Prediction of pediatric lower urinary tract dysfunction using uroflowmetry data and machine learning with explainable AI techniques

Abstract

<title>Abstract</title> Purpose To develop and evaluate a machine learning based approach for the interpretation of uroflowmetry in pediatric lower urinary tract dysfunction, aiming to reduce interobserver variability and improve diagnostic consistency while maintaining clinical interpretability. Methods In this single-center retrospective study, 2663 pediatric UF records (age 2–18 years, July 2019-March 2024) were analyzed. LUTD subtypes were assigned to the International Children's Continence Society terminology. Two settings were evaluated: three-class (overactive bladder (OAB), dysfunctional voiding (DV), normal (N)) and four-class (adding DV-OAB). Eight predictors (four numerical, four categorical) were used. Seven ML models (Decision Trees, Naive Bayes, Support Vector Machine, Efficiently Trained Linear Classifiers (Logistic Regression), Gaussian Kernel, Ensemble Classifiers, and Artificial Neural Networks) were trained with Bayesian hyperparameter optimization using stratified 10-fold cross-validation on a 90% − 10% train-test split. Weighted accuracy, precision, recall, and F1-score were reported. Feature importance was quantified with Shapley Additive Explanations (SHAP). Reporting followed TRIPOD-AI. Results In the three-class setting, the ensemble classifier achieved the best test performance (accuracy 86.09%, F1 86.03%) and all models exceeded 82% test accuracy. In the four-class setting, test accuracy dropped to 72.56–78.95%, and most models failed to predict the DV-OAB class owing to its low prevalence (2.55%). SHAP identified bladder capacity as the dominant feature for OAB, pelvic floor activity for DV, and residual volume as a secondary DV feature. Sex contributed zero predictive value. Conclusions Explainable ML enables interpretable classification of pediatric LUTD subtypes from UF data across a wide age range, including children under 5 years. External multicenter validation is required before clinical deployment.