基于钆塞酸二钠增强MRI特征构建肝细胞癌CK19表达术前预测模型

Machine Learning Models Based on Gd-EOB-DTPA Enhanced MRI Features for Preoperative Prediction of Cytokeratin 19 Expression in Hepatocellular Carcinoma

目的:基于钆塞酸二钠增强MRI影像特征构建肝细胞癌(HCC)细胞角蛋白19(CK19)表达情况的术前预测模型。方法:回顾性分析2015年5月至2022年8月期间经术后病理证实为HCC的226例患者临床、MRI影像特征,其中CK19阳性75例,阴性151例。将患者按照7∶3比例随机分为训练集及测试集,分别采用单变量方差分析方法(ANOVA_F)和基于随机森林的递归特征消除法(RFE?RF)进行特征筛选,然后分别采用逻辑回归(LR)和随机森林(RF)算法构建CK19表达术前预测模型,并对模型进行验证和评估。结果:通过ANOVA_F方法筛选出5个特征,包括肝炎、边缘、包膜凹陷征、马赛克征、瘤内动脉。ANOVA_F LR模型在训练集和测试集上的AUC(95%CI)分别为0.668(0.576~0.760)、0.654(0.518~0.789),ANOVA_F RF模型AUC(95%CI)分别为0.675(0.584~0.766)、0.647(0.505~0.790)。通过RF-RFE方法筛选出5个特征,包括年龄、log AFP、肿瘤大小、形态、包膜。RFE-RF LR模型在训练集和测试集上的AUC(95%CI)分别为0.624(0.532~0.716)、0.583(0.423~0.744),RFE-RF RF模型AUC(95%CI)分别为0.997(0.993~1.000)、0.706(0.568~0.845)。在测试集上,RFE?RF RF模型AUC值高于其他三种模型,其准确率、精准率和F1得分依次为70.60%、69.80%和70.10%,优于其他三种模型。决策曲线分析显示RFE_RF RF模型的临床应用价值高于其余三种模型。结论:基于术前临床、MRI影像特征建立的RFE_RF RF模型效能优于其他三种模型,可较好地术前预测HCC CK19表达情况。

Purpose:To develop preoperative prediction models for the expression of cytokeratin 19(CK19)in hepatocellular carcinoma(HCC)based on gadoxetic acid disodium(Gd-EOB-DTPA)enhanced MRI features.Methods:The clinical and MR data of 226 HCC patients from May 2015 to August 2022 were retrospectively analyzed,including 75 CK19 positive and 151 negative cases.The patients were randomly divided into a training set and a testing set at a ratio of 7∶3.Feature selection was performed using univariate statistical ANOVA_F,as well as the random forest-based recursive feature elimination method(RFE-RF),respectively.Subsequently,logistic regression(LR)and random forest(RF)algorithms were employed to build models,which were then validated and evaluated.Results:Five characteristics,namely hepatitis,tumor margin,capsular retraction,mosaic,and intratumoral vessels on arterial phase were selected by ANOVA_F.The ANOVA_F LR model achieved AUCs of 0.668(95%CI:0.576-0.760)and 0.654(95%CI 0.518-0.789)in the training and test sets,respectively,while the ANOVA_F RF model outperformed with AUCs of 0.675(95%CI 0.584-0.766)and 0.647(95%CI 0.505-0.790)in the same sets.Five characteristics were screened by RF-RFE,including age,log AFP level,tumor size,shape,and capsule.The RFE-RF LR model achieved AUCs of 0.624(95%CI 0.532-0.716)and 0.583(95%CI 0.423-0.744)in the training and test sets,respectively,while the RFE-RF RF model outperformed with AUCs of 0.997(95%CI 0.993-1.000)and 0.706(95%CI 0.568-0.845)in the same sets.The RFE-RF RF model exhibited a higher AUC than the other three models in the test set,demonstrating its superiority in terms of accuracy,precision,and F1 score with values of 70.60%,69.80%,and 70.10%,respectively.Decision curve analysis further confirmed the clinical usefulness of the models.Conclusions:The RFE-RF RF model established based on preoperative clinical and imaging features can be used to predict HCC CK19 expression effectively,demonstrating its superior efficacy compared to the other three models.