Postoperative Infection after Laparoscopic Cholecystectomy: Benchmarking Machine-Learning Models in a Real-World
Hongwei Zhao1, Jianzhe Wang1, Qi Chen1
1Department of General Surgery, Daqing People's Hospital, Daqing, China.
Objective:
To benchmark machine-learning (ML) models for predicting postoperative infection after laparoscopic cholecystectomy (LC) in a real-world cohort.
Study Design:
A descriptive study. Place and Duration of the Study: Department of General Surgery, Daqing People's Hospital, Daqing, China, from January 2016 to December 2025.
Methodology:
Consecutive patients undergoing LC were included. Postoperative infection was defined as a clinically diagnosed or culture-proven infection during index hospitalisation or within 30 days after surgery, using CDC/NHSN-based criteria where applicable. The cohort (n = 1,155; 186 infections) was randomly split into training (n = 816) and internal validation sets (n = 339). Four classifiers [decision tree (DT), support vector machine with a radial basis function kernel (SVM-RBF), random forest (RF), and naive Bayes (NB)] were developed. Performance was assessed using AUC, accuracy, sensitivity, and specificity at a prespecified probability threshold of 0.50, with precision-recall analysis, calibration, decision curve analysis, and permutation feature importance serving as complementary assessments.
Results:
On validation, NB and RF showed the best discrimination (AUCs = 0.824 and 0.809, respectively), while DT and SVM-RBF performed poorly (AUCs = 0.614 and 0.582, respectively). At the 0.50 threshold, SVM-RBF predicted all validation patients as non-infected, indicating complete classification failure for clinical use. RF achieved high specificity (0.974) and accuracy (0.903), whereas NB yielded higher sensitivity (0.417). The highest-ranked RF predictors were preoperative white blood cell group, age, perioperative acute cholecystitis, the preoperative length-of-stay group, obesity, and operation duration group.
Conclusion:
ML models showed heterogeneous performance for predicting postoperative infection after LC. NB and RF performed best on internal validation and may support early risk stratification when paired with model interpretability, clinical utility assessment, external validation, and threshold optimisation. SVM-RBF was unsuitable for this dataset at the fixed 0.50 threshold.
Key Words:
Laparoscopic cholecystectomy, Postoperative infection, Real-world cohort, Machine learning, Risk stratification.