SRNS

SRNS predicts steroid resistance in patients with idiopathic nephrotic syndrome (INS) using machine-learning analysis of clinical variables to identify biomarkers and stratify risk.

Key Features:

• Data collection and variable selection: Dataset comprised 91 subjects and 87 clinical variables, and a variable selection framework was used to identify informative predictors for SRNS.
• Variable selection framework: Combined a penalized regression approach (MLR+TLP) to detect linear effects, a nonparametric screening method (MAC) to detect nonlinear marginal or joint effects, and a stepwise refinement that accounts for correlations among clinical variables.
• Model construction: An initial support vector machine (SVM) trained on 26 selected clinical variables achieved 95.2% leave-one-out cross-validation (LOO-CV) accuracy, and a reduced SVM using eight variables (erythrocyte sedimentation rate, urine occult blood, percentage of neutrophils, immunoglobulin A, cholesterol, vinculin autoantibody, aspartate aminotransferase, prolonged prothrombin time) achieved 92.8% LOO-CV accuracy and 94.0% accuracy on the validation cohort with 90.0% sensitivity and 96.7% specificity.
• Incorporation of prior medical information: Integrated prior medical knowledge to account for correlations among selected variables and enhance prediction reliability.
• Statistical validation: Employed rigorous statistical testing with reported p-values <0.005 across initial and validation phases to demonstrate robustness and reduced overfitting risk.

Scientific Applications:

• SRNS risk prediction: Predicts steroid resistance in INS patients to stratify risk of progression toward end-stage renal disease.
• Biomarker identification: Identifies and prioritizes clinical biomarkers, including vinculin autoantibody as a podocyte-specific marker linearly associated with steroid responsiveness.
• Treatment selection support: Provides objective evidence to support selection of treatment strategies for children with nonhereditary SRNS.

Methodology:

Analysis used a dataset of 91 subjects with 87 clinical variables; variable selection combined MLR+TLP penalized regression, MAC nonparametric screening, and stepwise refinement; models were built using support vector machines with leave-one-out cross-validation and validated on an independent cohort, and statistical testing reported p-values <0.005 while incorporating prior medical information to consider variable correlations.