Development and Internal Multicenter Validation of a Deep Learning Model for Predicting Post-Hepatectomy Liver Failure in Patients with Hepatocellular Carcinoma: A Multicenter Study

TL;DR

A deep learning model was developed and validated across multiple hospitals to better predict liver failure after liver cancer surgery, potentially improving patient outcomes.

Contribution

A deep learning model was developed and internally validated across multiple centers to predict post-hepatectomy liver failure with higher accuracy than traditional methods.

Findings

The deep learning model achieved AUCs of 0.914, 0.892, and 0.906 in training, validation, and test sets.

Key predictors included ALBI and MELD scores, prothrombin time, intraoperative blood loss, and resection extent.

The model outperformed logistic regression and showed strong calibration and clinical utility.

Abstract

Post-hepatectomy liver failure is a serious complication that can occur after liver surgery for liver cancer and may lead to poor recovery or even death. Predicting which patients are at high risk before surgery remains difficult using traditional clinical tools. In this study, we developed a deep learning model that analyzes many clinical and surgical factors at the same time to provide a more accurate prediction of liver failure after surgery. The model was tested in patients from multiple hospitals and showed strong and consistent performance. By helping surgeons identify high-risk patients earlier, this model may support safer surgical planning and improve postoperative care.

Background/Objectives: Post-hepatectomy liver failure is one of the most serious complications after liver resection for hepatocellular carcinoma and is associated with high morbidity and mortality. Traditional clinical scoring systems and statistical models provide limited predictive accuracy. This study aimed to develop and internally validate a deep learning model for predicting post-hepatectomy liver failure using multicenter clinical data. Methods: A retrospective cohort of 498 patients from six centers undergoing curative-intent liver resection for hepatocellular carcinoma was analyzed. Preoperative biochemical parameters, intraoperative surgical variables, and tumor-related characteristics were incorporated into a deep neural network, with logistic regression as a baseline comparator. Data splitting was performed before preprocessing, and imputation/scaling parameters were fitted on the training set only to prevent information leakage. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC) and precision–recall (PR) curves; calibration was assessed using calibration plots and Brier score; and clinical utility was assessed using decision curve analysis (DCA). A sensitivity analysis using a preoperative-only feature set (excluding intraoperative variables) was also conducted. SHapley Additive exPlanations were used to determine variable importance. Results: The deep learning model achieved AUCs of 0.914, 0.892, and 0.906 in the training, validation, and test sets, outperforming logistic regression (0.782, 0.757, and 0.773). Key predictors included ALBI and MELD scores, prothrombin time, intraoperative blood loss, and resection extent. Calibration and decision curve analysis further supported the robustness and clinical utility of the model. Conclusions: The deep learning model provides improved predictive performance for post-hepatectomy liver failure compared with logistic regression in an internally validated multicenter cohort and may support perioperative risk stratification and surgical planning.

Figures