[Construction and validation of a prognostic prediction model for children-onset initial steroid-resistant nephrotic syndrome].

Objective: To analyze the risk factors for poor prognosis in children with steroid-resistant nephrotic syndrome (SRNS) and to construct and validate a prognostic model. Methods: A retrospective cohort study was conducted. Clinical data of 456 children with SRNS who were initially diagnosed and hospitalized at the Children's Hospital of Chongqing Medical University from January 2009 to December 2024 was collected, including general information, laboratory and pathological indicators, gene types, treatment, and prognosis. Follow-up was conducted for more than 12 months. The endpoint event was defined as a decrease in the estimated glomerular filtration rate (eGFR) of more than 30% compared to the baseline for three consecutive times (with an interval of at least one month between each measurement). The patients were divided into the event group and the non-event group based on whether the endpoint event occurred. Independent sample t-test, Mann-Whitney U test, test χ2, or Fisher's exact probability test were used for comparison between groups. Multivariate logistic regression was used to rank the importance of variables to screen for characteristic variables. The top 6 ranked characteristic variables were used to construct four machine learning models using Python: extreme gradient boosting, random forest, support vector machine, and logistic regression. The area under the receiver operating characteristic curve (AUC), accuracy, specificity, sensitivity, and F1 score were used to evaluate the discrimination and performance of the models. Calibration curves and clinical decision curves were drawn to assess the prediction accuracy and clinical net benefit of the models. The Shapley Additive Explanations (SHAP) method was applied to analyze the contribution of features. Results: Among the 456 children, 306 were male and 150 were female. The age of onset was 4.1 (2.3, 8.0) years, and the follow-up time was 2.2 (1.0, 4.0) years. There were 195 cases (42.7%) in the event group and 261 cases (57.3%) in the non-event group. The time of occurrence of the endpoint event in the event group was 1.0 (0.5, 2.5) years. Univariate analysis showed that there were statistically significant differences between the two groups in initial serum creatinine, initial eGFR, eGFR change rate after 3 months of treatment, gender, calcineurin inhibitor (CNI) treatment response, gene variation, white blood cell count, absolute monocyte count, blood urea nitrogen, and urine red blood cells (all P<0.05). The top 6 characteristic variables, including the eGFR change rate after 3 months of treatment, genetic variation, baseline eGFR, CNI treatment responsiveness, gender and baseline serum creatinine (the AUC decrease percentages of 20.1%, 9.4%, 1.6%, 1.0%, 0.3% and 0.1% respectively after permutation test). Among the four machine learning models, the random forest model performed the best (test set AUC was 0.77 (95%CI 0.73-0.81), accuracy 73.9%, specificity 78.5%, and sensitivity (71.2%, and F1 score 68.9%). The Hosmer-Lemeshow test showed that the predicted risk and actual risk of the random forest model were in good agreement (χ2=7.72, P=0.461); the clinical decision curve showed that the random forest model performed best and had the highest net benefit within a certain threshold range (0-0.5). Through SHAP, it was determined that the eGFR change rate after 3 months of treatment, initial eGFR, gene variation, and CNI treatment response were important predictors of poor prognosis in SRNS (average SHAP absolute values were 0.18, 0.08, 0.07 and 0.02, respectively). Conclusions: A random forest model was constructed based on machine learning and explained using the SHAP method. The results showed that the eGFR change rate after 3 months of treatment, initial eGFR, gene variation, and CNI treatment response were important factors affecting prognosis.