SMI-SWE多模态超声联合可解释机器学习模型在中晚期肝癌TACE-RFA术后复发预测中的应用

Application of an explainable multimodal SMI-SWE ultrasound-based machine learning model in predicting recurrence after TACE-RFA for intermediate to advanced liver cancer

摘要:

背景与目的 中晚期肝癌患者接受经导管动脉化疗栓塞（TACE）联合射频消融（RFA）序贯治疗后复发率仍较高，术前缺乏可靠的风险分层工具。本研究基于超微血管成像（SMI）与剪切波弹性成像（SWE）构建多模态超声联合机器学习模型，并结合SHAP算法进行可解释性分析，旨在预测中晚期肝癌患者TACE-RFA序贯治疗后的复发风险。方法 回顾性纳入2022年6月—2024年6月215例BCLC B/C期肝癌患者，按7∶3比例分为训练集（n=150）与测试集（n=65）。收集临床资料、生化指标及SMI、SWE超声特征。采用随机森林、支持向量机及极限梯度提升（XGB）进行变量重要性排序，结合交集筛选及双向逐步Logistic回归确定独立预测因子。构建7种机器学习模型，通过10折交叉验证优化参数。以曲线下面积（AUC）、敏感度、特异度、Brier评分及决策曲线分析（DCA）评价模型性能，并利用Tree SHAP解释模型输出。结果 术前肿瘤最大径、甲胎蛋白（AFP）、最大弹性模量（Emax）、弹性比值（Eratio）及血管形态为独立预测因子。XGB模型在训练集与测试集中的AUC分别为0.989和0.959，Brier评分分别为0.034和0.056，DCA显示其在较宽阈值范围内具有最佳净收益。SHAP分析表明，高Eratio、高AFP、血管形态不规则及肿瘤最大径较大显著增加复发风险，而较低Emax与复发风险降低相关。结论 基于SMI-SWE多模态超声构建的XGB模型可较准确预测中晚期肝癌患者TACE-RFA序贯治疗后的复发风险，具有良好的临床应用潜力。SHAP分析增强了模型的可解释性，为术前个体化风险评估和治疗决策提供了量化依据。

Abstract:

Background and Aims Recurrence remains common in patients with intermediate to advanced liver cancer after sequential transarterial chemoembolization (TACE) and radiofrequency ablation (RFA). Reliable preoperative risk stratification tools are lacking. This study aimed to develop and validate a multimodal ultrasound-based machine learning model integrating superb microvascular imaging (SMI) and shear wave elastography (SWE) to predict postoperative recurrence after sequential TACE-RFA therapy in patients with intermediate to advanced liver cancer, with model interpretability assessed using SHAP analysis.Methods A total of 215 patients with BCLC stage B/C liver cancer were retrospectively enrolled and randomly divided into a training set (n=150) and a testing set (n=65) at a 7∶3 ratio. Clinical characteristics, laboratory parameters, and SMI/SWE-derived features were collected. Feature importance was ranked using random forest, support vector machine, and extreme gradient boosting (XGB), followed by intersection selection and bidirectional stepwise logistic regression to identify independent predictors. Seven machine learning algorithms were trained with 10-fold cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, Brier score, and decision curve analysis (DCA). Tree SHAP was applied to interpret feature contributions.Results Preoperative tumor diameter, AFP level, maximum elastic modulus (Emax), elastic ratio (Eratio), and vascular morphology were identified as independent predictors of recurrence. The XGB model achieved the best performance, with AUCs of 0.989 in the training set and 0.959 in the testing set. Brier scores were 0.034 and 0.056, respectively. DCA demonstrated superior net benefit across a wide range of threshold probabilities. SHAP analysis indicated that higher Eratio, elevated AFP, irregular vascular morphology, and larger tumor diameter were associated with increased recurrence risk, whereas lower Emax was associated with reduced risk.Conclusion The proposed SMI-SWE-based XGB model provides accurate and clinically useful prediction of recurrence after sequential TACE-RFA therapy in intermediate to advanced liver cancer. SHAP-based interpretation enhances model transparency and supports individualized preoperative risk assessment.