临床及CT机器学习模型预测肝细胞癌患者首次TACE后急性肝功能恶化  

Clinical and CT machine learning model for predicting acute liver function deterioration in hepatocellular carcinoma patients after the first time TACE

作　　者：任泳年 唐昌乾 魏星博 李冬筱 王连才 李德宇 REN Yongnian;TANG Changqian;WEI Xingbo;LI Dongxiao;WANG Liancai;LI Deyu(Department of Hepatobiliary and Pancreatic Surgery,People’s Hospital of Zhengzhou University,Zhengzhou 450003,China;Department of Hepatobiliary and Pancreatic Surgery,Henan University People’s Hospital,Zhengzhou 450003,China;Department of Gastroenterology,People’s Hospital of Zhengzhou University,Zhengzhou 450003,China)

机构地区：[1]郑州大学人民医院肝胆胰腺外科,河南郑州450003 [2]河南大学人民医院肝胆胰腺外科,河南郑州450003 [3]郑州大学人民医院消化内科,河南郑州450003

出　　处：《中国介入影像与治疗学》2025年第3期153-158,共6页Chinese Journal of Interventional Imaging and Therapy

基　　金：河南省科技攻关项目(232102311024)。

摘　　要：目的观察基于治疗前临床及CT特征构建的机器学习(ML)模型预测肝细胞癌(HCC)患者首次TACE后急性肝功能恶化(ALFD)的价值。方法回顾性收集320例接受首次TACE的HCC患者,按4:1比例划分为训练集(n=256)与测试集(n=64);根据TACE后2周内临床、实验室及影像学所见评估有无ALFD。以单因素分析比较训练集有、无ALFD患者临床基线资料及TACE前CT所示病灶直径,基于差异有统计学意义的参数分别以9种ML算法构建ML模型,以测试集验证模型效能并筛选最优模型。于测试集评估最优模型校准度和临床价值,并以SHAP法分析各参数贡献度。结果训练集76例ALFD、180例非ALFD;测试集18例ALFD、46例非ALFD。9种ML模型中,极限梯度提升(XGBoost)模型在测试集的敏感度、特异度、准确率、曲线下面积、F1值及Kappa值分别为85.12%、89.34%、88.08%、0.927、0.811及0.725,为最优模型;其对于测试集的预测概率与实际概率的一致性良好,临床净获益较高;患者年龄、TACE前CT所示病灶直径、谷丙转氨酶、谷草转氨酶及TACE用时对该模型的贡献均较大。结论基于治疗前临床及CT特征构建的XGBoost模型能有效预测HCC患者首次TACE后ALFD。Objective To observe the value of machine learning(ML)models constructed based on pre-treatment clinical and CT features for predicting acute liver function deterioration(ALFD)in hepatocellular carcinoma(HCC)patients after the first time TACE.Methods Totally 320 HCC patients who underwent the first TACE were retrospectively enrolled and divided into training set(n=256)and test set(n=64)at the ratio of 4∶1.ALFD was evaluated according to clinical,laboratory and image findings within 2 weeks after TACE.Univariate analysis was performed to compare clinical baseline data and diameter of HCC on pre-TACE CT in training set,and parameters being statistical different between patients with and without ALFD were used to construct ML models using 9 different ML algorithms.The efficacy of each model for predicting ALFD in test set was evaluated,and the optimal model was selected.The calibration degree and clinical value of the optimal model were assessed in test set,and the contribution of each parameter was analyzed using SHAP method.Results In training set,76 cases were ALFD and 180 cases were non-ALFD,while in test set,18 cases were ALFD and 46 cases were non-ALFD.Among 9 ML models,the sensitivity,specificity,accuracy,area under the curve,F1 value and Kappa value of extreme gradient boosting(XGBoost)model in test set was 85.12%,89.34%,88.08%,0.927,0.811 and 0.725,respectively.XGBoost model was considered as the optimal one,with predicted probability in test set in good agreement with actual probability and high clinical net benefit.The contribution of patients’age,lesion diameter on pre-TACE CT,glutamic-pyruvic transaminase,glutamicoxaloacetic transaminase and TACE time were all great for XGBoost model.Conclusion XGBoost model based on pretreatment clinical and CT features could be used to effectively predict ALFD in HCC patients after the first time TACE.

关 键 词：癌 肝细胞 化学栓塞 治疗性 机器学习 急性肝功能恶化 

分 类 号：R735.7[医药卫生—肿瘤] R815[医药卫生—临床医学]