Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction

Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction Shiyou Guan, Yi Gan, Anxia Sun, Youfang He, Ping Yuan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7281320/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract [Background] Primary nephrotic syndrome (NS) in children is a common chronic kidney disease in pediatrics, characterized by complex pathogenesis, heterogeneous clinical manifestations, and easy recurrence. Existing clinical diagnosis mainly relies on symptoms and laboratory tests, lacking efficient and accurate early risk prediction tools, which limits the implementation of early intervention and individualized management. With the development of artificial intelligence technology, the construction of machine learning prediction models based on multidimensional clinical data has provided new possibilities for the early identification and precise intervention of NS. [Methods] This study retrospectively collected clinical data of 771 children with primary kidney diseases in the Pediatric Nephrology Ward of the Affiliated Hospital of Zunyi Medical University from 2009 to 2023, including 376 children with NS and 395 children with acute glomerulonephritis. The data were improved by preprocessing methods such as multiple imputation, standardization and coding, and general demographic characteristics, laboratory test indicators and renal pathological characteristics were screened as modeling variables. Four machine learning algorithms, GBDT, XGBoost, random forest (RF) and LightGBM, were used to construct a risk prediction model for the onset of the disease. The model performance was evaluated using five-fold cross validation, and the feature importance was explained by the SHAP method. [Results] All four models showed high predictive ability, among which the random forest model performed best, reaching an accuracy of 99.14%, precision of 99.13%, recall of 99.16%, F1 score of 0.9914 and AUC value of 0.9983 on the validation set. SHAP analysis results showed that indicators such as plasma IgG, total protein, complement C3, and ASO titer contributed significantly to model prediction and were highly consistent with the clinical pathological mechanism of NS, verifying the reliability and clinical interpretability of the model. [Conclusion] This study successfully constructed a risk prediction model for NS in children based on machine learning algorithms, which has high accuracy and good clinical interpretability, and provides strong data support for early screening and individualized treatment of NS. In the future, multi-center and multi-omics validation should be carried out to further improve the generalization ability and clinical application value of the model. Primary nephrotic syndrome in children machine learning random forest risk prediction SHAP interpretation Figures Figure 1 Figure 2 1 Introduction Primary nephrotic syndrome (NS) in children is one of the most common chronic kidney diseases in children, with an annual incidence of approximately 2–7/100,000, which seriously affects the quality of life and growth and development of children[1]. The disease is characterized by massive proteinuria, hypoalbuminemia, significant edema and hyperlipidemia. The course of the disease tends to relapse, and some children may progress to hormone-dependent or hormone-resistant nephropathy, which increases the difficulty of treatment and the risk of long-term renal insufficiency[2]. Although a variety of immunosuppressants and supportive therapies are currently used in the treatment of NS, its complex etiology and highly heterogeneous clinical manifestations still pose great challenges to early diagnosis and intervention. In recent years, researchers have gradually realized that the occurrence and development of NS may involve the combined effects of multiple factors such as genetics, immunity, metabolism, and environment [3]. Some studies suggest that certain immune indicators (such as IgG, complement C3), inflammatory factors, endocrine abnormalities, and renal tissue pathological changes all play an important role in the pathogenesis of NS. This complex etiology determines that the manifestations of NS in different children vary greatly. Therefore, the traditional "empirical judgment" diagnostic model can no longer meet the needs of precision medicine. For the early identification and risk stratification of NS, more systematic, efficient, and intelligent analysis tools are needed. With the widespread use of electronic medical records and hospital information systems, the acquisition of clinical data has become increasingly convenient. The multi-dimensional data of pediatric kidney disease patients (such as demographic information, biochemical indicators, immune parameters, pathological data, etc.) provide a solid data foundation for disease prediction modeling. How to accurately extract characteristic variables with diagnostic value from complex clinical data and construct a risk prediction model with generalization ability has become a hot topic in current research. Traditional statistical methods have limitations in dealing with nonlinearity, multivariate interactions, and high-dimensional features, which has also opened up new breakthroughs for the application of artificial intelligence technology, especially machine learning methods. As a core branch of artificial intelligence, machine learning has powerful data modeling and feature learning capabilities. It can automatically identify patterns from historical data and predict future trends without explicit programming. In recent years, algorithms such as support vector machines, random forests, gradient boosting trees, and XGBoost have been widely used in medical prediction tasks such as disease risk assessment, diagnostic classification, and prognosis analysis, and have achieved good results. Compared with traditional methods such as linear regression, machine learning models can better handle nonlinear feature relationships, outliers, and missing information, and are more suitable for risk modeling needs in complex clinical contexts. Against this background, this paper takes children with NS and acute glomerulonephritis admitted to the Affiliated Hospital of Zunyi Medical University between 2009 and 2023 as the research objects, integrates their multi-dimensional clinical, laboratory and pathological data, and attempts to use a variety of mainstream machine learning algorithms to construct a risk prediction model for the incidence of primary nephrotic syndrome in children, aiming to explore the feasibility and clinical value of this strategy in the early identification and precise intervention of NS in children. 2 Methods 2.1 Patient Enrollment This is a single-center retrospective study that included 771 children with primary kidney disease who were hospitalized in the Pediatric Nephrology Ward of the Affiliated Hospital of Zunyi Medical University from January 1, 2009 to December 31, 2023. Inclusion criteria included: (1) age 1–18 years; (2) primary kidney disease confirmed by renal puncture biopsy pathology based on clinical manifestations and laboratory test results; (3) complete clinical data, including demographic characteristics, laboratory test indicators, and renal histopathological examination results. According to the final diagnosis, the children were divided into a primary nephrotic syndrome (NS) group (n = 376) and an acute glomerulonephritis (AGN) group (n = 395) for subsequent model construction and comparative analysis. The basic information, laboratory tests and pathology reports of the children were all obtained from the hospital's electronic medical record system. To ensure the authenticity and consistency of the data, two independent researchers extracted and reviewed the data according to unified standards. For controversial case data, a third senior attending physician was invited to conduct the final review. This study strictly adhered to the ethical principles of the Declaration of Helsinki on research involving human subjects and was approved by the Ethics Committee of the Affiliated Hospital of Zunyi Medical University (Ethics Approval No.KLL-2025-070). All included cases signed informed consent during their initial hospitalization, or their legal guardians signed on their behalf, allowing their clinical data to be used for scientific research analysis. To protect patient privacy, all data were removed from personally identifiable information during use to ensure anonymization of the information. 2.2 Data Collection and Feature Selection This study collected multi-dimensional clinical information and laboratory test data from children, striving to fully cover potential risk factors associated with primary kidney disease. All data were retrieved and sorted through the hospital's electronic medical record system to ensure the integrity and consistency of the information. The main characteristic information collected includes the following aspects: General demographic characteristics mainly include sociodemographic parameters such as the child's age, gender, place of residence (urban/rural), and education level [4]. These factors can reflect the potential association between social environment, lifestyle and disease risk to a certain extent, and are of great reference value in epidemiological studies of childhood kidney disease. Secondly, laboratory test indicators are divided according to physiological systems, mainly including: (1) Hematological indicators: platelet count, fibrinogen level, to evaluate blood coagulation status and potential blood system abnormalities; (2) Immune function indicators: IgG level in urine and plasma, complement C3 level, hepatitis B surface antibody (HBsAb) status, antistreptococcal hemolysin O (ASO) titer, etc., reflecting the body's immune status and history of infection. These indicators are closely related to the immune pathogenesis of NS [4–5]; (3) Endocrine function indicators: thyroid stimulating hormone (TSH ) levels, which are used to assess the thyroid function status of children and indirectly reflect the metabolic regulation status; (4) biochemical metabolic indicators: including total protein, plasma albumin, prealbumin, urea nitrogen, blood creatinine, serum cystatin C, total cholesterol, vitamin D levels, etc., which systematically reflect the nutritional status, metabolic level and glomerular filtration function; (5) renal function indicators: such as urine protein quantitative, urine protein qualitative, hematuria (microscopic hematuria or macroscopic hematuria), hypertension, oliguria and edema. These indicators are important references for clinically assessing the severity of NS [6]. All enrolled children underwent renal puncture biopsy, and the results of renal tissue pathology were also included in the scope of data collection. Pathological characteristics mainly include the degree of glomerular lesions (such as mesangial hyperplasia, capillary wall changes, etc.), the degree of tubular interstitial damage (such as interstitial fibrosis, inflammatory cell infiltration), etc. These pathological parameters are of great significance for distinguishing different types of primary nephrotic syndrome and their prognosis assessment. After the data is sorted, all characteristic variables enter the subsequent data preprocessing and modeling process. 