Interpretable machine learning model for predicting kidney failure among CAKUT children in multicenter large-scale...

Area of Science:

• Nephrology
• Pediatric Medicine
• Artificial Intelligence in Healthcare

Background:

• Congenital anomalies of the kidney and urinary tract (CAKUT) represent the primary cause of kidney failure in children.
• Predicting individual patient outcomes and disease progression in CAKUT remains a significant clinical challenge.

Purpose of the Study:

• To develop and validate POCC, a machine learning model designed for predicting kidney failure risk at multiple time points (1, 3, and 5 years) in pediatric patients diagnosed with CAKUT.
• To assess the performance of both a general and a specialized version of the POCC model across different healthcare settings.

Main Methods:

• A multicenter study utilizing data from 2249 pediatric patients with CAKUT.
• Development of two machine learning model versions: a general model and a specialized model incorporating congenital-hereditary features.
• External validation of the models in independent cohorts from pediatric and general hospitals, alongside real-world deployment validation.

Main Results:

• The general POCC model demonstrated high internal AUCs (0.93-0.99) and strong external validation AUCs (0.89-0.98 and 0.81-0.90).
• The specialized POCC model achieved excellent internal AUCs (0.93-0.99) and external AUCs (0.91-0.96) in pediatric hospital settings.
• Online deployment of POCC showed 90.7% accuracy in real-world clinical validation.

Conclusions:

• POCC is the first tool capable of multi-timepoint kidney failure risk prediction for diverse CAKUT subphenotypes.
• The validated POCC model shows significant potential to enhance personalized management strategies for children with CAKUT.
• Accurate risk stratification using POCC can support timely clinical decisions and improve patient outcomes in pediatric nephrology.