Abstract:

Objective To improve the accuracy and clinical interpretability of recurrence prediction after liver transplantation in patients with hepatocellular carcinoma（HCC），and to explore the potential application of explainable machine learning models in integrating multimodal data. Methods This study included data from 138 liver transplant patients with liver cancer at the Third Medical Center of the Chinese People's Liberation Army General Hospital from December 2018 to December 2021. Preoperative contrast-enhanced CT radiomics features and clinical variables were extracted. Predictive models were developed using four machine learning methods ：LASSO regression，random forest，support vector machine（SVM），and neural network. Importance of variate was employed to identify key predictive factors. Model performance was evaluated using area under the receiver operating characteristiccurve（AUC），Brier score，and calibration curves. A nomogram was ultimately constructed based on important variables. Results The AUCs for the random forest model in predicting recurrence at 1，2，and 3 years were 0.881，0.906，and 0.915，respectively，significantly outperforming other models. The importance analysis identified five key imaging features，and the nomogram model combining these with clinical variables demonstrated good consistency and predictive capability. Conclusion Explainable machine learning models based on multimodal data can effectively improve the accuracy and transparency of recurrence prediction following liver transplantation in HCC patients. These models have strong clinical applicability and provide valuable support for individualized preoperative risk assessmentand treatment decision-making. 

Key words:

Hepatocellular carcinoma, Liver transplantation, Machine learning, Multimodal data

摘要：

目的 提高肝细胞癌（hepatocellular carcinoma，HCC）患者肝移植术后复发预测的准确性与临床可解释性，探索可解释机器学习模型在融合多模态数据中的应用潜力。 方法 本研究纳入中国人民解放军总医院第三医学中心 2018 年 12 月至 2021 年 12 月间 138 例肝癌肝移植患者资料，提取术前增强CT 影像组学特征和临床资料，采用 LASSO 回归、随机森林、支持向量机和神经网络构建预测模型，并通过变量重要性分析识别关键预测因子。通过 AUC、Brier 评分和校准曲线等指标评估模型性能，最终构建列线图。 结果 随机森林在 1、2、3 年复发预测中的 AUC 分别为 0.881、0.906 和 0.915，显著优于其他模型。重要性分析识别出 5 个关键影像特征，结合临床变量构建的列线图模型表现出良好的一致性和预测能力。 结论 基于多模态数据的可解释机器学习模型能有效提高肝癌肝移植术后复发预测的准确性与透明度，具有良好的临床适用性，为个体化术前风险评估与治疗决策提供支持。

关键词: 肝细胞癌 , 肝移植 , 机器学习 , 多模态数据