride as important predictors [8,12]. These results confirm the impact of several modifiable contributing factors, including proteinuria, blood levels of albumin, hemoglobin, and chloride, in addition to non-modifiable contributing factors such as baseline eGFR and age, and suggest the possibility of preventing or delaying the progression of pediatric CKD by correcting these laboratory abnormalities through more intensive therapeutic interventions at an earlier stage of the disease.

Although this study had several notable strengths, it also had some limitations. First, the relatively small dataset compared with those typically used in learning models may have limited our ability to develop a robust predictive tool [14]. Second, the inclusion of patients with extreme characteristics may have resulted in inaccurate estimates. Third, our model was based on a surrogate endpoint of eGFR decline, which may have weaker predictive power than a stronger endpoint such as KRT or mortality [31]. Finally, the underlying CKD etiology, which was previously considered as an important predictor of kidney function decline, did not appear to exert a significant influence on the model [16]. The observed discrepancy remains challenging to elucidate; however, it may be attributable to inherent demographic variations between the two cohorts, including racial disparities, the distinct timeframes predicted by the model, and the utilization of disparate definitions for the endpoint of eGFR decline. Consequently, we are planning further studies to compare the differences in risk factors for disease progression between these pediatric CKD cohorts. Concomitantly, we will employ the recently obtained KNOW-PedCKD dataset to further train and validate the model, thereby enhancing its robustness and generalizability. The KNOW-PedCKD study aims to develop a model that is clinically applicable and readily accessible online.

However, we endeavored to use our limited data effectively by applying MICE and Borderline-SMOTE as described in the ‘Variables and data processing’ section. For all prediction horizons (1–3 years), the combination of all techniques consistently achieved the highest performance across the AUROC, F1-score, and AUPRC (Table 3). Additionally, the KNOW-PedCKD study is a representative pediatric CKD cohort in Asia. It has been recruiting participants since 2011 and has the strengths of a large sample size and a long follow-up period with minimal dropouts, resulting in a wealth of baseline and follow-up data, including centrally measured biomarkers.

Table 3.

Ablation study of machine learning technique

In summary, we developed a tool for predicting short-term kidney function decline using data from the KNOW-PedCKD study. This tool is intended to support clinical decision-making in pediatric patients with CKD. The prediction model was trained using routinely measured data and was validated internally. However, further external validation using other pediatric CKD datasets is necessary.

Supplementary Materials

Article information Continued