基于极端梯度提升算法的肝细胞癌术后胆漏 预测模型构建与决策曲线分析

Prediction Model Construction and Decision Curve Analysis of Postoperative Bile Leakage in Hepatocellular Carcinoma Based on Extreme Gradient Boosting Algorithm

摘要: 目的：基于极端梯度提升(XGBoost)算法构建肝细胞癌(HCC)根治性肝切除术后胆漏风险预测模型，结合决策曲线分析(DCA)验证临床应用价值，为围手术期高危患者精准识别与分层管理提供量化依据。方法：采用回顾性队列研究设计，收集2015年12月至2025年12月350例HCC根治性肝切除患者临床资料，依据ISGLS标准分为胆漏组(42例)与非胆漏组(308例)。经单因素分析和LASSO回归降维筛选关键变量，按7:3分为训练集(245例)与验证集(105例)，构建XGBoost模型并五折交叉验证优化超参数。与LR、RF、GBM模型对比判别能力与临床净获益，以AUC、准确率、灵敏度、特异度评价判别能力，校准曲线与Brier评分验证校准度，DCA分析临床净获益，SHAP值解析特征重要度与交互效应。结果：350例HCC患者术后胆漏发生率12.0%，单因素联合LASSO筛选出9个关键变量。XGBoost模型训练集与验证集AUC分别为0.908、0.849，显著高于LR (0.807、0.763)、RF (0.872、0.817)及GBM (0.886、0.831)；验证集准确率、灵敏度、特异度分别为0.848、0.786、0.854，Brier评分0.11，校准曲线显示预测概率与真实发生率高度一致。DCA显示0.05~0.60风险阈值范围内XGBoost净获益持续高于LR及两种极端策略。SHAP分析表明大范围肝切除、胆道重建、术中失血量为三大核心危险因素，且存在协同交互效应。结论：基于XGBoost算法构建的HCC术后胆漏预测模型具有优异的判别能力与校准度，其临床净获益显著，可有效实现术前、术中对胆漏高危患者的精准识别，为围手术期手术策略优化、血流动力学管理、肝功能保护等干预措施的制定提供了靶向方向，具有重要的临床转化与应用价值。

Abstract: Objective: To develop a predictive model for the risk of bile leakage following curative hepatectomy for hepatocellular carcinoma (HCC) using the Extreme Gradient Boosting (XGBoost) algorithm, to validate its clinical utility through decision curve analysis (DCA), and to provide quantitative evidence for the precise identification and stratified management of high-risk patients during the perioperative period. Methods: A retrospective cohort study design was employed. Clinical data were collected from 350 patients who underwent curative hepatectomy for HCC between December 2015 and December 2025. Patients were classified into a bile leak group (42 cases) and a non-bile leak group (308 cases) according to ISGLS criteria. Key variables were screened through univariate analysis and LASSO regression for dimensionality reduction. The dataset was split 7:3 into a training set (245 cases) and a validation set (105 cases), and an XGBoost model was constructed with five-fold cross-validation to optimize hyperparameters. The discriminatory performance and clinical net benefit were compared with those of LR, RF, and GBM models. Discriminatory performance was evaluated using AUC, accuracy, sensitivity, and specificity; calibration was verified using ROC curves and Brier scores; clinical net benefit was analyzed using DCA; and feature importance and interaction effects were interpreted using SHAP values. Results: The incidence of postoperative bile leakage among 350 HCC patients was 12.0%. Univariate analysis combined with LASSO identified 9 key variables. The AUC values for the XGBoost model on the training set and validation set were 0.908 and 0.849, respectively, which were significantly higher than those of LR (0.807, 0.763), RF (0.872, 0.817), and GBM (0.886, 0.831); the validation set accuracy, sensitivity, and specificity were 0.848, 0.786, and 0.854, respectively, with a Brier score of 0.11. The calibration curve demonstrated high consistency between predicted probability and actual incidence. DCA analysis showed that within the risk threshold range of 0.05~0.60, the net benefit of XGBoost consistently exceeded that of LR and the two extreme strategies. SHAP analysis indicated that extensive liver resection, biliary reconstruction, and intraoperative blood loss were the three core risk factors, with synergistic interactions among them. Conclusion: The postoperative bile leak prediction model for HCC, constructed using the XGBoost algorithm, demonstrates excellent discriminatory power and calibration. It offers significant clinical net benefit and can effectively identify high-risk patients for bile leaks before and during surgery. This model provides a targeted direction for formulating perioperative interventions, including optimization of surgical strategies, hemodynamic management, and liver function protection, and holds significant clinical translation and application value.