Machine learning prediction of childhood nephrotic syndrome outcomes

Children with steroid-resistant, frequently relapsing, and steroid-dependent nephrotic syndrome experience high disease and treatment-related morbidity. There are few prediction models for childhood nephrotic syndrome outcomes. Our aim was to develop and internally validate outcome prediction models, using machine learning methods.

We analyzed data from Insight into Nephrotic Syndrome: Investigating Genes, Health, and Therapeutics, a prospective observational childhood nephrotic syndrome cohort. We included children (1-18 years) diagnosed from 1996 to 2023 from the Greater Toronto Area, Canada, with baseline data from within 90 days of diagnosis. Outcomes were frequent relapses or steroid dependence by 1 year, relapse occurrence by 1 year, steroid-sparing medication initiation by 2 years, and initial steroid resistance. We developed and compared the performance of nine different machine learning algorithms for each outcome. Predictors included sociodemographic, clinical, and laboratory features.

We included 515 children diagnosed with nephrotic syndrome. Of these, 484 (94%) were steroid-sensitive and 31 (6%) were steroid-resistant. Nine machine learning models were developed and optimized by hyperparameter tuning for each outcome. The random forest model had the best performance for predicting frequent relapses or steroid dependence, although its predictive ability was low (AUC 0.60). Machine learning models also had weak predictive ability for relapse occurrence (AUC 0.62; logistic regression with recursive feature elimination), steroid-sparing medication initiation (AUC 0.61; XGBoost), and steroid resistance (AUC 0.64; XGBoost).

Routinely collected sociodemographic, clinical, and laboratory features at nephrotic syndrome diagnosis are weakly predictive of subsequent relapses and treatment response, using machine learning methods. Discovery of novel biomarkers may improve future prediction and proactive treatment.