Abstract
BACKGROUND AND AIM: Idiopathic nephrotic syndrome (INS) is a glomerular disorder characterized by proteinuria, hypoalbuminemia, and edema, and relapse remains a major clinical challenge. Early prediction of relapse risk may facilitate individualized treatment and follow-up. This study aimed to develop and compare the performance of logistic regression, random forest, and deep learning models for predicting relapse in adult patients with INS using baseline clinical and laboratory data. METHODS: We conducted a retrospective cohort study of 562 adult patients with idiopathic nephrotic syndrome treated between January 2022 and January 2024. The primary outcome was the first relapse within 12 months after baseline assessment. Baseline demographic characteristics, clinical history, laboratory parameters, and treatment-related variables were collected. The dataset was randomly divided into training (70%), validation (15%), and test (15%) sets. Missing data were imputed, continuous variables were standardized as appropriate, and SMOTE was applied to the training set only to address class imbalance. Three predictive models were developed: logistic regression, random forest, and a deep learning-based neural network. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, and F1-score. RESULTS: Among the three models, the deep learning model showed the best predictive performance, with AUCs of 0.908, 0.900, and 0.883 in the training, validation, and test sets, respectively. The logistic regression model showed intermediate performance, whereas random forest showed the lowest discriminatory ability. The most influential predictors of relapse included steroid resistance, nephrotic-range proteinuria at baseline, prior relapse history/frequency, elevated ESR, and immunosuppressant use. CONCLUSIONS: Deep learning demonstrated better predictive performance than logistic regression and random forest for predicting 12-month relapse in adult patients with idiopathic nephrotic syndrome. These findings suggest that machine learning-based models, particularly deep learning, may serve as useful tools for relapse risk stratification. External validation in larger independent cohorts is needed before clinical implementation.