A Predictive Model Integrating Coagulation Profiles for Preoperative Risk Stratification of Difficult Single-Incision Laparoscopic Cholecystectomy: A Retrospective Cohort Study Based on Explainable Machine Learning

Background: Accurate preoperative prediction of surgical difficulty in single-incision laparoscopic cholecystectomy (SILC) remains challenging, as models for multiport surgery are not directly applicable. Methods: This retrospective single-center study included 520 patients, who were randomly divided into a training cohort (n=364) and an internal validation cohort (n=156). Six predictors were identified, guiding 8 machine learning models. Performance was assessed using the area under the curve (AUC), calibration, decision curve analysis, and classification metrics. The optimal model was interpreted through SHapley Additive exPlanations and deployed online. Results: The optimal neural network achieved excellent validation performance (AUC: 0.867, accuracy: 0.833, specificity: 0.941, NPV: 0.855), with good calibration (Brier 0.118) and net benefit across clinical thresholds. Ablation analysis showed AUC decreased by 0.043 when coagulation parameters were excluded; correlation analyses suggested they provide complementary predictive value. Exploratory sensitivity analysis using a hard-event endpoint maintained good discrimination (AUC: 0.877), supporting robustness of the composite outcome. The model is publicly available as an online calculator at https://newprediction.shinyapps.io/predict/. Conclusion: We developed an interpretable machine learning model incorporating coagulation parameters to identify patients at increased risk of difficult SILC. The model demonstrated favorable predictive performance and may support preoperative risk stratification, patient selection, and individualized surgical planning.