2.3 Data Preprocessing Before model construction, in order to ensure data quality and algorithm performance, this study systematically preprocessed the original clinical data. First, in order to solve the problem of missing values in some variables, the chained multiple imputation method (Multiple Imputation by Chained Equations, MICE) was used to fill in the missing values in order to retain the sample information to the greatest extent and reduce data bias. Subsequently, outliers were detected for each continuous variable through box plots and variable distribution diagrams, and the source of outliers was determined by combining Z scores and IQR ranges. Obviously erroneous data were eliminated, and reasonable extreme values were retained and recorded for subsequent model robustness assessment. Considering the dimensional differences between the variables, in order to avoid affecting the convergence efficiency and learning effect of the model, all numerical variables were Z-score standardized to make their mean 0 and standard deviation 1, which was adapted to the input requirements of various machine learning algorithms. At the same time, categorical variables such as gender and place of residence were converted into binary vectors by one-hot encoding to achieve effective recognition and calculation of the algorithm. The above preprocessing steps provide a clean, structurally unified, high-quality data foundation for subsequent feature selection and modeling. 2.4 Model Construction Based on the preprocessed clinical data, this study used four mainstream ensemble learning algorithms to construct a risk prediction model for primary nephrotic syndrome in children, namely, Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Random Forest (RF) and Light GBM (LGBM) [ 7 ] . In the process of building the GBDT model, the base learner was set as the CART regression tree, the initial learning rate was set to 0.1, the maximum tree depth (max_depth) was limited to 5 layers to prevent overfitting, and the number of weak learners (n_estimators) was initially set to 100, which was dynamically adjusted through cross-validation. As an enhanced boosting algorithm, XGBoost is further optimized in parameter settings, using undersampling ratio (subsample = 0.8), column sampling ratio (colsample_bytree = 0.8), L2 regularization coefficient (lambda = 1), maximum tree depth max_depth set to 6, and learning rate set to 0.05 to improve generalization ability and prevent overfitting; at the same time, the early stopping mechanism (early_stopping_rounds = 30) is used to control the number of training rounds to avoid redundant calculations. The random forest model uses a decision tree ensemble based on the bagging strategy, setting the number of decision trees (n_estimators) to 100, the maximum number of features (max_features) to the square root strategy (sqrt), and no limit on tree depth (max_depth = None). By increasing the number of trees, the variance is reduced and the model stability is improved. The LightGBM model was constructed using a leaf-wise intelligent splitting strategy, with the initial learning rate set to 0.05, the maximum tree depth set to 7, the subsample ratio set to 0.8, and the minimum number of leaf nodes (min_data_in_leaf) set to 20 to improve computational efficiency and accuracy. To ensure optimal model performance, this study used a five-fold cross-validation (k = 5) strategy to evaluate the performance of different models on the training set and validation set, and further fine-tuned the parameters through grid search combined with Bayesian optimization [ 8 ] . 2.5 Model Evaluation and Validation In order to comprehensively evaluate the performance of various machine learning models in the task of predicting the risk of primary nephrotic syndrome in children, this study used a variety of classification performance indicators for quantitative analysis. The main evaluation indicators include accuracy, recall, precision, F1-score, and area under the receiver operating characteristic curve (AUC). Among them, accuracy reflects the proportion of correct classification of the model in all predictions, which is used to measure the overall classification ability; recall represents the proportion of all true positive samples successfully detected by the model, focusing on evaluating the model's recognition sensitivity for sick children; precision refers to the proportion of true positive samples in all predicted positive samples, which is used to measure the reliability of the model's prediction results; F1 score is the harmonic mean of precision and recall, taking into account the accuracy and comprehensiveness of the model; AUC value reflects the overall ability of the model to distinguish between positive and negative samples by drawing the receiver operating characteristic curve (ROC curve). The closer the AUC value is to 1, the better the model's identification performance. 3 Results 3.1 Model Comparison 3.1.1 Model Evaluation Metrics Based on five-fold cross validation, this study systematically evaluated the performance of four machine learning models, GBDT, LightGBM, random forest (RF) and XGBoost, in the task of predicting the risk of primary nephrotic syndrome in children. The model evaluation indicators include accuracy, precision, recall and F1-Score. The specific comparison results are as follows: the GBDT model has an accuracy of 0.9828, a precision of 0.9827, a recall of 0.983, and an F1-Score of 0.9828; the LightGBM model performs slightly better, with an accuracy of 0.9871, a precision of 0.9872, a recall of 0.9869, and an F1-Score of 0.9871; the Random Forest model performs best in all indicators, with an accuracy of 0.9914, a precision of 0.9913, a recall of 0.9916, and an F1-Score of 0.9914; the XGBoost model has similar indicators to GBDT, with an accuracy of 0.9828, a precision of 0.9827, a recall of 0.9827, and an F1-Score of 0.9827. Comprehensive comparison shows that all four models show high prediction accuracy, but the random forest model is slightly higher than other models in accuracy, precision, recall and F1 score, showing better overall performance, becoming the preferred model for subsequent in-depth analysis and application discussion. The specific results are shown in Table 1 . 3.1.2 Model AUC Comparison To further evaluate the ability of each model to distinguish the risk of primary nephrotic syndrome in children, this study plotted the receiver operating characteristic (ROC) curves of four machine learning models and calculated the area under the curve (AUC) value (Fig. 1 ). The results showed that all four models showed extremely high discrimination ability, and the AUC values were close to 1, indicating that the performance of the models in positive and negative sample recognition was very stable. Among them, the LightGBM model had the highest AUC value of 0.9992, followed by the GBDT model (AUC = 0.9990), the random forest model was slightly lower (AUC = 0.9983), and the XGBoost model AUC was 0.9982. The ROC curve shapes were significantly better than the random guessing reference line (diagonal line), and maintained a very high true positive recognition rate under low false alarm rates, further verifying that each model had excellent discrimination performance in this research task. In general, although the differences in AUC among the four models are small, the LightGBM and GBDT models are slightly better in that the ROC curve is closer to the upper left corner, suggesting that they have stronger sensitivity and applicability in clinical risk screening. Table 1 Performance evaluation results of different machine learning algorithms Model Accuracy Precision Recall F1-Score GBDT 0.9828 0.9827 0.983 0.9828 LGBM 0.9871 0.9872 0.9869 0.9871 RF 0.9914 0.9913 0.9916 0.9914 XGB 0.9828 0.9827 0.9827 0.9827 3.2 SHAP analysis visualization This study conducted a feature importance analysis on the final model based on the SHAP (SHapley Additive exPlanations) method. As a feature attribution method based on game theory, SHAP can quantify the positive and negative impact direction of each feature on a single prediction result and its contribution, thereby providing an intuitive explanation for the internal mechanism of the "black box model". Figure 2 shows the feature importance ranking and distribution based on SHAP values. It can be seen that plasma IgG is the variable that most significantly affects the output of the model, and its SHAP value is significantly negatively distributed, indicating that the reduced IgG level has a higher weight in the model prediction of NS (Nephrotic Syndrome), which is consistent with the immune disorder mechanism of massive IgG loss in children with NS. Secondly, microscopic hematuria and total protein levels also showed strong predictive ability, among which hypoproteinemia, as a classic feature of NS, is closely related to model decision-making. In addition, hypertension (HTN), antistreptococcal hemolysin O (ASO) titer, gross hematuria, and platelet count also played a significant role in the model. The significance of immune indicators such as ASO and C3 further supports the immune-related pathogenesis of NS. The SHAP values of renal function and metabolism-related indicators such as plasma albumin, complement C3, and urea nitrogen also show that they have a moderate contribution to the model output. It is worth noting that some non-traditional core features, such as variables such as hospital stay days, weight, and onset, also affect the model judgment to a certain extent. This may be related to the course of the disease, the severity of the disease, and treatment compliance, suggesting that the model not only captures the direct manifestations of the disease, but also takes into account information at the clinical management level. Overall, the results of SHAP analysis show that the model construction not only has high predictive ability, but also shows good clinical interpretability. Most of the key variables it relies on are consistent with the pathophysiological mechanism, clinical manifestations and diagnostic criteria of NS, providing a trust basis for the application of the model in actual medical scenarios. 