Background: The purpose of this study was to accurately predict pediatric choledocholithiasis with clinical data using a computational machine learning algorithm.
Methods: A multicenter retrospective cohort study was performed on children <18 years of age who underwent cholecystectomy between 2016 to 2019 at 10 pediatric institutions. Demographic data, clinical findings, laboratory, and ultrasound results were evaluated by bivariate analyses. An Extra-Trees machine learning algorithm using k-fold cross-validation was used to determine predictive factors for choledocholithiasis. Model performance was assessed using the area under the receiver operating characteristic curve on a validation dataset.
Results: A cohort of 1,597 patients was included, with an average age of 13.9 ± 3.2 years. Choledocholithiasis was confirmed in 301 patients (18.8%). Obesity was the most common comorbidity in all patients. Choledocholithiasis was associated with the finding of a common bile duct stone on ultrasound, increased common bile duct diameter, and higher serum concentrations of aspartate aminotransferase, alanine transaminase, lipase, and direct and peak total bilirubin. Nine features (age, body mass index, common bile duct stone on ultrasound, common bile duct diameter, aspartate aminotransferase, alanine transaminase, lipase, direct bilirubin, and peak total bilirubin) were clinically important and included in the machine learning algorithm. Our 9-feature model deployed on new patients was found to be highly predictive for choledocholithiasis, with an area under the receiver operating characteristic score of 0.935.
Conclusion: This multicenter study uses machine learning for pediatric choledocholithiasis. Nine clinical factors were highly predictive of choledocholithiasis, and a machine learning model trained using medical and laboratory data was able to identify children at the highest risk for choledocholithiasis.
Copyright © 2023 The Authors. Published by Elsevier Inc. All rights reserved.