Introduction
Predicting choledocholithiasis in acute pancreatitis is challenging. Existing predictive models typically rely on the common bile duct (CBD) width, a parameter difficult to assess reliably. We aimed to develop a model based solely on easily available clinical and laboratory markers, excluding CBD width.
Aims & Methods
We conducted a retrospective cohort study of 348 patients treated for acute pancreatitis at the LMU University Hospital in Munich, Germany, between 2005 and 2021. A Random Forrest (RF) model was developed using the TensorFlow Decision Forests (TFDF) library with 14 clinical and laboratory features, excluding CBD width. Model performance was assessed using sensitivity, specificity, predictive values, and area under the receiver operating characteristic curve (AUC-ROC).
Results
Of 348 patients with acute biliary pancreatitis, 167 (48.0%) had CBD stones. Our model achieved a sensitivity of 92.3%, a negative predictive value (NPV) of 83.3%, and an AUC-ROC of 0.739. Compared to using CBD width ≥8 mm or bilirubin ≥4 mg/dl as predictors (as guidelines recommend), our model demonstrated superior sensitivity (92.3% vs. 79.5% vs. 30.8%, respectively) and NPV (83.3% vs. 74.3% vs. 54.8%, respectively). The rate of missed intraductal concrements was substantially lower with our model (4.3% vs. 41.4%). In subgroup analysis, our model performed better in predicting stones (AUC=0.8) than microlithiasis or sludge (AUC=0.623).
Conclusion
We developed the first machine learning-based score to predict choledocholithiasis in acute pancreatitis without relying on CBD width. The model's high sensitivity and NPV make it a promising tool for clinical decision-making, specifically for hospitals without the option of endosonography.
Disclosure
SS is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 413635475 - and the LMU Munich Clinician Scientist Program (MCSP).