4 Discussion This study was based on the data of 771 cases of pediatric kidney disease admitted to the Affiliated Hospital of Zunyi Medical University from 2009 to 2023. It integrated demographic information, laboratory tests, immune metabolic indicators and renal pathological characteristics, and attempted for the first time to use multiple ensemble learning algorithms to construct a risk prediction model for the onset of primary nephrotic syndrome (NS) in children. In the process of data preprocessing, multiple imputation, standardization and unique hot encoding strategies were used to improve data quality. Subsequently, four mainstream machine learning methods, GBDT, XGBoost, random forest (RF) and LightGBM, were used to establish a classification model, and its performance was systematically evaluated by cross-validation and independent validation sets. The results showed that the random forest model performed best in terms of accuracy, precision, recall, F1 score and AUC, and had good generalization ability and robustness. At the same time, the SHAP method was used to visualize the model features, further verifying that the prediction basis of the model was consistent with the clinical mechanism of NS. This study not only achieved the construction of a high-performance automated risk prediction model, but also provided a practical data-driven tool for the early identification and precise intervention of NS in clinical practice. This article systematically analyzed clinical indicators such as plasma IgG, antistreptolysin O (ASO), complement C3 levels, and gross hematuria during the onset of PNS in children, revealing the dynamic changes of these indicators during the disease progression and their clinical significance [ 9 ] . Studies have found that immune system disorders play a key role in the occurrence and development of PNS. First, the significant decrease in immunoglobulin levels, especially IgG levels, reflects the abnormality of B cell function and humoral immune status. Studies have shown that the decrease in IgG levels may be closely related to B cell production disorders and impaired humoral immune function. This change may be associated with the pathological characteristics of massive proteinuria and hypoproteinemia in the course of the disease. Secondly, the dynamic changes of the complement system, especially C3 levels, have also become an important indicator for evaluating immune function. Studies have shown that C3 levels show a downward trend [ 10 ] . This phenomenon may be related to the deposition of immune complexes in the glomeruli and changes in the permeability of the filtration barrier. In addition, the abnormal fluctuation of ASO levels has also attracted the attention of researchers. The increase of this indicator may reflect the increased risk of infection in children, and it has important clinical value as an evaluation indicator of antibody levels. From the perspective of renal pathophysiology, children with PNS often show a series of significant renal structural and functional abnormalities. Abnormally elevated blood pressure levels may be due to the combined effects of systemic hypertension or hypertension within the kidney [ 11 ] ; while increased urine red blood cell content may reflect impaired glomerular filtration function. Macroscopic hematuria is a typical clinical manifestation of PNS, and its mechanism of occurrence may be related to the deposition of immune complexes in the glomeruli. Changes in the plasma protein profile, including decreased levels of total protein and prealbumin, not only reveal disease-related protein loss [ 12 ] , but also suggest that children may be malnourished [ 13 ] . These changes may be closely related to pathological features such as hypoproteinemia and hypercholesterolemia. In addition, the fluctuation of platelet count has also attracted the attention of researchers in children with PNS, and its potential pathological significance still needs further research and confirmation. These findings have important scientific value and clinical significance. First, by revealing the dynamic changes of immune system indicators and renal function indicators, this study provides a new pathophysiological perspective for the pathogenesis of PNS (primary nephrotic syndrome). In particular, based on the phenomena of significantly reduced levels of immunoglobulins (such as IgG), disordered activity of the complement system (such as C3), and increased titers of antistreptococcal hemolysin O (ASO), it is suggested that immune imbalance may play a key role in children with NS, which is highly consistent with the trend of the shift from the "kidney-derived hypothesis" to the "immune hypothesis" in recent years regarding the pathogenesis of NS. Secondly, the identification of these immune and metabolic-related indicators not only enhances the depth of understanding of the disease, but also provides theoretical support for the establishment of new diagnostic methods, personalized treatments, and efficacy prediction models. For example, risk stratification management combined with the dynamic levels of IgG and C3 may help to accurately formulate immunosuppressant use plans, avoid the side effects and waste of resources caused by generalized treatment, and thus achieve "precision medicine" in the true sense. At the same time, this study also further emphasized the importance of real-time dynamic monitoring of key biomarkers. This monitoring method can capture tiny signal changes in the development of the disease, providing a scientific basis for early identification of the disease, early warning of progression, and tracking of efficacy. It is particularly suitable for children, a special group with rapid disease progression and large individual differences. From a more macro perspective, the in-depth development of PNS-related research has not only broadened people's understanding of this complex pediatric disease, but also provided a feasible direction for establishing a predictive medical model for high-risk children. Through long-term tracking and intelligent analysis of the changing trajectories of these important biomarkers, clinical workers are expected to identify potential pathological trends in the subclinical stage, thereby intervening in advance, significantly improving the medium- and long-term prognosis of children, and ultimately promoting the pediatric kidney disease prevention and control system to move towards an integrated "prediction-prevention-intervention" direction. Although this study has achieved relatively positive results in data integration, model construction and clinical interpretation, there are still several limitations that need to be further optimized and overcome in subsequent studies. First, this study is a single-center retrospective study, and all the data used are from the Affiliated Hospital of Zunyi Medical University. There are certain geographical limitations and sample representativeness deviations, which may limit the generalization ability of the model in other regions or multi-center clinical scenarios. Secondly, although this study covers multidimensional variables in routine clinical examinations, it has not yet incorporated high-throughput molecular feature information such as genetics, metabolomics, and proteomics. The model is still based on "phenotypic data" and cannot fully reflect the molecular mechanism and individual biological differences of primary nephrotic syndrome in children. In addition, due to the lack of individual variables in some historical cases, although multiple interpolation methods were used to fill them, this method is essentially based on conditional inference between variables, which may still introduce information errors and affect the accuracy and interpretability of model predictions. Furthermore, although machine learning models are superior to traditional statistical methods in performance, their algorithm parameter settings and feature selection are somewhat subjective, and the performance differences of different models in different tasks may affect the repeatability of their clinical promotion. Finally, although the evaluation indicators in the model construction phase have covered multidimensional performance, they have not yet been prospectively verified in real clinical situations. Further verification is needed in larger samples, multiple time points, and real disease course tracking to ensure their stability and practical value. 5 Conclusion Based on real-world multidimensional clinical data, this study constructed a variety of machine learning models suitable for the risk prediction of primary nephrotic syndrome (NS) in children, and systematically evaluated the performance of algorithms such as GBDT, XGBoost, random forest and LightGBM. The results showed that the random forest model showed the best performance in terms of accuracy, precision, recall, F1 score and AUC, and had good generalization ability and clinical adaptability. Through the feature interpretability analysis of the SHAP method, the key variables on which the model depends highly fit the immunopathological mechanism and clinical diagnostic characteristics of NS, further enhancing the credibility and practical value of the model. Overall, this study provides a data-driven decision-making tool for the early identification and precise intervention of NS, which has important clinical translation potential. However, considering that the study still has limitations such as single sample, missing molecular data and lack of prospective verification, multi-center, multi-omics joint modeling and clinical embedded verification research should be carried out in the future [ 14 ] to further improve the stability, interpretability and wide applicability of the model and promote the development of pediatric kidney disease risk prediction system towards intelligence and individualization. Abbreviations Table 2 AGN acute glomerulonephritis ASO antistreptococcal hemolysin O AUC area under the curve C3 Complement component 3 GBDT Gradient Boosting Decision Tree HBsAb hepatitis B surface antibody HTN hypertension IgG Immunoglobulin G LGBM Light Gradient Boosting Machine MICE Multiple Imputation by Chained Equations NS nephrotic syndrome PNS primary nephrotic syndrome RF Random Forest SHAP Shapley Additive exPlanations TSH thyroid stimulating hormone XGBoost Extreme Gradient Boosting Ethics and Consent to Participate Declarations Ethics and Consent to Participate Ethical Approval: This study was approved by the Ethics Committee of the Affiliated Hospital of Zunyi Medical University (Approval No. KLL-2025-070). All procedures strictly adhered to the ethical principles outlined in the Declaration of Helsinki for research involving human subjects. Informed Consent: Written informed consent was obtained from all participants or their legal guardians during initial hospitalization, authorizing the use of clinical data for scientific research. Data Anonymization: All data generated during this study are included in this published article and its supplementary information files. To protect participant privacy, all personally identifiable information was removed prior to analysis, ensuring complete anonymization of the dataset. Consent for publication: Author consent for publication: The authors consent to the publication of the manuscript entitled "Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction" in the BMC Medical Informatics and Decision Making, including any supplementary materials. We futher agree that the publisher may exercise its rights to publish, distribute, and archive the manuscript in all forms and media, and that the work may be made available under an open-access license if selected or required. Patient consent for publication: Patients give their informed consent for the publication of clinical details, images, and/or data related to my case in "Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction". Patients understand that such information may be publicly accessible but confirm that their identity will be protected as far as possible. Patients have been informed of the purpose and potential implications of this publication and voluntarily consent without coercion. Availability of data and materials: All data generated or analysed during this study are included in this published article and its supplementary information files. Competing Interests: The authors declare that they have no competing interests as defined by BMC, or other interests that might be perceived to influence the results and/or discussion reported in this paper. Funding: Support By Key Advantageous Discipline Construction project of Guizhou Provincial Health Commission in 2023. Authors' contributions This manuscript reflects a collaborative effort, with each author contributing significantly to the research design, execution, analysis, and dissemination. The specific contributions are outlined below using author initials: Research Conceptualization & Clinical Expertise •Y.G. (Yi Gan) As a pediatric nephrology specialist and associate chief physician, Y.G. provided critical clinical insights into the study design, patient cohort selection, and pathological