Machine learning model for predicting severe infection in children with idiopathic nephrotic syndrome: multicenter retrospective study

Background Infection is a common complication of idiopathic nephrotic syndrome (INS), and early identification of severe infection can improve patient outcome. Methods This multicenter retrospective study developed and validated machine learning (ML) models that predict severe infection in children with INS. The derivation cohort ( n = 2357) consisted of INS patients at one institution, and was separated into a training set and testing set. The external validation set ( n = 372) consisted of INS patients from three other hospitals. Data were collected for 41 variables, and ten of them were then selected by univariate analysis and Least Absolute Shrinkage and Selection Operator (LASSO) regression. Ten ML models were compared, and the best one was identified using receiver operating characteristic (ROC) analysis and other methods. Results The incidence rate of severe infection was 6.8% in the derivation cohort. The Light Gradient Boosting Machine (LightGBM) model had the best predictive performance (accuracy: 0.843, precision: 0.843, recall: 0.842, F1: 0.843, sensitivity: 0.842, specificity: 0.844, AUROC:0.912, AUPRC:0.915). The ten predictors were C-reactive protein, hemoglobin, white blood cells, activated partial thromboplastin time, creatinine, high-density lipoprotein, corrected serum calcium, complement 3, and number of immunosuppressants, and incidence of SRNS. This model had an AUROC of 0.979 and AUPRC of 0.842 in the external validation cohort. Conclusion A LightGBM model for predicting severe infection in patients with INS had excellent performance. Future applications of this model may provide an effective, convenient, and cost-effective approach for early identification of severe infection in children with INS.