Machine learning-based prediction for incidence of endoscopic retrograde cholangiopancreatography after emergency laparoscopic cholecystectomy: A retrospective, multicenter cohort study

Background: Laparoscopic cholecystectomy is the preferred treatment for symptomatic cholelithiasis and acute cholecystitis, with increasing applications even in severe cases. However, the possibility of postoperative endoscopic retrograde cholangiopancreatography (ERCP) to manage choledocholithiasis or biliary injuries poses significant clinical challenges. This study aimed to develop a predictive model for ERCP incidence following emergency laparoscopic cholecystectomy using advanced machine learning techniques.

Methods: We conducted a retrospective cohort study using the Tokushukai Medical Database, which includes data from 42 hospitals. The study population consisted of adult patients undergoing emergency laparoscopic cholecystectomy. We used four machine learning models-logistic regression, random forest, gradient-boosting decision trees (GBDTs), and multilayer perceptrons on a dataset divided into training/validation and testing groups. We also calculated Shapley additive explanation values for GBDTs to identify variables with larger feature importance.

Results: Of 9,695 patients from July 2010 to June 2020, 8,854 met the inclusion criteria. The incidence of postoperative ERCP was 5.7% (362/6,377) and 6.4% (158/2477) in the training/validation and testing datasets, respectively. The GBDT demonstrated superior performance, with the highest predictive capacity for postoperative ERCP. Significant predictors identified included common bile duct dilatation on CT or ultrasound, serum albumin, and lactate dehydrogenase levels, which showed larger feature importance.

Conclusion: This study successfully developed a robust predictive model for ERCP following emergency laparoscopic cholecystectomy.