relevance. Guided the integration of clinical indicators (e.g., IgG, C3, ASO) with NS pathophysiology. Supervised ethical compliance and clinical validation of model outputs. Data Curation & Methodology •S.G. (Shiyou Guan) Led data collection from electronic medical records (2009–2023), including demographic, laboratory, and renal pathological variables. Designed feature selection strategies and implemented preprocessing pipelines (missing value imputation via MICE, standardization, one-hot encoding). Coordinated cross-validation protocols. Model Development & Algorithm Implementation •P.Y. (Ping Yuan) Spearheaded machine learning model construction using GBDT, XGBoost, RF, and LightGBM. Optimized hyperparameters (e.g., tree depth, learning rate, regularization) and implemented Bayesian optimization for fine-tuning. Developed the SHAP-based interpretability framework and visualized feature importance (Figure 2). Statistical Analysis & Validation •A.S. (Anxia Sun) Performed five-fold cross-validation, evaluated model performance metrics (accuracy, precision, recall, F1-score, AUC), and generated ROC curves (Figure 1). Conducted robustness assessments for outlier handling and data distribution analysis. Clinical Data Annotation & Preliminary Analysis •Y.H. (Youfang He) Annotated clinical records, extracted laboratory indicators (e.g., proteinuria, hematuria, hypertension), and assisted in pathological data collation. Supported SHAP interpretation by aligning model features with NS clinical manifestat Acknowledgments The authors acknowledge the financial support from the Key Advantageous Discipline Construction Project of Guizhou Provincial Health Commission (2023). This funding was instrumental in facilitating data collection, laboratory analyses, and computational resources required for the development of the machine learning prediction model. We also extend our appreciation to the medical staff at Affiliated Hospital of Zunyi Medical University for their assistance in clinical data curation. The authors confirm that the work described is original, has not been published previously, All necessary permissions have been obtained for the use of any copyrighted material, including figures, tables, or text experts from other sources, and appropriate acknowledgments have been included. References Yu Z, Xia Zhengkun. Research progress on biological treatment of primary and secondary nephrotic syndrome in children[J]. Chin J Gen Pract. 2014;17(27):17–21. Wang Shan W, Hong. Clinical analysis of 40 cases of primary nephrotic syndrome in children[J]. J Bengbu Med Coll. 2014;39(03):71–3. Li H. Clinical characteristics and treatment progress of primary nephrotic syndrome in children[J]. Chin J Contemp Pediatr. 2022;24(04):321–6. Johnson RA, Smith JD. The Role of Environmental Factors in the Development of Childhood Nephrotic Syndrome. Environ Health Perspect. 2021;129(3):31001–10. Green EA, Thompson RF. Clinical Management of Steroid-Resistant Nephrotic Syndrome in Children. Clin Pediatr. 2019;58(7):789–802. Brown KM, White PC. Immunological Markers in the Diagnosis of Childhood Nephrotic Syndrome. Pediatr Res. 2020;87(3):456–63. Chen Ming W. Application of machine learning in the diagnosis of nephrotic syndrome in children[J]. Chin J Lab Diagnosis. 2021;25(05):839–42. Smith JD, Johnson LR. Machine Learning Techniques for Predicting Nephrotic Syndrome in Pediatric Patients. J Pediatr Nephrol. 2021;29(2):123–35. Liu Tao Z, Min. Progress in immunotherapy for primary nephrotic syndrome in children[J]. Chin J Immunol. 2020;36(08):997–1001. Li H, Zhang W. Clinical Characteristics and Treatment Progress of Primary Nephrotic Syndrome in Children. J Contemp Pediatr. 2022;24(4):321–6. Shen Xiaoyu S, Sha Y, Lei, et al. Clinical analysis of the characteristics of dynamic blood pressure changes in children with primary nephrotic syndrome[J]. Diagnostics Theory Pract. 2022;21(05):613–8. Miller CA, Davis MB. Pathophysiology of Nephrotic Syndrome in Children: A Review. Pediatr Nephrol. 2017;32(4):587–601. Lee JH, Kim SY. Nutritional Management in Children with Nephrotic Syndrome. Nutr Rev. 2020;78(5):354–65. Nguyen TQ, Lee JH. Genetic Factors in the Development of Childhood Nephrotic Syndrome. Genet Sci. 2018;20(5):547–54. Additional Declarations No competing interests reported. Supplementary Files RawDataofMachineLearningforPredictiveDiagnosisofProteinuriaDiseasea.xlsx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers invited by journal 27 Nov, 2025 Editor invited by journal 07 Oct, 2025 Editor assigned by journal 01 Oct, 2025 Submission checks completed at journal 30 Sep, 2025 First submitted to journal 30 Sep, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7281320","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":551758534,"identity":"a4bcb84a-c75a-4b31-8ded-3fae4ede558b","order_by":0,"name":"Shiyou Guan","email":"","orcid":"","institution":"Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Shiyou","middleName":"","lastName":"Guan","suffix":""},{"id":551758535,"identity":"7e730d43-563c-435e-b7bc-4d63dcfad0f3","order_by":1,"name":"Yi Gan","email":"","orcid":"","institution":"Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yi","middleName":"","lastName":"Gan","suffix":""},{"id":551758536,"identity":"69915cb2-ad53-43c5-a446-a706f8a3ad49","order_by":2,"name":"Anxia Sun","email":"","orcid":"","institution":"Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Anxia","middleName":"","lastName":"Sun","suffix":""},{"id":551758537,"identity":"1994d144-b248-40ee-90f7-c6b8f396f67e","order_by":3,"name":"Youfang He","email":"","orcid":"","institution":"Hospital of Zunyi Medical University","correspondingAuthor":false,"prefix":"","firstName":"Youfang","middleName":"","lastName":"He","suffix":""},{"id":551758538,"identity":"c0cd9438-0bc2-4be3-93b5-dc2169dc7138","order_by":4,"name":"Ping Yuan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIie3RvQrCMBDA8SuCLkEdL1SiLyAcBJykvopB0NXRsaWQLj6A4Hs4t3ToK1R0MItzBwcHwY/JMRkF85vvzyUcgOf9pCAGoCnrdZLENO7Jein4tkwlum9qSkn1SveZyzCdUm1u1FLx0WhAiMQ4tiXnIpMDaqtkr/RlDQs5yW1JrXSIxFQaqowQcnVwTFBpXmhkrglviCTDwDHh50KHQHOB77dJJIe/dE/Zld8fTzarKmOaTSSsyagGaH3PQZbxj+Huff+7w6Dned4fewEoDUYv6mc+HAAAAABJRU5ErkJggg==","orcid":"","institution":"Hospital of Zunyi Medical University","correspondingAuthor":true,"prefix":"","firstName":"Ping","middleName":"","lastName":"Yuan","suffix":""}],"badges":[],"createdAt":"2025-08-03 05:08:12","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7281320/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7281320/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":97553833,"identity":"ff94c796-d420-420c-8df6-00d5bd9094b3","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":364977,"visible":true,"origin":"","legend":"","description":"","filename":"InterpretableMLModelforPediatricPNSRiskPrediction.docx","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/91d701f10bfa14edfda8cbca.docx"},{"id":97673280,"identity":"f7270159-8e50-45e2-a6a2-4320e255a7f8","added_by":"auto","created_at":"2025-12-08 09:39:47","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":9170,"visible":true,"origin":"","legend":"","description":"","filename":"67cfe1f63c094335a8fcfb099e012717.json","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/1b684740f4e6290f6f6f52f1.json"},{"id":97672537,"identity":"c0113bfb-e14d-4af3-815a-ea8b4685b58d","added_by":"auto","created_at":"2025-12-08 09:38:15","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":156750,"visible":true,"origin":"","legend":"","description":"","filename":"RawDataofMachineLearningforPredictiveDiagnosisofProteinuriaDiseasea.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/2c8d00dc207a81af8800299c.xlsx"},{"id":97553838,"identity":"12e5355a-5f01-43f1-a49f-dd4a33d9dac2","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":72354,"visible":true,"origin":"","legend":"","description":"","filename":"67cfe1f63c094335a8fcfb099e0127171enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/bfd1b9ed5686ccd883d5bcd7.xml"},{"id":97553830,"identity":"9c1d5f48-d3ce-45fc-9784-d0705aff3106","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40835,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/63c97b43a4f6ded98d8b76c2.png"},{"id":97553836,"identity":"fd8bfbe1-def0-4f53-8716-34679a69d8ef","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44430,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/86a6ba11fe148957801b8979.png"},{"id":97670689,"identity":"7d93e295-1bb8-487e-8315-8012132e98f5","added_by":"auto","created_at":"2025-12-08 09:31:11","extension":"xml","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":65473,"visible":true,"origin":"","legend":"","description":"","filename":"67cfe1f63c094335a8fcfb099e0127171structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/44d765b7d3b833908c5484df.xml"},{"id":97553839,"identity":"0de94fbf-5ee3-4b64-868c-a6e752cee329","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"html","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":77315,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/2a89a5c2975024f01759276f.html"},{"id":97553831,"identity":"54aa841f-f6ec-4645-a3c7-3342defe953e","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":142042,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves and AUC values of each model, where A-D are the results of GBDT, LGBM, RF, and XGB models, respectively.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/b70e4b7006436812e2f78b19.png"},{"id":97553835,"identity":"c4a932d2-5ed6-407e-bae2-2cc3a2b23311","added_by":"auto","created_at":"2025-12-05 18:01:56","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":178217,"visible":true,"origin":"","legend":"\u003cp\u003eSHAP analysis visualization\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/c693d436a6e28bef3fb1a65f.png"},{"id":97678642,"identity":"2f3687fa-204c-40af-8849-18e3166dc941","added_by":"auto","created_at":"2025-12-08 09:55:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":713808,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/3b8222cd-9d7f-47da-8fc5-9dfafaa3d40e.pdf"},{"id":97553840,"identity":"19c55a63-7f58-412c-9dbd-c07d9be6bafc","added_by":"auto","created_at":"2025-12-05 18:01:59","extension":"xlsx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":156750,"visible":true,"origin":"","legend":"","description":"","filename":"RawDataofMachineLearningforPredictiveDiagnosisofProteinuriaDiseasea.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7281320/v1/7c14fc4e5439d27081b5dc4e.xlsx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003ePrimary nephrotic syndrome (NS) in children is one of the most common chronic kidney diseases in children, with an annual incidence of approximately 2\u0026ndash;7/100,000, which seriously affects the quality of life and growth and development of children[1]. The disease is characterized by massive proteinuria, hypoalbuminemia, significant edema and hyperlipidemia. The course of the disease tends to relapse, and some children may progress to hormone-dependent or hormone-resistant nephropathy, which increases the difficulty of treatment and the risk of long-term renal insufficiency[2]. Although a variety of immunosuppressants and supportive therapies are currently used in the treatment of NS, its complex etiology and highly heterogeneous clinical manifestations still pose great challenges to early diagnosis and intervention.