【摘要】 目的 采用机器学习方法构建并验证乳腺癌多柔比星+环磷酰胺+多西他赛（TAC）化学治疗（简称化疗）方案致骨髓抑制的风险预测模型。方法 回顾性收集医院2019年1月至2023年10月收治的接受TAC化疗方案的女性乳腺癌患者的病历资料。对数据进行预处理后，综合运用随机森林、极限梯度提升模型进行特征筛选，并结合单因素分析与多因素Logistic回归分析确定最佳特征变量。基于筛选出的特征，构建支持向量、逻辑回归、随机森林、极限梯度提升和类别型特征梯度提升5种预测模型，在测试集中进行性能验证与比较，筛选最优模型。采用Shapley加法解释法对最优模型进行可解释性分析。结果 共纳入240例患者，其中发生骨髓抑制170例（70.83%）。最佳特征变量为化疗前白细胞计数、淋巴细胞数、血清白蛋白水平、预防性使用集落刺激因子、化疗周期、体质量指数及基础疾病。热图分析结果显示，各特征间相关性较弱，均适宜纳入模型。随机森林模型为最优预测模型，受试者工作特征曲线下面积为0.965，预测性能与校准度良好。结论 基于机器学习方法构建的随机森林模型在预测乳腺癌TAC化疗方案致骨髓抑制方面具有较好的性能与临床实用性，可作为个体化风险评估的辅助工具。更多还原

【Abstract】 Objective To construct and validate a risk prediction model for myelosuppression induced by the doxorubicin+cyclophosphamide + docetaxel(TAC)chemotherapy regimen in breast cancer based on the machine learning approach.Methods Medical records of female patients with breast cancer admitted to the hospital from January 2019 to October 2023 and received TAC chemotherapy regimen were collected. After data preprocessing,the feature selection was performed by the Random Forest(RF) and Extreme Gradient Boosting(XGB) models,and the optimal feature variables were determined through the univariate analysis and multivariate Logistic regression analysis. Based on the selected features,five prediction models [support vector machine(SVM),Logistic Regression(LR),RF,XGB,and Categorical Boosting(CatBoost)] were constructed. Performance validation and comparison were conducted on the test set to select the optimal model. The interpretability of the optimal model was further analyzed using Shapley’s additive interpretation(SHAP) method.Results A total of 240 patients were included,among whom 170 patients(70. 83%) experienced myelosuppression. The optimal feature variables were pre-chemotherapy leukocyte and lymphocyte counts,serum albumin level,prophylactic use of colony-stimulating factors,chemotherapy cycles,body mass index,and underlying diseases.Heatmap analysis results showed weak correlations among the features,all of which were suitable for model construction. The RF model was identified as the optimal predictive model,with an area under the receiver operating characteristic curve(AUC) of 0. 965,it had good predictive performance and calibration,which could serve as a reference for clinical decision-making.Conclusion The RF model constructed based on the machine learning approach has good performance and clinical utility in predicting myelosuppression induced by TAC chemotherapy regimens for breast cancer,and it can be used as a supplementary tool for personalized risk assessment.更多还原