Predicting pathological response to neoadjuvant chemotherapy in breast cancer patients based on imbalanced clinical data
Predicting pathological response to neoadjuvant chemotherapy in breast cancer patients based on imbalanced clinical data
复制标题
基于不平衡临床数据预测乳腺癌患者新辅助化疗的病理反应
DOI：
10.1007/s00779-018-1144-3
复制
发表时间：
2018-10-01
影响因子：
--
通讯作者：
Han, Bing
中科院分区：
文献类型：
--
作者：
Gao, Ting;Hao, Yaguang;Han, Bing
新辅助化疗（NAC）可能有助于一些乳腺癌患者随后的手术或放疗。然而，与NAC相关的风险是存在的。为了降低风险，可以使用机器学习方法根据临床数据辅助乳腺肿瘤的诊断。本研究通过实际临床数据探讨了集成机器学习模型在预测乳腺癌患者对NAC的病理反应中的应用。确定了集合k-最近邻（EKNN）模型来预测病理反应。回顾性分析NAC患者不平衡的临床资料，从所有特征中选取11个临床病理变量，建立简洁的EKNN模型。共有259例患者的临床资料被纳入模型。每个k近邻（KNN）的训练集和测试集分别包含27例和9例患者。数据库中共有259例乳腺癌患者，其中病理完全缓解36例，部分缓解157例，病情稳定66例。为了解决临床数据不平衡的问题，设计了一个集成学习的EKNN，其中基本学习器中每个类别的样本数量设置为最小数量36。结果表明，EKNN模型对乳腺癌NAC术后病理反应的分类准确率为81.48%，Kappa系数为0.72，鲁棒性和泛化性均优于单一KNN模型的平均预测能力（单一KNN模型的平均准确率为62.22%，Kappa系数为0.43）。根据实际临床数据，选择重要的临床病理变量，并通过集成EKNN模型很好地解决了不平衡问题。该模型提高了在临床数据不平衡的情况下预测病理反应的稳健性和通用性。它表明，集成机器学习在协助癌症阶段诊断和精准医疗方面可能具有实际应用。
Neoadjuvant chemotherapy (NAC) may help some breast cancer patients with subsequent surgery or radiotherapy. However, there are certain risks associated with NAC. To lower the risks, machine-learning methods can be used to assist the diagnosis of breast tumors based on clinical data. This study investigated the use of ensemble machine-learning models in the prediction of pathological response to NAC for breast cancer patients with actual clinical data. The ensemble k-nearest neighbor (EKNN) model was determined to predict pathological responses. The imbalanced clinical data of patients with NAC were reviewed retrospectively, and 11 clinicopathological variables were selected from all features to establish succinct EKNN model. A total of 259 patients’ clinical data was included in the model. The training and testing set for each single k-nearest neighbor (KNN) contained 27 and 9 patients, respectively. A total of 259 breast cancer patients in the database included 36 cases of pathological complete response, 157 cases of partial response, and 66 cases of stable disease. To solve the imbalanced clinical data problem, an ensemble-learning EKNN was designed, where the number of samples for each class in a base learner is set to equal to the minimum number 36. It showed that the classification accuracy of pathological response for breast cancer patients after NAC was 81.48% by EKNN model and the Kappa coefficient was 0.72, indicating that the robustness and generalization were better than the average prediction ability of single KNN model (average accuracy of single KNN model was 62.22% and Kappa coefficient was 0.43). Based on actual clinical data, important clinicopathological variables are selected, and the imbalanced problem are well solved by the ensemble EKNN model. The model improved the robustness and generalization for predicting the pathological response with imbalanced clinical data. It suggested that ensemble machine learning has possible practical applications for assisting cancer stage diagnoses and precision medicine.