\u003c/p\u003e\u003cp\u003eIn recent years, researchers have gradually realized that the occurrence and development of NS may involve the combined effects of multiple factors such as genetics, immunity, metabolism, and environment [3]. Some studies suggest that certain immune indicators (such as IgG, complement C3), inflammatory factors, endocrine abnormalities, and renal tissue pathological changes all play an important role in the pathogenesis of NS. This complex etiology determines that the manifestations of NS in different children vary greatly. Therefore, the traditional \"empirical judgment\" diagnostic model can no longer meet the needs of precision medicine. For the early identification and risk stratification of NS, more systematic, efficient, and intelligent analysis tools are needed.\u003c/p\u003e\u003cp\u003eWith the widespread use of electronic medical records and hospital information systems, the acquisition of clinical data has become increasingly convenient. The multi-dimensional data of pediatric kidney disease patients (such as demographic information, biochemical indicators, immune parameters, pathological data, etc.) provide a solid data foundation for disease prediction modeling. How to accurately extract characteristic variables with diagnostic value from complex clinical data and construct a risk prediction model with generalization ability has become a hot topic in current research. Traditional statistical methods have limitations in dealing with nonlinearity, multivariate interactions, and high-dimensional features, which has also opened up new breakthroughs for the application of artificial intelligence technology, especially machine learning methods.\u003c/p\u003e\u003cp\u003eAs a core branch of artificial intelligence, machine learning has powerful data modeling and feature learning capabilities. It can automatically identify patterns from historical data and predict future trends without explicit programming. In recent years, algorithms such as support vector machines, random forests, gradient boosting trees, and XGBoost have been widely used in medical prediction tasks such as disease risk assessment, diagnostic classification, and prognosis analysis, and have achieved good results. Compared with traditional methods such as linear regression, machine learning models can better handle nonlinear feature relationships, outliers, and missing information, and are more suitable for risk modeling needs in complex clinical contexts.\u003c/p\u003e\u003cp\u003eAgainst this background, this paper takes children with NS and acute glomerulonephritis admitted to the Affiliated Hospital of Zunyi Medical University between 2009 and 2023 as the research objects, integrates their multi-dimensional clinical, laboratory and pathological data, and attempts to use a variety of mainstream machine learning algorithms to construct a risk prediction model for the incidence of primary nephrotic syndrome in children, aiming to explore the feasibility and clinical value of this strategy in the early identification and precise intervention of NS in children.\u003c/p\u003e"},{"header":"2 Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 Patient Enrollment\u003c/h2\u003e\u003cp\u003eThis is a single-center retrospective study that included 771 children with primary kidney disease who were hospitalized in the Pediatric Nephrology Ward of the Affiliated Hospital of Zunyi Medical University from January 1, 2009 to December 31, 2023. Inclusion criteria included: (1) age 1\u0026ndash;18 years; (2) primary kidney disease confirmed by renal puncture biopsy pathology based on clinical manifestations and laboratory test results; (3) complete clinical data, including demographic characteristics, laboratory test indicators, and renal histopathological examination results. According to the final diagnosis, the children were divided into a primary nephrotic syndrome (NS) group (n\u0026thinsp;=\u0026thinsp;376) and an acute glomerulonephritis (AGN) group (n\u0026thinsp;=\u0026thinsp;395) for subsequent model construction and comparative analysis.\u003c/p\u003e\u003cp\u003eThe basic information, laboratory tests and pathology reports of the children were all obtained from the hospital's electronic medical record system. To ensure the authenticity and consistency of the data, two independent researchers extracted and reviewed the data according to unified standards. For controversial case data, a third senior attending physician was invited to conduct the final review.\u003c/p\u003e\u003cp\u003e This study strictly adhered to the ethical principles of the Declaration of Helsinki on research involving human subjects and was approved by the Ethics Committee of the Affiliated Hospital of Zunyi Medical University (Ethics Approval No.KLL-2025-070). All included cases signed informed consent during their initial hospitalization, or their legal guardians signed on their behalf, allowing their clinical data to be used for scientific research analysis. To protect patient privacy, all data were removed from personally identifiable information during use to ensure anonymization of the information.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 Data Collection and Feature Selection\u003c/h2\u003e\u003cp\u003eThis study collected multi-dimensional clinical information and laboratory test data from children, striving to fully cover potential risk factors associated with primary kidney disease. All data were retrieved and sorted through the hospital's electronic medical record system to ensure the integrity and consistency of the information. The main characteristic information collected includes the following aspects:\u003c/p\u003e\u003cp\u003eGeneral demographic characteristics mainly include sociodemographic parameters such as the child's age, gender, place of residence (urban/rural), and education level [4]. These factors can reflect the potential association between social environment, lifestyle and disease risk to a certain extent, and are of great reference value in epidemiological studies of childhood kidney disease. Secondly, laboratory test indicators are divided according to physiological systems, mainly including: (1) Hematological indicators: platelet count, fibrinogen level, to evaluate blood coagulation status and potential blood system abnormalities; (2) Immune function indicators: IgG level in urine and plasma, complement C3 level, hepatitis B surface antibody (HBsAb) status, antistreptococcal hemolysin O (ASO) titer, etc., reflecting the body's immune status and history of infection. These indicators are closely related to the immune pathogenesis of NS [4\u0026ndash;5]; (3) Endocrine function indicators: thyroid stimulating hormone (TSH ) levels, which are used to assess the thyroid function status of children and indirectly reflect the metabolic regulation status; (4) biochemical metabolic indicators: including total protein, plasma albumin, prealbumin, urea nitrogen, blood creatinine, serum cystatin C, total cholesterol, vitamin D levels, etc., which systematically reflect the nutritional status, metabolic level and glomerular filtration function; (5) renal function indicators: such as urine protein quantitative, urine protein qualitative, hematuria (microscopic hematuria or macroscopic hematuria), hypertension, oliguria and edema. These indicators are important references for clinically assessing the severity of NS [6].\u003c/p\u003e\u003cp\u003eAll enrolled children underwent renal puncture biopsy, and the results of renal tissue pathology were also included in the scope of data collection. Pathological characteristics mainly include the degree of glomerular lesions (such as mesangial hyperplasia, capillary wall changes, etc.), the degree of tubular interstitial damage (such as interstitial fibrosis, inflammatory cell infiltration), etc. These pathological parameters are of great significance for distinguishing different types of primary nephrotic syndrome and their prognosis assessment. After the data is sorted, all characteristic variables enter the subsequent data preprocessing and modeling process.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 Data Preprocessing\u003c/h2\u003e\u003cp\u003eBefore model construction, in order to ensure data quality and algorithm performance, this study systematically preprocessed the original clinical data. First, in order to solve the problem of missing values in some variables, the chained multiple imputation method (Multiple Imputation by Chained Equations, MICE) was used to fill in the missing values in order to retain the sample information to the greatest extent and reduce data bias. Subsequently, outliers were detected for each continuous variable through box plots and variable distribution diagrams, and the source of outliers was determined by combining Z scores and IQR ranges. Obviously erroneous data were eliminated, and reasonable extreme values were retained and recorded for subsequent model robustness assessment. Considering the dimensional differences between the variables, in order to avoid affecting the convergence efficiency and learning effect of the model, all numerical variables were Z-score standardized to make their mean 0 and standard deviation 1, which was adapted to the input requirements of various machine learning algorithms. At the same time, categorical variables such as gender and place of residence were converted into binary vectors by one-hot encoding to achieve effective recognition and calculation of the algorithm. The above preprocessing steps provide a clean, structurally unified, high-quality data foundation for subsequent feature selection and modeling.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Model Construction\u003c/h2\u003e\u003cp\u003eBased on the preprocessed clinical data, this study used four mainstream ensemble learning algorithms to construct a risk prediction model for primary nephrotic syndrome in children, namely, Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), Random Forest (RF) and Light GBM (LGBM) \u003csup\u003e[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]\u003c/sup\u003e. In the process of building the GBDT model, the base learner was set as the CART regression tree, the initial learning rate was set to 0.1, the maximum tree depth (max_depth) was limited to 5 layers to prevent overfitting, and the number of weak learners (n_estimators) was initially set to 100, which was dynamically adjusted through cross-validation. As an enhanced boosting algorithm, XGBoost is further optimized in parameter settings, using undersampling ratio (subsample\u0026thinsp;=\u0026thinsp;0.8), column sampling ratio (colsample_bytree\u0026thinsp;=\u0026thinsp;0.8), L2 regularization coefficient (lambda\u0026thinsp;=\u0026thinsp;1), maximum tree depth max_depth set to 6, and learning rate set to 0.05 to improve generalization ability and prevent overfitting; at the same time, the early stopping mechanism (early_stopping_rounds\u0026thinsp;=\u0026thinsp;30) is used to control the number of training rounds to avoid redundant calculations. The random forest model uses a decision tree ensemble based on the bagging strategy, setting the number of decision trees (n_estimators) to 100, the maximum number of features (max_features) to the square root strategy (sqrt), and no limit on tree depth (max_depth\u0026thinsp;=\u0026thinsp;None). By increasing the number of trees, the variance is reduced and the model stability is improved. The LightGBM model was constructed using a leaf-wise intelligent splitting strategy, with the initial learning rate set to 0.05, the maximum tree depth set to 7, the subsample ratio set to 0.8, and the minimum number of leaf nodes (min_data_in_leaf) set to 20 to improve computational efficiency and accuracy. To ensure optimal model performance, this study used a five-fold cross-validation (k\u0026thinsp;=\u0026thinsp;5) strategy to evaluate the performance of different models on the training set and validation set, and further fine-tuned the parameters through grid search combined with Bayesian optimization \u003csup\u003e[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]\u003c/sup\u003e .