超声影像组学联合临床特征的机器学习模型预测HER2阳性乳腺癌新辅助治疗病理完全缓解的价值

Clinical value of a machine learning model integrating ultrasound radiomics and clinical features for predicting pathological complete response to neoadjuvant therapy in HER2-positive breast cancer

摘要:

背景与目的 人表皮生长因子受体2（HER2）阳性乳腺癌具有高度侵袭性及明显异质性，新辅助化疗（NAC）联合抗HER2靶向治疗虽可提高病理完全缓解（pCR）率，但不同患者治疗反应差异较大。构建治疗前无创预测模型，对筛选潜在获益人群及优化个体化治疗具有重要意义。本研究探讨基于超声影像组学及临床病理特征构建的机器学习模型预测HER2阳性乳腺癌患者NAC联合抗HER2靶向治疗后pCR的临床价值。方法 回顾性纳入2019年10月—2024年10月湖南省肿瘤医院及中南大学湘雅三医院经病理证实并接受NAC联合抗HER2靶向治疗的HER2阳性乳腺癌患者353例。其中湖南省肿瘤医院275例作为内部数据集，中南大学湘雅三医院78例作为外部验证集。基于治疗前超声图像提取影像组学特征，并结合临床病理特征，经相关性分析及LASSO回归筛选关键特征后，分别采用极端梯度提升（XGBoost）、随机森林及朴素贝叶斯算法构建临床模型、超声影像组学模型及联合模型。采用受试者工作特征曲线评估模型预测效能，采用决策曲线分析（DCA）评价模型临床获益，并利用SHAP算法解释最优模型特征重要性。结果 共筛选出12个临床病理特征及7个超声影像组学特征用于建模。三种算法中，XGBoost模型整体预测性能最佳。与单纯临床模型或超声影像组学模型相比，联合模型显示出更优预测效能。XGBoost联合模型在训练集、内部验证集及外部验证集中的曲线下面积分别为0.964（95% CI=0.893~0.965）、0.857（95% CI=0.822~0.960）和0.771（95% CI=0.765~0.946）。DCA结果显示，该模型在多数阈值概率范围内均具有较高净获益。SHAP分析提示HER2表达状态、靶向治疗方式、雌激素状态及部分纹理特征对模型预测贡献较大。结论 基于超声影像组学联合临床病理特征构建的XGBoost机器学习模型能够较准确地预测HER2阳性乳腺癌患者NAC联合抗HER2靶向治疗后的pCR状态，具有较好的临床应用潜力，可为个体化治疗决策提供参考。

Abstract:

Background and Aims HER2-positive breast cancer is characterized by strong invasiveness and marked heterogeneity. Although neoadjuvant chemotherapy (NAC) combined with anti-HER2 targeted therapy significantly improves the pathological complete response (pCR) rate, therapeutic responses vary substantially among patients. Therefore, developing a noninvasive pretreatment predictive model is of great importance for individualized treatment optimization. This study aimed to evaluate the clinical value of machine learning models based on ultrasound radiomics and clinicopathological features in predicting pCR in HER2-positive breast cancer after NAC combined with anti-HER2 targeted therapy.Methods A total of 353 patients with pathologically confirmed HER2-positive breast cancer who received NAC combined with anti-HER2 targeted therapy between October 2019 and October 2024 were retrospectively enrolled from Hunan Cancer Hospital and the Third Xiangya Hospital of Central South University. Among them, 275 patients from Hunan Cancer Hospital were assigned to the internal dataset, while 78 patients from the Third Xiangya Hospital were used as the external validation dataset. Ultrasound radiomics features were extracted from pretreatment ultrasound images and combined with clinicopathological features. After correlation analysis and least absolute shrinkage and selection operator (LASSO) regression for feature selection, Extreme Gradient Boosting (XGBoost), Random Forest, and Na?ve Bayes algorithms were applied to construct clinical models, radiomics models, and combined models. Receiver operating characteristic curves were used to evaluate predictive performance, decision curve analysis (DCA) was performed to assess clinical utility, and Shapley Additive Explanation (SHAP) analysis was used to interpret feature importance in the optimal model.Results Twelve clinicopathologic features and seven ultrasound radiomics features were finally selected for model construction. Among the three algorithms, the XGBoost model demonstrated the best predictive performance. Compared with the clinical model or radiomics model alone, the combined model achieved superior predictive efficacy. The AUCs of the XGBoost combined model for predicting pCR were 0.964 (95% CI=0.893-0.965), 0.857 (95% CI=0.822-0.960), and 0.771 (95% CI=0.765-0.946) in the training, internal validation, and external validation cohorts, respectively. DCA demonstrated favorable net clinical benefits across a wide range of threshold probabilities. SHAP analysis revealed that HER2 expression status, targeted therapy strategy, estrogen receptor status, and several texture features contributed substantially to model prediction.Conclusion The XGBoost machine learning model integrating ultrasound radiomics and clinicopathological features can effectively predict pCR after NAC combined with anti-HER2 targeted therapy in HER2-positive breast cancer, showing promising clinical utility for individualized therapeutic decision-making.