Pediatric urology, among many other specialties, has witnessed significant advancements in efforts to integrate artificial intelligence (AI) into everyday clinical practice. In particular, machine learning (ML) applications have the potential to revolutionize shared decision-making and advance personalized medicine. These technologies hold great promise in modernizing how we diagnose, monitor, and manage several conditions, including hydronephrosis. This article chronicles the evolution of AI and ML applications in pediatric patients with hydronephrosis. We discuss key models, their contributions and challenges, and ethical considerations surrounding their use.

Our group generated one of the earliest reports exploring the use of ML for hydronephrosis in 2019.¹ We described using cloud-based analytics to predict the likelihood of surgical interventions for infants with hydronephrosis. This study leveraged Microsoft Azure Machine Learning Studio to analyze a large dataset of clinical variables, including Society for Fetal Urology grades and differential renal function. The model primarily employed decision jungles and neural network algorithms, achieving an impressive area under the curve (AUC) of 0.9. As early proof-of-concept work, we demonstrated the feasibility of employing AI to enhance decision-making processes and predict surgical probability. The delivery of information in real time during a clinic visit means clinicians could better counsel families and potentially reduce unnecessary procedures, thereby optimizing patient outcomes and resource utilization. However, a vital issue detected with this approach is the variability in managing these patients, highlighting problems with external validation and widespread acceptance by providers and families.

Since then, models for hydronephrosis have become more sophisticated with a move toward using images for their inputs instead of, or in addition to, clinical variables. This has the potential to improve predictions’ strength and address reproducibility issues due to variable human interpretation. Most of these models focus on kidney ultrasound images to extract features to make predictions. Maximizing the data obtained from ultrasounds is an attractive concept as kidney ultrasound images are abundantly available, and ultrasound is relatively inexpensive, noninvasive, free from radiation exposure, and the most common imaging modality we use. To date, several groups^2–4 have harnessed this technology to automatically dichotomize hydronephrosis severity into low or high grades or assign Society of Fetal Urology grades to ultrasound images with varying degrees of accuracy. While these efforts are exciting and represent the initial stepping stones of image analysis in pediatric urology, the clinical utility of these models remains to be proven. One major drawback is the inherent subjectivity in the initial interpretation of ultrasound images in assigning grades. For a model to be able to automatically assign a hydronephrosis grade, it must first be trained with images and the respective human-assigned grades, therefore introducing this subjectivity into the model.

Another theme that has become an area of interest in ML hydronephrosis research is decreasing the burden and invasiveness of monitoring. Many children with hydronephrosis are investigated with voiding cystourethrograms and nuclear renography to document the presence of vesicoureteral reflux or further assess obstruction. These tests require invasive, painful procedures such as catheterization and intravenous access, as well as incurring radiation exposure. While a subset of children with hydronephrosis would benefit from these tests, many may likely be able to safely avoid them, guided by diagnostic and predictive data from readily available clinical and ultrasound data. Wang et al⁵ sought to predict the presence of dilating vesicoureteral reflux from early ultrasound features using an optimal classification tree model and reporting an AUC of 0.81, representing good accuracy. Khondker et al⁶ explored the value of using clinical and ultrasound image features to predict the drainage time on nuclear scans to determine obstruction. A random forest classifier generated decision trees to predict drainage times classified as unlikely obstruction, intermediate risk, and high risk with an area under the receiver operator curve of 0.74. These models show great promise; once validated and deployed, they can significantly aid in reducing the number of children who are investigated with invasive tests once they have been deemed low risk with this technology.

The final theme in this research domain is arguably the fastest growing. It involves using ML to predict clinical outcomes in seconds, either from clinical variables or imaging, and is an essential step in the movement toward personalized medicine. Drysdale et al⁷ built a 2-stage model to predict the probability of recurrent obstruction (rUPJO) after pyeloplasty whereby patients were classified as having no obstruction or rUPJO, including an individualized time to rUPJO for each patient. This model achieved an area under the receiver operator curve of 0.86 and included a web-based application and freely available code. Weaver et al⁸ built a convolutional neural network from furosemide nuclear scan variables to determine which patients would develop a loss of function of > 5% within 6 months of an equivocal scan with an AUC of 0.67. Erdman et al⁹ used a custom-built 7-layer Siamese convolutional neural network to predict patients at the highest risk for obstruction and who could be safely discharged from care using only 2 ultrasound images as inputs with an AUC of 0.93 and multi-institutional external validation. These models have the potential to streamline care and aid in risk stratification for hydronephrosis patients, particularly those from remote or rural communities.

While these projects are exciting, they represent early efforts and only the beginning of the era of ML work within pediatric urology. As such, it is essential to consider some concerning implications as this work gains momentum. In addition to ensuring that accessible scientific and technological tools are employed appropriately, it is the responsibility of those conducting ML research that is patient-facing to ensure that it is fair and free from bias. There are endless opportunities for biases to enter ML research, including overfitting, automation, and selection, as well as a recently described bias of induced belief revision.¹⁰ In addition, we are responsible for ensuring model fairness—that is, ensuring that any produced models are generalizable and widely accessible. This means that representatives of the diverse patient population are included in all aspects of model training, testing, and validation. These steps are critical in ensuring that historical systemic issues, such as underrepresentation, are not perpetuated.

In conclusion, the integration of AI and ML in pediatric hydronephrosis represents a significant leap forward in the field of pediatric urology. These technologies offer promising avenues for improving patient care, from early predictive models to complex algorithms for risk stratification and personalized treatment. However, it is crucial to address the ethical implications and ensure continuous validation of these models to maintain their reliability and trustworthiness. As we seek progress, the collaborative efforts of clinicians, researchers, and technologists will be essential in harnessing the full potential of AI in urology (Figure).