ºÚÁÏÉçÇø

Abstract: Antenatal HN is the most common congenital anomaly, affecting up to 5% of pregnancies and identified as a dilatation of the urinary tract on prenatal imaging. While the majority of cases are physiologic and self-resolving, a subset necessitate surgical intervention. Currently, infants are monitored with serial ultrasounds and may undergo invasive and burdensome investigations, including catheterisation and radionuclide imaging, which contribute to patient distress and healthcare costs.

The authors sought to develop a non-invasive, image-based risk stratification tool using routinely acquired renal ultrasound images by leveraging a multi-phase machine learning (ML) pipeline—comprising algorithm development, a prospective silent trial, and a forthcoming clinical trial. Through this process, we have created the Hydronephrosis Severity Index (HSI): a deep learning model that outputs a continuous risk score indicating the likelihood of obstruction. HSI scores effectively stratified children into low-, medium-, and high-risk categories for surgical obstruction, achieving robust generalizability when externally validated at three major North American centers (Stanford, CHOP, and UIowa), with AUROCs ≥ 0.90 at all sites.

The authors' objective in this talk is to highlight:

1. The clinical motivations for developing the HSI.

2. The ML model design and training process, including preprocessing strategies and model architecture.

3. The validation results across diverse institutions.

4. Key ethical, logistical, and clinical considerations in preparing for real-world deployment, including integration into clinical workflows and potential to reduce invasive testing.

Mandy Rickard, NP, is a nurse practitioner at The Hospital for Sick Children (SickKids) in Toronto, Canada.  Her research interest is focused on hydronephrosis and the use of artificial intelligence (AI) models to predict the severity of hydronephrosis cases.

Dr Lauren Erdman is a PhD computer scientist and assistant professor at Cincinnati Children’s Hospital in the James M. Anderson Center for Health Systems Excellence. Her research is primarily focused on developing and applying machine learning methods to improve pediatric clinical management.