Artificial intelligence (AI), particularly machine learning (ML) algorithms, is becoming an essential tool in predicting outcomes for pediatric hydronephrosis by identifying risk factors and patterns that support early intervention. This systematic review examined studies on the efficacy of ML models specifically designed for pediatric hydronephrosis prediction, focusing on the accuracy and types of models used. A systematic search was conducted on PubMed and Google Scholar following PRISMA guidelines, ultimately selecting 6 out of 32 studies that met inclusion criteria: focus on pediatric hydronephrosis, use of area under the receiver operating characteristics curve (AUC) for performance measurement, predictive rather than diagnostic analysis, and exclusion of other urological conditions like posterior urethral valves.
Among the selected studies, 7 ML models were evaluated, most using clinical data (5 models) and a few using imaging data, particularly ultrasound (2 models). Sample sizes in the studies varied widely, from 50 to 1,645 records. Decision trees were the most common model type used (3 models), followed by convolutional neural networks (CNN) and support vector machines (SVM). Clinical data models often included key variables such as patient age, sex, and ultrasound findings, with AUC values ranging from 0.67 to 0.96, indicating varying predictive accuracy. Models based on imaging data, specifically renal ultrasounds, yielded higher AUCs between 0.93 and 0.98, suggesting strong predictive potential.
This review underscores the variability in ML model performance, highlighting the need for further research to optimize these models in pediatric urology. Understanding the impact of sample size, data types, and model selection on performance could significantly enhance AI’s role in managing pediatric hydronephrosis and other urological conditions.
