基于机器学习算法构建瑞戈非尼对晚期肝细胞癌患者门静脉血流动力学影响的预测模型

Building a Predictive Model for the Hemodynamic Effects of Regorafenib on Portal Vein in Patients with Advanced Hepatocellular Carcinoma Based on Machine Learning Algorithms

摘要: 目的：探讨瑞戈非尼对晚期肝细胞癌患者门静脉血流动力学的影响，并基于机器学习算法利用治疗前基线临床资料构建瑞戈非尼治疗后门静脉高压变化的预测模型以指导临床实践。方法：回顾分析了2017年至2023年在青岛大学附属医院接受瑞戈非尼治疗的202例晚期肝细胞癌患者，比较了治疗前后门静脉充血指数的变化。利用患者的临床数据，多因素分析筛选出门静脉压力降低的独立风险因素。构建了六种预测模型，根据对模型训练集、测试集的综合评估(ROC曲线、校准曲线、临床决策曲线)以确定表现最佳的预测模型。结果：门静脉压力降低组的患者预后明显优于门静脉压力升高组(P = 0.029)。通过多因素分析和SHAP算法，筛选出门静脉压力降低的独立风险因素：PT百分活度、纤维蛋白原、γ-谷氨酰转肽酶(GGT)、年龄、血小板计数、白蛋白和预后营养指数水平。随机森林模型在训练/测试队列中表现最佳，AUC分别为0.829/0.889。临床决策曲线分析也显示该模型的预测能力和临床效益优于其他模型，校准曲线进一步证实了其良好的预测一致性。结论：本研究基于治疗前基线临床资料构建的随机森林机器学习模型，能够准确预测瑞戈非尼治疗后门静脉高压的变化，从而为临床治疗提供指导。

Abstract: Objective: To investigate the effects of regorafenib on the hemodynamics of the portal vein in patients with advanced hepatocellular carcinoma (HCC) and to construct a predictive model for changes in portal hypertension after regorafenib treatment based on pre-treatment baseline clinical data using machine learning algorithms, thereby guiding clinical practice. Methods: A retrospective analysis was conducted on 202 patients with advanced HCC who received regorafenib treatment at the Affiliated Hospital of Qingdao University from 2017 to 2023. Changes in portal vein congestion index before and after treatment were compared. Multivariate analysis was performed to identify independent risk factors associated with reduced portal vein pressure. Six predictive models were constructed, and the performance of these models was evaluated comprehensively using the training and testing sets (ROC curve, calibration curve, clinical decision curve) to identify the best-performing predictive model. Results: Patients in the portal vein pressure reduction group had a significantly better prognosis than those in the portal vein pressure increase group (P = 0.029). Through multivariate analysis and the SHAP algorithm, independent risk factors for reduced portal vein pressure were identified: prothrombin time percentage activity (PT %), fibrinogen, gamma-glutamyl transferase (GGT), age, platelet count, albumin, and prognostic nutritional index levels. The random forest model exhibited the best performance in both training and testing cohorts, with AUC values of 0.829 and 0.889, respectively. Clinical decision curve analysis also demonstrated that this model’s predictive ability and clinical benefit were superior to those of other models, and calibration curves further confirmed its excellent predictive consistency. Conclusion: The random forest machine learning model developed based on pre-treatment baseline clinical data can accurately predict changes in portal hypertension after regorafenib treatment, thereby providing guidance for clinical management.