Indication model for laparoscopic repeat liver resection in the era of artificial intelligence: machine learning prediction of surgical indication.

Background: Laparoscopic repeat liver resection (LRLR) is still a challenging technique and requires a careful selection of indications. However, the current difficulty scoring system is not suitable for selecting indications. The purpose of this study is to develop the indication model for LRLR using machine learning and to identify factors associated with open conversion (OC).

Methods: Patients who underwent repeat hepatectomy (2017-2021) at Samsung Medical Center 2021 were investigated. Multiple indication models were developed using machine learning techniques (random forest, SVM, XGB) and logistic regression. The predictive performance of these models was compared, and risk factors associated with OC were analyzed.

Results: Among 221 patients (110 LRLR, 111 ORLR), the ORLR group had a higher previous open approach rate (75.7% vs. 38.2%, p<0.001). Twice previous abdominal surgery was the only independent OC risk factor (OR 6.56, p=0.009). The indication model showed moderate predictive power (random forest AUC=0.779, logistic regression AUC=0.725, p=0.710). Important variables were previous laparoscopic approach, present subsegmentectomy, and left-sided tumor location.

Conclusion: The performance of the indication model for LRLR showed moderate predictive power in both machine learning and logistic regression. The important variables for LRLR were previous laparoscopic approach, present subsegmentectomy, and left side location.