To construct a predictive model for early identification of post-hepatectomy liver failure (PHLF) for liver cancer using the extreme gradient boosting (XGBoost), and evaluate its predictive efficiency.

Clinical data of 583 patients with liver cancer who underwent hepatectomy in Nanfang Hospital from November 2018 to January 2022 were retrospectively analyzed. Among them, 504 patients were male and 79 female, aged from 23 to 77 years, with a median age of 54 years. Clinical indexes of all enrolled patients before, during and after hepatectomy were collected, and relevant clinical scores were calculated. The data set was randomly divided into the training and validation sets according to a ratio of 8∶2. Based on XGBoost, the PHLF prediction model was constructed by using 5-fold cross validation and loss function to adjust the hyperparameters. The prediction efficiency of XGBoost model was evaluated by the ROC curve and compared with traditional scoring system. Meantime, the importance of the feature variables was ranked. Shapley additive explanation (SHAP) was used to visually explain the model.

Among 583 patients, 467 cases were assigned into the training set and 63 cases developed PHLF. 116 cases were allocated in the validation set and 15 developed PHLF. Univariate analysis and Lasso regression analysis showed that 8 clinical indexes including preoperative INR, AST, ALB, operation time, extensive hepatectomy, INR, AST and TB at postoperative 1 d (D1) were significantly associated with PHLF in the training set. The area under the ROC curve (AUC) of XGBoost model in training and validation sets was 0.973 and 0.904, respectively. SHAP value was employed to quantify the impact of each feature on the prediction results of the model, D1 INR had the largest weight, and high D1 AST was positively correlated with the increase of PHLF risk. Based on INR and AST, the PHLF prediction score formula was constructed, and the PHLF prediction score=-13.395+1.2×preoperative AST(U/L)/100+9.236×D1 INR. The AUC of the scoring model was 0.838, the sensitivity was 0.825 and the specificity was 0.748, respectively.

The PHLF prediction model based on XGBoost yields high accuracy and robustness in both the training and validation sets of liver cancer, which has the potential as an auxiliary tool for clinical decision-making, contributes to promptly identifying patients with high-risk PHLF and delivering immediate interventions, thereby improving clinical prognosis of patients with liver cancer.