Development of an Interpretable Machine Learning Model for Predicting Clavien-Dindo Grade ≥2 Complications after Unilateral Minimally Invasive Pyeloplasty in UPJO: A Retrospective Cohort Study
Related Experiment Videos
Haotian Pan1,2,3,4, Runwu Wang1,2,3,4, Mengnan Jiang1,2,3,4
1Department of Urology, Traditional Chinese Medicine Innovation Team, Laboratory for Targeted Delivery Application of Traditional Chinese Medicine, Ministry of Education Key Laboratory of Child Development and Disorders, Chongqing, P.R. China.
Background:
Minimally invasive pyeloplasty (MIP), encompassing both conventional laparoscopy and robot-assisted approaches, has become the primary treatment for pediatric ureteropelvic junction obstruction (UPJO). However, these procedures remain associated with considerable postoperative complications that affect surgical outcomes. This study employed the Clavien-Dindo classification system and interpretable machine learning (ML) algorithms to predict Clavien-Dindo grade ≥2 complications using strictly preoperative clinical and ultrasound parameters. Furthermore, an online calculator was developed to assist clinicians in real-time risk stratification.
Methods:
A retrospective analysis was conducted on 533 pediatric UPJO patients admitted between January 2020 and December 2024. Feature selection was performed using recursive feature elimination. Ten ML algorithms were developed and compared to identify the optimal predictive model.
Results:
A total of 533 children undergoing unilateral minimally invasive pyeloplasty were included, among whom 127 (23.8%) developed Clavien-Dindo grade ≥2 complications. Seven preoperative clinical and imaging features were prioritized for model construction: age, affected-to-unaffected kidney volume ratio, increased calyx diameters, blood neutrophil count, cystatin C, anteroposterior pelvic diameter, and white blood cell (WBC) count. Based on comparisons of the area under the receiver operating characteristic curve and decision curve analysis in training and testing sets, the light gradient boosting machine (LightGBM) model demonstrated superior performance. SHapley Additive exPlanations (SHAP) analysis enhanced interpretability, and an online calculator was developed to improve clinical applicability.
Conclusion:
The LightGBM model can effectively predict the occurrence of Clavien-Dindo grade ≥2 complications following unilateral MIP. It may assist clinicians in evaluating postoperative risks and developing personalized follow-up strategies for early intervention.
