Machine learning based prediction model for bile leak following hepatectomy for liver cancer.

Objective: We sought to develop a machine learning (ML) preoperative model to predict bile leak following hepatectomy for primary and secondary liver cancer.

Methods: An eXtreme Gradient Boosting (XGBoost) model was developed to predict post-hepatectomy bile leak using data from the ACS-NSQIP database. The model was externally validated using data from hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (ICC) multi-institutional databases.

Results: Overall, 20,570 and 2253 patients were identified from the ACS-NSQIP and multi-institutional databases, respectively. The incidence rates of bile leak were 7.0 %, 6.3 % and 10.2 % in the ACS-NSQIP, HCC and ICC databases, respectively. The XGBoost model achieved areas under receiver operating characteristic curves (AUROC) of 0.748, 0.719 and 0.711 in the training, testing and external validation cohorts, respectively. The SHAP algorithm demonstrated that the factors most strongly predictive of postoperative bile leak were serum alkaline phosphatase, surgical approach and cancer diagnosis. An online tool was developed for ease-of-use and clinical applicability (https://altaf-pawlik-bileleak-calculator.streamlit.app/).

Conclusion: A novel ML model demonstrated strong discrimination power to preoperatively identify patients at high risk of developing bile leak post-hepatectomy. The online calculator may be used as a clinical tool to inform preoperative surgical planning, intraoperative decision-making, and postoperative recovery protocols for patients undergoing hepatectomy.