Predicting congenital anomalies of the kidney and urinary tract (CAKUT) in prenatal hydronephrosis via machine learning.

PURPOSE: Urinary tract dilation (UTD) classification provides reliable predictive value for congenital anomalies of the kidney and urinary tract (CAKUT) but does not clearly define when advanced imaging is required. This study aimed to identify which specific UTD parameters most effectively predict the presence of CAKUT using multiple machine learning (ML) algorithms. METHODS: This retrospective cohort study included infants diagnosed with prenatal hydronephrosis (HN). Data on demographics, ultrasound findings, and final diagnoses were collected. Multiple supervised ML algorithms were trained and compared in Azure Machine Learning Studio, and the three best-performing models-Random Forest (RF), Light Gradient Boosting Machine (LightGBM), and Extreme Gradient Boosting (XGBoost)-were selected for detailed analysis. The dataset was randomly divided into training (70%), testing (20%), and validation (10%) subsets, and each model was iterated 1000 times. Mean performance metrics were reported to enhance robustness. Model performance was evaluated by accuracy, precision, recall, F1 score, and area under the ROC curve (AUC). RESULTS: Among 206 kidney units analyzed, 78 (37.9%) were diagnosed with CAKUT. CAKUT occurrence was significantly higher in kidneys with central or ureteral dilation (p < 0.001 for both). Central dilation consistently ranked as the most influential predictor, followed by ureteral dilation, anterior–posterior (AP) diameter, and age at first ultrasound. All three algorithms demonstrated excellent discrimination (AUC: 0.915–0.922). CONCLUSIONS: Central and ureteral dilation were the strongest determinants of CAKUT across all algorithms. The comparable performance of RF, LightGBM, and XGBoost supports the robustness of these findings and may guide more selective imaging strategies in low-risk infants.