\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5 Model Evaluation and Validation\u003c/h2\u003e\u003cp\u003eIn order to comprehensively evaluate the performance of various machine learning models in the task of predicting the risk of primary nephrotic syndrome in children, this study used a variety of classification performance indicators for quantitative analysis. The main evaluation indicators include accuracy, recall, precision, F1-score, and area under the receiver operating characteristic curve (AUC). Among them, accuracy reflects the proportion of correct classification of the model in all predictions, which is used to measure the overall classification ability; recall represents the proportion of all true positive samples successfully detected by the model, focusing on evaluating the model's recognition sensitivity for sick children; precision refers to the proportion of true positive samples in all predicted positive samples, which is used to measure the reliability of the model's prediction results; F1 score is the harmonic mean of precision and recall, taking into account the accuracy and comprehensiveness of the model; AUC value reflects the overall ability of the model to distinguish between positive and negative samples by drawing the receiver operating characteristic curve (ROC curve). The closer the AUC value is to 1, the better the model's identification performance.\u003c/p\u003e\u003c/div\u003e"},{"header":"3 Results","content":"\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Model Comparison\u003c/h2\u003e\u003cdiv id=\"Sec10\" class=\"Section3\"\u003e\u003ch2\u003e3.1.1 Model Evaluation Metrics\u003c/h2\u003e\u003cp\u003eBased on five-fold cross validation, this study systematically evaluated the performance of four machine learning models, GBDT, LightGBM, random forest (RF) and XGBoost, in the task of predicting the risk of primary nephrotic syndrome in children. The model evaluation indicators include accuracy, precision, recall and F1-Score. The specific comparison results are as follows: the GBDT model has an accuracy of 0.9828, a precision of 0.9827, a recall of 0.983, and an F1-Score of 0.9828; the LightGBM model performs slightly better, with an accuracy of 0.9871, a precision of 0.9872, a recall of 0.9869, and an F1-Score of 0.9871; the Random Forest model performs best in all indicators, with an accuracy of 0.9914, a precision of 0.9913, a recall of 0.9916, and an F1-Score of 0.9914; the XGBoost model has similar indicators to GBDT, with an accuracy of 0.9828, a precision of 0.9827, a recall of 0.9827, and an F1-Score of 0.9827. Comprehensive comparison shows that all four models show high prediction accuracy, but the random forest model is slightly higher than other models in accuracy, precision, recall and F1 score, showing better overall performance, becoming the preferred model for subsequent in-depth analysis and application discussion. The specific results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\u003ch2\u003e3.1.2 Model AUC Comparison\u003c/h2\u003e\u003cp\u003eTo further evaluate the ability of each model to distinguish the risk of primary nephrotic syndrome in children, this study plotted the receiver operating characteristic (ROC) curves of four machine learning models and calculated the area under the curve (AUC) value (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The results showed that all four models showed extremely high discrimination ability, and the AUC values were close to 1, indicating that the performance of the models in positive and negative sample recognition was very stable. Among them, the LightGBM model had the highest AUC value of 0.9992, followed by the GBDT model (AUC\u0026thinsp;=\u0026thinsp;0.9990), the random forest model was slightly lower (AUC\u0026thinsp;=\u0026thinsp;0.9983), and the XGBoost model AUC was 0.9982. The ROC curve shapes were significantly better than the random guessing reference line (diagonal line), and maintained a very high true positive recognition rate under low false alarm rates, further verifying that each model had excellent discrimination performance in this research task. In general, although the differences in AUC among the four models are small, the LightGBM and GBDT models are slightly better in that the ROC curve is closer to the upper left corner, suggesting that they have stronger sensitivity and applicability in clinical risk screening.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance evaluation results of different machine learning algorithms\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eF1-Score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGBDT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9828\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.9827\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.983\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9828\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLGBM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9871\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.9872\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9869\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9871\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9914\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.9913\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9914\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eXGB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.9828\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.9827\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.9827\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.9827\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e3.2 SHAP analysis visualization\u003c/h2\u003e\u003cp\u003eThis study conducted a feature importance analysis on the final model based on the SHAP (SHapley Additive exPlanations) method. As a feature attribution method based on game theory, SHAP can quantify the positive and negative impact direction of each feature on a single prediction result and its contribution, thereby providing an intuitive explanation for the internal mechanism of the \"black box model\". Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows the feature importance ranking and distribution based on SHAP values. It can be seen that plasma IgG is the variable that most significantly affects the output of the model, and its SHAP value is significantly negatively distributed, indicating that the reduced IgG level has a higher weight in the model prediction of NS (Nephrotic Syndrome), which is consistent with the immune disorder mechanism of massive IgG loss in children with NS. Secondly, microscopic hematuria and total protein levels also showed strong predictive ability, among which hypoproteinemia, as a classic feature of NS, is closely related to model decision-making.\u003c/p\u003e\u003cp\u003eIn addition, hypertension (HTN), antistreptococcal hemolysin O (ASO) titer, gross hematuria, and platelet count also played a significant role in the model. The significance of immune indicators such as ASO and C3 further supports the immune-related pathogenesis of NS. The SHAP values of renal function and metabolism-related indicators such as plasma albumin, complement C3, and urea nitrogen also show that they have a moderate contribution to the model output. It is worth noting that some non-traditional core features, such as variables such as hospital stay days, weight, and onset, also affect the model judgment to a certain extent. This may be related to the course of the disease, the severity of the disease, and treatment compliance, suggesting that the model not only captures the direct manifestations of the disease, but also takes into account information at the clinical management level.\u003c/p\u003e\u003cp\u003eOverall, the results of SHAP analysis show that the model construction not only has high predictive ability, but also shows good clinical interpretability. Most of the key variables it relies on are consistent with the pathophysiological mechanism, clinical manifestations and diagnostic criteria of NS, providing a trust basis for the application of the model in actual medical scenarios.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"4 Discussion","content":"\u003cp\u003eThis study was based on the data of 771 cases of pediatric kidney disease admitted to the Affiliated Hospital of Zunyi Medical University from 2009 to 2023. It integrated demographic information, laboratory tests, immune metabolic indicators and renal pathological characteristics, and attempted for the first time to use multiple ensemble learning algorithms to construct a risk prediction model for the onset of primary nephrotic syndrome (NS) in children. In the process of data preprocessing, multiple imputation, standardization and unique hot encoding strategies were used to improve data quality. Subsequently, four mainstream machine learning methods, GBDT, XGBoost, random forest (RF) and LightGBM, were used to establish a classification model, and its performance was systematically evaluated by cross-validation and independent validation sets. The results showed that the random forest model performed best in terms of accuracy, precision, recall, F1 score and AUC, and had good generalization ability and robustness. At the same time, the SHAP method was used to visualize the model features, further verifying that the prediction basis of the model was consistent with the clinical mechanism of NS. This study not only achieved the construction of a high-performance automated risk prediction model, but also provided a practical data-driven tool for the early identification and precise intervention of NS in clinical practice.\u003c/p\u003e\u003cp\u003eThis article systematically analyzed clinical indicators such as plasma IgG, antistreptolysin O (ASO), complement C3 levels, and gross hematuria during the onset of PNS in children, revealing the dynamic changes of these indicators during the disease progression and their clinical significance \u003csup\u003e[\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]\u003c/sup\u003e. Studies have found that immune system disorders play a key role in the occurrence and development of PNS. First, the significant decrease in immunoglobulin levels, especially IgG levels, reflects the abnormality of B cell function and humoral immune status. Studies have shown that the decrease in IgG levels may be closely related to B cell production disorders and impaired humoral immune function. This change may be associated with the pathological characteristics of massive proteinuria and hypoproteinemia in the course of the disease. Secondly, the dynamic changes of the complement system, especially C3 levels, have also become an important indicator for evaluating immune function. Studies have shown that C3 levels show a downward trend \u003csup\u003e[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]\u003c/sup\u003e. This phenomenon may be related to the deposition of immune complexes in the glomeruli and changes in the permeability of the filtration barrier. In addition, the abnormal fluctuation of ASO levels has also attracted the attention of researchers. The increase of this indicator may reflect the increased risk of infection in children, and it has important clinical value as an evaluation indicator of antibody levels.