超声与机器学习预测乳腺癌新辅助化疗疗效
本研究回顾性纳入463例接受新辅助化疗的乳腺癌患者,基于超声与临床病理特征构建机器学习模型预测病理完全缓解(pCR)。XGBoost模型在测试集中表现最佳,AUC为0.8955。研究提示该模型可作为疗效评估工具,但需多中心数据验证。
要点
- XGBoost模型预测pCR效能较高,测试集AUC为0.8955。
- ER阴性、PR阴性、肿瘤短径及HER2阳性对pCR预测贡献大于15%。
- 该模型为无创预测工具,有助于个体化治疗规划。
核心结论
- 结合超声影像与临床病理特征的XGBoost模型对乳腺癌新辅助化疗后pCR具有较高的预测准确性。
研究设计与样本
- 患者随机分为训练集277例和测试集118例,采用合成少数类过采样技术处理类别不平衡。
- 通过最小绝对收缩和选择算子回归筛选出肿瘤短径、最大弹性、年龄组等10个关键特征。
- 构建随机森林、逻辑回归、极端梯度提升(XGBoost)等7种机器学习模型,采用十折交叉验证优化参数。
关键影像表现/征象
- 肿瘤短径和最大弹性为模型中的超声影像关键特征。
诊断标准与关键数值
- XGBoost模型测试集AUC为0.8955(95%CI:0.8409-0.9601),敏感度0.8095,特异度0.8026。
- Shapley加性解释分析显示ER阴性、PR阴性、肿瘤短径、HER2阳性为pCR显著预测因子,贡献大于15%。
- 校准曲线显示预测与实测一致性良好,斜率接近1,Brier评分为0.11。
临床应用与阅片要点
- 该模型可用于乳腺癌新辅助化疗过程中无创预测pCR,辅助个体化治疗决策。
- 决策曲线分析显示XGBoost在10%-90%临床决策阈值范围内提供最高净获益。
局限性
- 研究为单中心回顾性设计,需多中心数据验证模型的普适性与临床应用价值。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
OBJECTIVE: This study aimed to develop a machine learning model based on ultrasonography (US) and clinicopathological features to predict pathological complete response (pCR) following neoadjuvant chemotherapy (NAC) in patients with breast cancer. The goal was to establish a non-invasive prediction tool to facilitate individualized treatment planning.
METHODS: A retrospective analysis was conducted on data from 463 patients with breast cancer who underwent NAC at Shanxi Bethune Hospital between January 2018 and December 2024. Patients were randomly allocated into a training set (n = 277) and a test set (n = 118). To address class imbalance, the Synthetic Minority Over-sampling Technique algorithm was applied. Ten key features, including tumor short diameter, maximum elasticity, and age group, were selected through Least Absolute Shrinkage and Selection Operator regression. Seven machine learning models were constructed, including Random Forest, Logistic Regression, and Extreme Gradient Boosting (XGBoost). Model parameters were optimized through ten-fold cross-validation. Performance evaluation involved receiver operating characteristic (ROC) curves, decision curve analysis (DCA), and calibration curves.
RESULTS: Among the developed models, XGBoost demonstrated superior performance, achieving an area under the ROC curve of 0.8955 (95% confidence interval: 0.8409-0.9601), sensitivity of 0.8095, and specificity of 0.8026 in the test set. Shapley Additive Explanations analysis identified ER-negative, PR-negative, tumor short diameter, and HER2-positive as significant predictors of pCR (contribution > 15%). DCA indicated that XGBoost provided the highest net benefit within clinical decision thresholds (10%-90%), and the calibration curve demonstrated good consistency between predicted and observed outcomes, with a slope approaching 1 (Brier score = 0.11).
CONCLUSION: The XGBoost model, incorporating US imaging and clinicopathological features, demonstrated high accuracy in predicting pCR following NAC in patients with breast cancer. These findings indicate that the model may serve as a valuable tool for efficacy evaluation. Further validation with multi-center data is necessary to confirm generalizability and support clinical application.
引用
- Jia, Li, Xiao, et al. Predicting the Efficacy of Breast Cancer Neoadjuvant Chemotherapy Using Ultrasonography and Machine Learning. Ultrasound Med Biol. 2026. PMID: 41580311. doi:10.1016/j.ultrasmedbio.2025.12.009
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。