Construction of an Intelligent Decision-Making Model for Laparoscopic Common Bile Duct Repair

Purpose: This study aimed to develop a machine learning-based decision-support model to assist surgeons in intraoperatively selecting the optimal repair strategy (primary duct closure [PDC] versus T-tube drainage [TTD]) following laparoscopic common bile duct exploration (LCBDE). Methods: Clinical data from 117 patients with common bile duct stones (CBDS) who underwent LCBDE were retrospectively analyzed. Patients were categorized into PDC ( n = 66) and TTD ( n = 51) groups. After baseline comparison using SPSS, the dataset was standardized and split. Feature selection was performed using SelectKBest, Recursive Feature Elimination (RFE), and RFE with Cross-Validation (RFECV). Several machine learning classifiers were trained and evaluated based on accuracy, F1 score, precision, recall, and the area under the ROC curve (AUC). Results: Baseline characteristics were similar between groups, except for higher total protein and more frequent purulent bile in the TTD group. RFE identified five optimal predictive features: age, white blood cells, C-reactive protein, total protein, and albumin. The Random Forest classifier, combined with RFE feature selection, demonstrated the best predictive performance, achieving an AUC of 0.83, an accuracy of 0.72, a precision of 0.87, and a recall of 0.62. Conclusion: We successfully constructed an intelligent decision-making model for laparoscopic bile duct repair. Utilizing RFE and Random Forest, the model identifies key clinical features to provide personalized surgical recommendations, which may enhance precision treatment for patients with CBDS.