\u003c/p\u003e\u003cp\u003eFrom the perspective of renal pathophysiology, children with PNS often show a series of significant renal structural and functional abnormalities. Abnormally elevated blood pressure levels may be due to the combined effects of systemic hypertension or hypertension within the kidney \u003csup\u003e[\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]\u003c/sup\u003e ; while increased urine red blood cell content may reflect impaired glomerular filtration function. Macroscopic hematuria is a typical clinical manifestation of PNS, and its mechanism of occurrence may be related to the deposition of immune complexes in the glomeruli. Changes in the plasma protein profile, including decreased levels of total protein and prealbumin, not only reveal disease-related protein loss \u003csup\u003e[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]\u003c/sup\u003e, but also suggest that children may be malnourished \u003csup\u003e[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]\u003c/sup\u003e. These changes may be closely related to pathological features such as hypoproteinemia and hypercholesterolemia. In addition, the fluctuation of platelet count has also attracted the attention of researchers in children with PNS, and its potential pathological significance still needs further research and confirmation.\u003c/p\u003e\u003cp\u003eThese findings have important scientific value and clinical significance. First, by revealing the dynamic changes of immune system indicators and renal function indicators, this study provides a new pathophysiological perspective for the pathogenesis of PNS (primary nephrotic syndrome). In particular, based on the phenomena of significantly reduced levels of immunoglobulins (such as IgG), disordered activity of the complement system (such as C3), and increased titers of antistreptococcal hemolysin O (ASO), it is suggested that immune imbalance may play a key role in children with NS, which is highly consistent with the trend of the shift from the \"kidney-derived hypothesis\" to the \"immune hypothesis\" in recent years regarding the pathogenesis of NS. Secondly, the identification of these immune and metabolic-related indicators not only enhances the depth of understanding of the disease, but also provides theoretical support for the establishment of new diagnostic methods, personalized treatments, and efficacy prediction models. For example, risk stratification management combined with the dynamic levels of IgG and C3 may help to accurately formulate immunosuppressant use plans, avoid the side effects and waste of resources caused by generalized treatment, and thus achieve \"precision medicine\" in the true sense.\u003c/p\u003e\u003cp\u003eAt the same time, this study also further emphasized the importance of real-time dynamic monitoring of key biomarkers. This monitoring method can capture tiny signal changes in the development of the disease, providing a scientific basis for early identification of the disease, early warning of progression, and tracking of efficacy. It is particularly suitable for children, a special group with rapid disease progression and large individual differences. From a more macro perspective, the in-depth development of PNS-related research has not only broadened people's understanding of this complex pediatric disease, but also provided a feasible direction for establishing a predictive medical model for high-risk children. Through long-term tracking and intelligent analysis of the changing trajectories of these important biomarkers, clinical workers are expected to identify potential pathological trends in the subclinical stage, thereby intervening in advance, significantly improving the medium- and long-term prognosis of children, and ultimately promoting the pediatric kidney disease prevention and control system to move towards an integrated \"prediction-prevention-intervention\" direction.\u003c/p\u003e\u003cp\u003eAlthough this study has achieved relatively positive results in data integration, model construction and clinical interpretation, there are still several limitations that need to be further optimized and overcome in subsequent studies. First, this study is a single-center retrospective study, and all the data used are from the Affiliated Hospital of Zunyi Medical University. There are certain geographical limitations and sample representativeness deviations, which may limit the generalization ability of the model in other regions or multi-center clinical scenarios. Secondly, although this study covers multidimensional variables in routine clinical examinations, it has not yet incorporated high-throughput molecular feature information such as genetics, metabolomics, and proteomics. The model is still based on \"phenotypic data\" and cannot fully reflect the molecular mechanism and individual biological differences of primary nephrotic syndrome in children. In addition, due to the lack of individual variables in some historical cases, although multiple interpolation methods were used to fill them, this method is essentially based on conditional inference between variables, which may still introduce information errors and affect the accuracy and interpretability of model predictions. Furthermore, although machine learning models are superior to traditional statistical methods in performance, their algorithm parameter settings and feature selection are somewhat subjective, and the performance differences of different models in different tasks may affect the repeatability of their clinical promotion. Finally, although the evaluation indicators in the model construction phase have covered multidimensional performance, they have not yet been prospectively verified in real clinical situations. Further verification is needed in larger samples, multiple time points, and real disease course tracking to ensure their stability and practical value.\u003c/p\u003e"},{"header":"5 Conclusion","content":"\u003cp\u003eBased on real-world multidimensional clinical data, this study constructed a variety of machine learning models suitable for the risk prediction of primary nephrotic syndrome (NS) in children, and systematically evaluated the performance of algorithms such as GBDT, XGBoost, random forest and LightGBM. The results showed that the random forest model showed the best performance in terms of accuracy, precision, recall, F1 score and AUC, and had good generalization ability and clinical adaptability. Through the feature interpretability analysis of the SHAP method, the key variables on which the model depends highly fit the immunopathological mechanism and clinical diagnostic characteristics of NS, further enhancing the credibility and practical value of the model. Overall, this study provides a data-driven decision-making tool for the early identification and precise intervention of NS, which has important clinical translation potential. However, considering that the study still has limitations such as single sample, missing molecular data and lack of prospective verification, multi-center, multi-omics joint modeling and clinical embedded verification research should be carried out in the future \u003csup\u003e[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]\u003c/sup\u003e to further improve the stability, interpretability and wide applicability of the model and promote the development of pediatric kidney disease risk prediction system towards intelligence and individualization.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"2\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAGN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eacute glomerulonephritis\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eASO\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eantistreptococcal hemolysin O\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003earea under the curve\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eC3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eComplement component 3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGBDT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eGradient Boosting Decision Tree\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHBsAb\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ehepatitis B surface antibody\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eHTN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ehypertension\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIgG\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eImmunoglobulin G\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLGBM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLight Gradient Boosting Machine\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMICE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMultiple Imputation by Chained Equations\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNS\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003enephrotic syndrome\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePNS\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eprimary nephrotic syndrome\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRandom Forest\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSHAP\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eShapley Additive exPlanations\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTSH\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ethyroid stimulating hormone\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eExtreme Gradient Boosting\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"2\"\u003eEthics and Consent to Participate\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003eEthics and Consent to Participate\u003c/p\u003e\n\u003cp\u003eEthical Approval:\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of the Affiliated Hospital of Zunyi Medical University (Approval No. KLL-2025-070). All procedures strictly adhered to the ethical principles outlined in the Declaration of Helsinki for research involving human subjects.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; Informed Consent:\u003c/p\u003e\n\u003cp\u003eWritten informed consent was obtained from all participants or their legal guardians during initial hospitalization, authorizing the use of clinical data for scientific research.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; Data Anonymization:\u003c/p\u003e\n\u003cp\u003eAll data generated during this study are included in this published article and its supplementary information files. To protect participant privacy, all personally identifiable information was removed prior to analysis, ensuring complete anonymization of the dataset.\u003c/p\u003e\n\u003cp\u003eConsent for publication:\u003c/p\u003e\n\u003cp\u003eAuthor consent for publication:\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; The authors consent to the publication of the manuscript entitled \u0026nbsp;\u0026quot;Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction\u0026quot; in the BMC Medical Informatics and Decision Making, including any supplementary materials.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; We futher agree that the publisher may exercise its rights to publish, distribute, and archive the manuscript in all forms and media, and that the work may be made available under an open-access license if selected or required.