Machine learning models for predicting steroid-resistant of nephrotic syndrome

Area of Science:

• Nephrology
• Medical Informatics
• Biostatistics

Background:

• Steroid-resistant nephrotic syndrome (SRNS) poses a significant risk for end-stage renal disease in children.
• Early identification of steroid responsiveness in idiopathic nephrotic syndrome (INS) is crucial due to poor SRNS outcomes.

Purpose of the Study:

• To develop a predictive model for identifying steroid-resistant nephrotic syndrome (SRNS) in patients with idiopathic nephrotic syndrome (INS).
• To identify informative clinical variables for predicting SRNS and improve early diagnosis and treatment strategies.

Main Methods:

• A novel variable selection framework combining penalized regression (MLR+TLP) for linear effects and a nonparametric method (MAC) for nonlinear effects was employed.
• A stepwise method was used to build the final prediction model, considering correlations between selected variables.
• Statistical testing was performed to assess model overfitting.

Main Results:

• An initial Support Vector Machine (SVM) model using 26 variables achieved 95.2% leave-one-out cross-validation (LOO-CV) accuracy.
• A reduced SVM model with eight key variables (including vinculin autoantibody) demonstrated high predictive performance (92.8% LOO-CV, 94.0% validation accuracy).
• The reduced model achieved 90.0% sensitivity and 96.7% specificity, with vinculin autoantibody being a key predictor.

Conclusions:

• The developed SRNS prediction model offers a comprehensive evaluation of patient conditions, aiding in treatment selection for nonhereditary SRNS.
• The model provides scientific guidance for managing children with nonhereditary SRNS.
• A user-friendly web tool is available for accessing the SRNS prediction model.