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; Patient consent for publication:\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; Patients give their informed consent for the publication of clinical details, images, and/or data related to my case in \u0026quot;Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction\u0026quot;. Patients \u0026nbsp;understand that such information may be publicly accessible but confirm that their identity will be protected as far as possible. Patients have been informed of the purpose and potential implications of this publication and voluntarily consent without coercion.\u003c/p\u003e\n\u003cp\u003eAvailability of data and materials:\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; All data generated or analysed during this study are included in this published article and its supplementary information files.\u003c/p\u003e\n\u003cp\u003eCompeting Interests:\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; The authors declare that they have no competing interests as defined by BMC, or other interests that might be perceived to influence the results and/or discussion reported in this paper.\u003c/p\u003e\n\u003cp\u003eFunding:\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; Support By Key Advantageous Discipline Construction project of Guizhou Provincial Health Commission in 2023.\u003c/p\u003e\n\u003cp\u003eAuthors\u0026apos; contributions\u003c/p\u003e\n\u003cp\u003eThis\u0026nbsp;manuscript\u0026nbsp;reflects\u0026nbsp;a\u0026nbsp;collaborative\u0026nbsp;effort,\u0026nbsp;with\u0026nbsp;each\u0026nbsp;author\u0026nbsp; contributing\u0026nbsp;significantly\u0026nbsp;to\u0026nbsp;the\u0026nbsp;research\u0026nbsp;design,\u0026nbsp;execution,\u0026nbsp;analysis,\u0026nbsp;and\u0026nbsp;dissemination.\u0026nbsp;The\u0026nbsp; specific\u0026nbsp;contributions\u0026nbsp;are\u0026nbsp;outlined\u0026nbsp;below\u0026nbsp;using\u0026nbsp;author\u0026nbsp;initials:\u003cbr\u003e\u0026nbsp;Research Conceptualization \u0026amp; Clinical Expertise\u003cbr\u003e\u0026nbsp;\u0026bull;Y.G. (Yi Gan)\u003cbr\u003e\u0026nbsp;As a pediatric nephrology specialist and associate chief physician, Y.G. provided critical clinical insights into the study design, patient cohort selection, and pathological relevance. Guided the integration of clinical indicators (e.g., IgG, C3, ASO) with NS pathophysiology. Supervised ethical compliance and clinical validation of model outputs.\u0026nbsp;\u003cbr\u003e\u0026nbsp;Data Curation \u0026amp; Methodology\u003cbr\u003e\u0026nbsp;\u0026bull;S.G. (Shiyou Guan)\u003cbr\u003e\u0026nbsp;Led data collection from electronic medical records (2009\u0026ndash;2023), including demographic, laboratory, and renal pathological variables. Designed feature selection strategies and implemented preprocessing pipelines (missing value imputation via MICE, standardization, one-hot encoding). Coordinated cross-validation protocols.\u0026nbsp;\u003cbr\u003e\u0026nbsp;Model Development \u0026amp; Algorithm Implementation\u003cbr\u003e\u0026nbsp;\u0026bull;P.Y. (Ping Yuan)\u003cbr\u003e\u0026nbsp;Spearheaded machine learning model construction using GBDT, XGBoost, RF, and LightGBM. Optimized hyperparameters (e.g., tree depth, learning rate, regularization) and implemented Bayesian optimization for fine-tuning. Developed the SHAP-based interpretability framework and visualized feature importance (Figure 2).\u0026nbsp;\u003cbr\u003e\u0026nbsp;Statistical Analysis \u0026amp; Validation\u003cbr\u003e\u0026nbsp;\u0026bull;A.S. (Anxia Sun)\u003cbr\u003e\u0026nbsp;Performed five-fold cross-validation, evaluated model performance metrics (accuracy, precision, recall, F1-score, AUC), and generated ROC curves (Figure 1). Conducted robustness assessments for outlier handling and data distribution analysis.\u0026nbsp;\u003cbr\u003e\u0026nbsp;Clinical Data Annotation \u0026amp; Preliminary Analysis\u003cbr\u003e\u0026nbsp;\u0026bull;Y.H. (Youfang He)\u003cbr\u003e\u0026nbsp;Annotated clinical records, extracted laboratory indicators (e.g., proteinuria, hematuria, hypertension), and assisted in pathological data collation. Supported SHAP interpretation by aligning model features with NS clinical manifestat\u003c/p\u003e\n\u003cp\u003eAcknowledgments\u003c/p\u003e\n\u003cp\u003eThe authors acknowledge the financial support from the Key Advantageous Discipline Construction Project of Guizhou Provincial Health Commission (2023). This funding was instrumental in facilitating data collection, laboratory analyses, and computational resources required for the development of the machine learning prediction model. We also extend our appreciation to the medical staff at Affiliated Hospital of Zunyi Medical University for their assistance in clinical data curation.\u003c/p\u003e\n\u003cp\u003eThe authors confirm that the work described is original, has not been published previously, All necessary permissions have been obtained for the use of any copyrighted material, including figures, tables, or text experts from other sources, and appropriate acknowledgments have been included.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eYu Z, Xia Zhengkun. Research progress on biological treatment of primary and secondary nephrotic syndrome in children[J]. Chin J Gen Pract. 2014;17(27):17\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang Shan W, Hong. Clinical analysis of 40 cases of primary nephrotic syndrome in children[J]. J Bengbu Med Coll. 2014;39(03):71\u0026ndash;3.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi H. Clinical characteristics and treatment progress of primary nephrotic syndrome in children[J]. Chin J Contemp Pediatr. 2022;24(04):321\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJohnson RA, Smith JD. The Role of Environmental Factors in the Development of Childhood Nephrotic Syndrome. Environ Health Perspect. 2021;129(3):31001\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGreen EA, Thompson RF. Clinical Management of Steroid-Resistant Nephrotic Syndrome in Children. Clin Pediatr. 2019;58(7):789\u0026ndash;802.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBrown KM, White PC. Immunological Markers in the Diagnosis of Childhood Nephrotic Syndrome. Pediatr Res. 2020;87(3):456\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen Ming W. Application of machine learning in the diagnosis of nephrotic syndrome in children[J]. Chin J Lab Diagnosis. 2021;25(05):839\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSmith JD, Johnson LR. Machine Learning Techniques for Predicting Nephrotic Syndrome in Pediatric Patients. J Pediatr Nephrol. 2021;29(2):123\u0026ndash;35.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu Tao Z, Min. Progress in immunotherapy for primary nephrotic syndrome in children[J]. Chin J Immunol. 2020;36(08):997\u0026ndash;1001.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi H, Zhang W. Clinical Characteristics and Treatment Progress of Primary Nephrotic Syndrome in Children. J Contemp Pediatr. 2022;24(4):321\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShen Xiaoyu S, Sha Y, Lei, et al. Clinical analysis of the characteristics of dynamic blood pressure changes in children with primary nephrotic syndrome[J]. Diagnostics Theory Pract. 2022;21(05):613\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMiller CA, Davis MB. Pathophysiology of Nephrotic Syndrome in Children: A Review. Pediatr Nephrol. 2017;32(4):587\u0026ndash;601.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLee JH, Kim SY. Nutritional Management in Children with Nephrotic Syndrome. Nutr Rev. 2020;78(5):354\u0026ndash;65.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNguyen TQ, Lee JH. Genetic Factors in the Development of Childhood Nephrotic Syndrome. Genet Sci. 2018;20(5):547\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Primary nephrotic syndrome in children, machine learning, random forest, risk prediction, SHAP interpretation","lastPublishedDoi":"10.21203/rs.3.rs-7281320/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7281320/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e[Background] Primary nephrotic syndrome (NS) in children is a common chronic kidney disease in pediatrics, characterized by complex pathogenesis, heterogeneous clinical manifestations, and easy recurrence. Existing clinical diagnosis mainly relies on symptoms and laboratory tests, lacking efficient and accurate early risk prediction tools, which limits the implementation of early intervention and individualized management. With the development of artificial intelligence technology, the construction of machine learning prediction models based on multidimensional clinical data has provided new possibilities for the early identification and precise intervention of NS.\u003c/p\u003e\n\u003cp\u003e[Methods] This study retrospectively collected clinical data of 771 children with primary kidney diseases in the Pediatric Nephrology Ward of the Affiliated Hospital of Zunyi Medical University from 2009 to 2023, including 376 children with NS and 395 children with acute glomerulonephritis. The data were improved by preprocessing methods such as multiple imputation, standardization and coding, and general demographic characteristics, laboratory test indicators and renal pathological characteristics were screened as modeling variables. Four machine learning algorithms, GBDT, XGBoost, random forest (RF) and LightGBM, were used to construct a risk prediction model for the onset of the disease. The model performance was evaluated using five-fold cross validation, and the feature importance was explained by the SHAP method.\u003c/p\u003e\n\u003cp\u003e[Results] All four models showed high predictive ability, among which the random forest model performed best, reaching an accuracy of 99.14%, precision of 99.13%, recall of 99.16%, F1 score of 0.9914 and AUC value of 0.9983 on the validation set. SHAP analysis results showed that indicators such as plasma IgG, total protein, complement C3, and ASO titer contributed significantly to model prediction and were highly consistent with the clinical pathological mechanism of NS, verifying the reliability and clinical interpretability of the model.\u003c/p\u003e\n\u003cp\u003e[Conclusion] This study successfully constructed a risk prediction model for NS in children based on machine learning algorithms, which has high accuracy and good clinical interpretability, and provides strong data support for early screening and individualized treatment of NS. In the future, multi-center and multi-omics validation should be carried out to further improve the generalization ability and clinical application value of the model.\u003c/p\u003e","manuscriptTitle":"Interpretable Machine Learning Model for Pediatric Primary Nephrotic Syndrome Risk Prediction","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-05 18:01:51","doi":"10.21203/rs.3.rs-7281320/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewersInvited","content":"","date":"2025-11-27T12:18:35+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2025-10-07T04:16:45+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-01T05:27:49+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-09-30T16:07:20+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2025-09-30T15:15:00+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"a0915371-38eb-4afc-91b3-6b67039d41dc","owner":[],"postedDate":"December 5th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-12-05T18:01:51+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-05 18:01:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7281320","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7281320","identity":"rs-7281320","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}